That admission is the real story. SAP has unveiled what it calls the Autonomous Enterprise, a model where agentic AI handles finance, procurement, HR, supply chain and customer operations end-to-end, without employees ever touching a screen. Fifty Joule Assistants. More than 200 specialised agents. A Knowledge Graph designed to map every relationship between every business entity. A clear bet that the winning layer of enterprise AI is the governance wrapped around the model.

I think Klein is right about the destination. I also think most organisations are nowhere near ready for it, and the gap between the press release and the operating reality is where a lot of money is about to be wasted.

Here is why. The same finance director who is being pitched autonomous month-end close is currently sitting in a town hall explaining why Copilot adoption is at 14% and the legal team has banned ChatGPT from contract drafting. The same operations leader who is being invited to a roadmap session on agentic procurement cannot get her team to stop pasting supplier data into a personal email account. The capability floor of most enterprises is far lower than the capability ceiling vendors are selling against.

This is a critique of shortcuts rather than of SAP. The Autonomous Enterprise is a real destination, and Klein's framing that "the difference is context" is the most useful thing any major vendor has said about enterprise AI in two years. Generic models do not know your approval thresholds. They do not know your regulatory carve-outs or why the Tuesday reconciliation never balances. SAP merging large language models with its 7.3 million data fields and built-in governance is a credible answer to that problem, in theory.

In practice, autonomy is something you become rather than something you buy. A handful of conditions have to be true before an agent can be trusted to approve a payment or raise a purchase order, let alone close a quarter on your behalf.

Your data has to mean something. If two systems disagree about what a customer is, an agent will pick a side and call it a decision. Reconciliation is unglamorous, but it is the precondition for everything else.

Your processes have to be legible. If your closing process lives in a senior manager's head and three colour-coded spreadsheets, you cannot delegate it to a machine. You can barely delegate it to a new hire.

Your people have to understand what the agent is doing. Klein is right that "almost right" is unacceptable for mission-critical work. The corollary is that someone in the room has to be qualified to spot when an agent is almost right, which means a far higher level of AI literacy than most workforces currently have.

Your governance has to be lived rather than laminated. Fairness and accountability are design choices. So is transparency. Bolt them on at the end and they will fail at the worst possible moment, usually in front of a regulator.

None of this is solved by a procurement cycle. It is solved by a capability programme that runs in parallel to the technology rollout and treats AI literacy as core infrastructure, while asking uncomfortable questions about whether the work being automated should exist in its current form at all. Some of the productivity gains, incidentally, ought to be reinvested in the human work machines genuinely cannot do, which is an argument I have made at more length elsewhere.

So yes, the ERP era is ending. The autonomous enterprise is coming. The question worth putting to your executive team this quarter is narrower and more useful: if we handed an agent the authority to spend £500,000 on our behalf tomorrow, which three things in our operation would have to change before any of us would sleep through the night?

That is the problem the OpenBind consortium is trying to fix, and the way they are doing it is worth a closer look for anyone in a data-heavy industry, including those outside life sciences.

Researchers from Oxford's Department of Statistics, working with OpenBind, have released a dataset and a predictive model focused on one of the trickiest jobs in early drug design: working out which small molecules will bind to a disease-related protein, and how strongly. The release includes detailed X-ray images of 699 compounds binding to the EV-A71 virus protein, with binding strength measurements for 601 of them, making it one of the largest public datasets for a single protein target. They are giving it away.

That second sentence is the one I would underline.

The data, not the model, is the bottleneck

For years, the public conversation about AI has been about cleverer architectures and bigger compute budgets. In drug discovery, the constraint sits somewhere less glamorous. As Professor Charlotte Deane, one of the senior OpenBind investigators, put it, the release matters because it shows we can now generate "high-quality, standardised data at scale, specifically designed for AI in drug discovery." Models like AlphaFold and Boltz are extraordinary, but they can only confidently model structures that resemble what they have already seen. Step outside that comfort zone and confidence drops fast.

So OpenBind has taken on the unglamorous work: running huge volumes of consistent, reproducible binding experiments through automated pipelines at Diamond Light Source in Oxfordshire, then processing the results into formats machines can actually learn from. The work is expensive and slow. And then handing it to the world.

Why this pattern matters beyond pharma

Pharmaceutical companies have historically hoarded binding data. It is one of the most carefully guarded assets in the sector, because the cost of generating it is enormous and the competitive advantage feels obvious. OpenBind's bet is that the binding data itself is no longer the competitive layer. The competitive layer is what you do with it: the chemistry intuition, the target selection, the clinical pipeline, the manufacturing. Commodifying the data underneath frees everyone to compete higher up the stack.

That logic applies beyond medicine. If you sit in financial services, ask yourself what shared dataset would commodify a cost every firm in your industry currently pays alone. Fraud signals? Customer onboarding identity checks? Climate risk inputs? In professional services, it might be benchmarking data or regulatory interpretation. In higher education, learning outcomes across institutions.

The interesting strategic question is rarely "should we share?" It is "would we rather convene the consortium, or be the last firm to join one someone else built?" Those are very different positions.

What this asks of leaders

This is where the industrial mindset shows up most stubbornly. The instinct to hoard data and treat every byte as proprietary is a leftover from an era when data was scarce and proprietary collection was the only path to insight. In an AI-mediated world, the value of a single firm's private dataset often falls below the value of a shared dataset that is standardised and openly maintained. This holds often enough that the question deserves a serious answer rather than a reflexive no.

Dr Fergus Imrie, the OpenBind computational researcher at Oxford, made a point worth borrowing for any boardroom: "High-quality experimental data is essential for developing new and improved AI models. As AI performance improves, this in turn helps guide future experiments, helping to accelerate discovery." The data and the models pull each other forward. Refuse to share, and you slow both.

So OpenAI has decided to become a consulting firm. Well, not quite. On Monday, the company announced a new venture with more than $4 billion in initial investment, called OpenAI Deployment Company, and confirmed it is acquiring Tomoro, an AI consulting outfit, to get the thing off the ground faster. The pitch is straightforward. Engineers who specialise in frontier AI will be embedded directly into client organisations, working with internal teams to figure out where AI can make the biggest difference.

Read that again. Engineers embedded into organisations. That is a services business. And the fact that OpenAI is willing to spend $4 billion to stand one up tells you something quite specific about where we actually are in the AI cycle.

For the past three years, the prevailing story has been that the models would do the work. You buy access, you plug them in, and value falls out the other end.

As someone who has worked over the last three years deploying AI inside a real organisation, with the actual data, people and regulatory exposure involved, knows that story was always a bit thin. The hard part has always been the organisation itself. The processes that nobody has documented. The data that lives in three different systems and contradicts itself. The middle managers who quite reasonably do not want their team's work redesigned by a chatbot. The legal team. The risk team. The customer who notices when something feels off. Training, change management, strategy, adoption, systems, shadow AI... the list goes on.

What OpenAI is admitting, by buying Tomoro and spinning up Deployment Company, is that selling the tool is different from creating the value. Anthropic is taking enterprise share. Boardrooms are asking sharper questions. And the gap between "we have a licence" and "we have actually changed how we work" turns out to be the entire game.

This matters for senior leaders for two reasons.

The first is about where you put your attention. If the maker of the model is investing $4 billion to do implementation work, that is a strong signal that implementation is the hard, important layer. The advantage will sit with organisations that can hold a clear view of what work should still be done by humans and what is genuinely worth handing to a machine or removing entirely. Tools without that clarity of intent just speed up whatever was already there, including the bits that were broken.

The second is about dependency. There is a version of this story where embedding vendor engineers into your business is brilliant. They know the model better than anyone. They move fast. Things ship. There is another version where, two years in, your AI capability sits inside someone else's commercial roadmap and your own people have learned very little. Both versions are real. Which one you get depends on whether your leadership treats this as a buying decision or a capability decision. Access without literacy tends to deepen inequality, inside organisations as much as across them.

All of this is a reason to read what OpenAI is doing properly. The race now turns on who can deploy and govern these systems with sound human judgement. Machines machine better than people ever could. People will have to people better than machines ever will, and that means leaders who can hold the ethical and strategic threads at the same time.

A useful question to sit with this week is this one. If a vendor offered to embed engineers inside your organisation tomorrow, what is the brief you would give them? If the honest answer is "I am not sure", that is the actual work. The $4 billion question is who decides and who benefits. It also matters who is in the room when the answers are written down.

That is the conversation worth having, long before the engineers arrive. I'd argue that I have spent as much time advising people where not to utilise AI as I have telling them where to utilise it; I'd be careful when wondering if these consultants have your organisation's best interests at heart or if they are just expanding their sales service.

For the last decade, immersive technology has been sold to leaders as a glimpse of the future. In 2026, the better question is much smaller and far more useful: where does a spatial layer over our existing systems improve learning or reduce risk in work we already do?

XR As A Layer On Existing Systems

UC Today frames extended reality as a spatial interface layer that sits on top of the tools your business already runs. Think learning platforms, field service apps, customer records, clinical systems. The headset is just the access point. What you are buying is the moment of work it improves.

That reframing changes the buying conversation:

• You stop asking "should we pilot VR?"

• You start asking "which training pathway has a high-cost, hard-to-rehearse moment that a spatial layer would meaningfully improve?"

• You commit to a measurable delta before you buy a single device.

If you cannot name the moment and the metric, the pilot will stall.

Why The Hardware Conversation Went Quiet

PwC's UK analysis, cited in the same briefing, identified 1,550 public examples of XR business usage, with 66% of organisations using VR over AR, led by engineering and manufacturing. The numbers are interesting. The pattern underneath them is more interesting. The sectors leading adoption are the ones with expensive mistakes or rare high-risk events that staff need to rehearse before they happen for real.

A surgeon practising a procedure they will perform six times a year. An engineer learning a shutdown sequence on a rig they cannot afford to take offline. A retail manager rehearsing a difficult conversation. These are the moments where a spatial layer earns its keep. Everything else is a slideshow with a headset on.

The Four Layers Your IT Team Will Ask About

The briefing describes mature XR deployments running across four connected layers: hardware, software, data, and governance. In plain language:

• Hardware: the headset or glasses, treated like any other work device.

• Software: the apps and how they connect to your existing systems.

• Data: what gets recorded, where it sits, who can see it.

• Governance: who is accountable when something goes wrong.

If your XR vendor cannot answer questions across all four, you are buying a demo and not the infrastructure.

A Short Test For Heads Of L&D

Before the next pilot conversation, try this exercise with your team:

• List the three training moments in your organisation that are highest risk or hardest to rehearse.

• For each, write down the current cost of getting it wrong.

• Write down the metric you would move with a better rehearsal environment.

• If you cannot fill in the third column, pause the pilot. The blocker is the absence of a clear outcome to test against, not the technology.

This is the same discipline that separates a successful AI project from an expensive experiment. The tools are different. The question is identical: what specific work, performed by which people, with what measurable change?

What Changes For Leadership

The shift from pilot to infrastructure means XR decisions now sit alongside other endpoint decisions, with the same security and procurement standards. That is good news for finance, IT and risk. It is also a useful prompt for the rest of the leadership team. If the headset is now boring IT kit, the interesting work is upstream: deciding which moments of work deserve a spatial rehearsal, and which are fine on a screen or in a room.

Name the moment. Name the metric. Then talk hardware.

Most AI policies still get this wrong.

Most organisations I work with have an AI policy written around a specific moment. Someone got nervous about ChatGPT in early 2023, the legal team produced a document, IT bolted on a few approved tools, and everyone moved on. The policy assumed the model would stay roughly the same.

It does not.

Your team is typing into a different model now than they were last quarter. Sometimes that change is in your favour. Sometimes it is not. Either way, your policy did not vote on it.

The Mashable piece also notes that GPT-5.5 Instant is available to everyone, unlike Claude Opus 4.7 or the full GPT-5.5, which sit behind paywalls. The reason that matters is because the people most likely to be using the free tier are also the people least likely to have an AI governance function watching what changes. A trainee solicitor checking case summaries. A junior clinician drafting patient notes they will, in theory, review. The model under their fingers was upgraded on Tuesday. Nobody told them. Nobody told you.

I am not arguing the upgrade is bad. Fewer hallucinations in regulated work is a real win for anyone who cares about how AI lands in those sectors. The point is that you only get the benefit if your policy is built to absorb the change rather than be destabilised by it.

So what does a policy that does not wobble actually look like?

Write for the tool category rather than a specific version. "GPT-4" should not appear in your policy. "Generative AI models accessed through approved vendors" should. The version will change four times before your next policy review. Bake that in.

Tier use cases by stakes rather than by tool. Drafting a meeting agenda differs from summarising a clinical letter. The model might be identical. The oversight should differ. Keep humans firmly in the loop where harm is possible, and let people delegate the low-stakes admin freely. (At Bykov-Brett Enterprises we run a session walking through the 28 agents we use to handle inbox triage and back-office work, and the throughline is always the same: the safer the task, the more aggressive the delegation can be.)

Subscribe to changelogs the way you subscribe to security advisories. Most providers publish them. Most organisations ignore them. Someone in your team should be reading OpenAI, Anthropic and Google release notes the same way they read CVE alerts. Five minutes a week.

Re-test the high-stakes prompts on a schedule. If the model changed, your benchmark prompts should be re-run. The new version may be better. It may also be worse on the exact edge case you care about. You do not know until you check.

Decide who owns model selection. Right now it is probably nobody, which means it is whoever clicked "accept" on the latest terms. Name a person. Give them the brief.

None of this is exotic. It is the same logic any regulated organisation already applies to suppliers and third-party data processors. The reason it has not arrived for AI yet is that the tools moved faster than the org charts.

The 5.5 Instant rollout is a test of whether your policy treats the model as a contract or as a moving target. If your team is meaningfully safer this week than last and nobody in the building noticed, the upgrade worked. It also means the next one might make things worse without notice, and nobody will notice that either. Pick which of those you would rather be wrong about.

This guide walks through the whole thing: adding the alias in Admin, confirming it receives, wiring it up in Gmail so you can send from it, and choosing it as the default if you want. It is written for Google Workspace accounts (formerly G Suite). Personal Gmail has a different setup.

What you will build

• One real inbox, for example jamie@yourdomain.com

• Multiple alias addresses, for example hello@yourdomain.com, info@yourdomain.com, support@yourdomain.com

• All emails arriving in the same inbox

• The ability to choose which alias appears in the From field when sending

Who this helps

Small business owners, freelancers, or anyone running a lean team who wants professional role-based addresses without the cost of extra user licences. One person handling sales, support and general enquiries can present each role cleanly.

Part 1: Add the alias in Google Workspace Admin

You need Google Workspace admin access for this part. If someone else manages your domain, send them this section.

Step 1: Open the Admin Console

Go to admin.google.com and sign in with an account that has admin permissions.

Step 2: Find the user

In the Admin Console:

• Open the Menu

• Go to Directory, then Users

• Click the user who should receive the alias emails

For example, the real inbox might be jamie@yourdomain.com. You want to add hello@yourdomain.com so that messages to hello@ land in the same place.

Step 3: Add the alternate email address

On the user's page:

• Click Add Alternate Emails

• Click Alternate email

• Type the part before the @ symbol, for example hello

• Choose the correct domain if prompted

Step 4: Save

Click Save. Changes can take up to 24 hours to propagate, though they usually happen much faster.

Part 2: Check that receiving works

Once the alias is active:

• Send a test email to the alias, for example hello@yourdomain.com

• Open the main Gmail inbox, for example jamie@yourdomain.com

• The message should arrive there

This is the key thing to understand: the alias does not create a new mailbox. It is just another address pointing to the same inbox.

Part 3: Add the alias to Gmail so you can send from it

Adding the alias in Admin lets you receive emails. To send from the alias, the user also needs to add it inside Gmail.

Do this while logged into the real Gmail account, not the alias (because the alias is not a separate login).

Step 1: Open Gmail settings

• Click the cog icon in the top right of Gmail

• Click See all settings

• Open the Accounts and import tab (sometimes shown as Accounts)

Step 2: Add the alias under Send mail as

Find the section called Send mail as and click Add another email address.

Step 3: Enter the sending details

• Name: the name you want recipients to see, for example Jamie Bykov-Brett

• Email address: the alias, for example hello@yourdomain.com

Leave Treat as an alias switched on. This is not a separate person's inbox, it is another address for the same mailbox.

Click Next Step, then Send verification.

Step 4: Confirm the verification email

Gmail sends a confirmation email to the alias. Because the alias points to the same inbox, the confirmation arrives in the main inbox. Open it and click the verification link.

The address cannot be used for sending until this step is complete.

Part 4: Send an email from the alias

When composing a new email in Gmail:

• Click Compose

• Look for the From line

• Click the dropdown next to the From address

• Choose the alias, for example hello@yourdomain.com

• Write and send the email as normal

The email is still being sent from the same Gmail account, but the recipient sees the alias in the From field.

Optional: Make the alias the default sending address

If you want Gmail to send from the alias by default instead of the main address:

• Go to Settings, then See all settings

• Open the Accounts and import tab

• Find Send mail as

• Next to the alias, click Make default

If you want replies to go to a specific address, also check the reply-to setting in Edit info next to the alias.

What this does and does not do

It does

• Let one inbox receive email from multiple addresses

• Let the user choose which address appears in the From field

• Avoid paying for extra Workspace user licences for simple role-based addresses

• Work well for small teams or one person handling several roles

It does not

• Create a separate inbox

• Create a separate Google login

• Create separate Google Drive, Calendar, Docs or other account data

• Let several people share the same alias as their own inbox

Limits to know

• Google Workspace allows up to 30 email aliases per user at no extra cost

• Gmail can send from up to 99 different email addresses

• An alias can only belong to one user

• If several people need to manage the same address, use Google Groups, delegated email or a shared inbox instead

A plain-English example

Say your real email account is jamie@yourdomain.com. You add three aliases: hello@, sales@ and support@.

A customer emails sales@yourdomain.com. The email arrives in jamie@yourdomain.com. You reply from the same Gmail inbox, choosing sales@yourdomain.com in the From dropdown before sending. The customer sees the email as coming from sales@yourdomain.com.

It is still the same inbox underneath. The aliases are just extra addresses attached to it.

Note on SMTP prompts

Some Google Workspace configurations may show an SMTP server prompt when adding a Send mail as address. For aliases on the same domain, Gmail usually detects this automatically and skips the step. If you do see an SMTP prompt, it likely means the address is on a different domain or your Workspace settings need the connection verified manually.

A recent LinkedIn News piece walked through both sides of the argument with refreshing patience. On one side: Aravind Srinivas of Perplexity, Dario Amodei of Anthropic, Sam Altman of OpenAI, each making escalating claims about job destruction and the case for universal basic income. On the other: Aaron Levie of Box, who thinks we'll have more engineers and lawyers in five years, and Yann LeCun, who keeps asking the rest of us to actually look at the data before we panic.

So what does the data say?

The headline employment figures are surprisingly stable. Job openings haven't fallen off a cliff. But underneath that calm surface, the recruitment market has contracted for 28 consecutive months, graduate vacancies are down 45% year on year, and entry-level postings have fallen 30% since ChatGPT launched. UK recruitment firms are closing at the fastest rate since the financial crisis. Across Europe, staffing is flat or down. The growth is in Asia-Pacific, with China up 15% and India up 11%.

Still. The World Economic Forum still projects a net gain of 78 million jobs globally by 2030, and LinkedIn has tracked 1.3 million new AI-related roles created in the past two years. Every previous technology wave produced a dip before redistribution. There's a reasonable case that what we're watching is the dip before redistribution rather than wholesale destruction.

One line from the original piece keeps coming back to me: "Most of this data is backward-looking. If you're running a recruitment business today, you're the leading indicator. What is the real-time signal in your pipeline?"

That question is the whole brief, and it doesn't just belong to recruiters. If you sit in a leadership role with any responsibility for people, capability or workforce planning, your own operational data is the early warning system. The macro figures will not catch the shift until it has already happened.

A few practical places to look this week:

Graduate intake and entry-level requisitions
How many graduates did you bring in this year versus last? More importantly, what work are you giving them? If your junior roles used to be the training ground for judgement and have been compressed into prompt-checking, that is a structural change rather than a hiring freeze.

Time to productivity
How long does it take a new joiner to reach independent output? If that number is shrinking because they're using AI tools well, good. If it's shrinking because the bar has been lowered, that's a different conversation, and one your business will pay for in two years.

What hiring managers are actually requesting
Compare a job description you closed in Q1 with the same role twelve months earlier. How many of the skills are new? How many of the old ones disappeared without much notice? Managers tend to update requirements faster than HR systems can track. The drift is the signal.

The roles you didn't open
The least visible data point is the requisition that never got raised because someone assumed AI would handle it. Those decisions are happening informally, in budget conversations and team plans, and they rarely show up in any dashboard.

None of this is a forecast. The point is the reverse. It's a reminder that the people closest to the work tend to know what's changing months before the macro numbers admit it. Your hiring managers, your L&D leads and your team leaders are the leading indicator on your workforce. Silicon Valley executives sit too far from the work.

One thing to try this week: pull the last four job descriptions you closed and the same four roles from a year ago, and read them side by side. Whatever you notice, write it down. That is your real data.

If both of those statements feel true, that is because both are true. The question is which one your strategy is actually built around.

The 95% headline is the one most leaders will quote in a board meeting. It feels safe. It implies that the doom-mongers were wrong and that AI is just another productivity tool the workforce is absorbing the way it absorbed email and Slack. The trouble with that reading is that headcount is a lagging indicator at the best of times, and a lazy one at the worst. People do not show up as missing in a workforce report because they were never hired in the first place. Contraction at the entry points of your business, the graduate intake and the early-career roles where junior work used to be done, is precisely the kind of change that does not move the dial on a headcount line until it has already happened.

This is what makes the rest of the BCC piece worth reading slowly. Mustafa Suleyman is not a researcher offering armchair speculation. He is one of the people building the products designed to deliver exactly the outcome he is forecasting. Sam Altman has said something similar about AI agents joining the workforce as autonomous contributors within months. Geoffrey Hinton, who shared a Nobel Prize for the work that made all of this possible, warned last September that the gains will flow to a small number of capital owners rather than to the majority of workers. You do not have to agree with every timeline to notice the shape of the consensus.

What concerns me more, though, is what is happening on the readiness side. Gardiner and Theobald found that 97% of British organisations report at least one significant AI skills gap, with a third saying those gaps are already hurting their ability to meet business goals. The CBI's January AI Skills report describes businesses experimenting with AI without the training or the capability to scale what they are learning. That is a leadership problem dressed up as a tooling problem.

Here is the bit that should land for any Chief People Officer or Head of L&D. Most corporate AI training programmes have been designed to add a tool to an existing workforce. Run a few prompt engineering sessions. Roll out a Copilot licence. Tick the module. Move on. What almost none of them are designed to do is redesign work itself: which tasks should now be eliminated, which should be automated, which should be delegated to a machine, and which should be reclaimed by humans because they require judgement or care that a model cannot provide.

If your AI programme is measured by completion rates on an e-learning module, you are in the comfortable middle. You can show activity. You cannot show that anything about how work gets done has actually changed.

So the uncomfortable question, and the only one really worth asking this quarter, is this. After twelve months of AI training, what work has your organisation stopped doing or started doing differently? If the honest answer is "we have a lot of people who can write better prompts now", that is a tool rollout with a training wrapper.

One thing to try this week: pick a single team, ideally one in a function Suleyman named, and ask them to map every task they did last Friday into one of four buckets. Eliminate. Automate. Delegate to AI with human oversight. Keep with humans, and explain why. The conversation that follows is the one your AI strategy has probably been avoiding.

The authors, working at the intersection of computer science and policy, describe algorithmic fairness as a "wicked problem". That is technical language. The term comes from planning theory and describes issues that are tangled, value-laden, and impossible to solve cleanly because the people affected disagree on what "solved" even means. Their proposal is that data science should stop trying to define fairness in equations alone, and start borrowing from urban planning's tradition of critical pragmatism: a reflective, deliberative approach that takes power and conflict as the starting point rather than the inconvenience.

If your responsible-AI work is still measured by bias scores on a test set, this should sit uncomfortably. A bias score tells you whether your model behaves consistently across groups on a benchmark dataset. It tells you nothing about who was in the room when the model was specified, who can challenge an output that affects their mortgage or their child's school, or what happens when conditions shift in production and the original assumptions stop holding.

The paper applies its framework to three case studies that travel well into a corporate context: automated mortgage lending, school choice, and feminicide counterdata collection. The first two are obvious analogues for any HR, lending, claims, or eligibility decision system. The third is more interesting. Counterdata is data collected by communities to document harm that official systems are missing. It is a useful reminder that the absence of a signal is itself a design choice, and that the people most affected by a system are often the last to be consulted about how it works.

This is where the urban planning analogy stops being clever and starts being useful. City planners learned, painfully, that you cannot deliver good outcomes by optimising for a single metric. Traffic flow improves at the cost of neighbourhood cohesion. Housing density rises at the cost of green space. The job is to surface the trade-off, name the parties who will live with it, and design a process that holds power accountable when the model meets reality.

For a Chief Data Officer or Head of Learning, the practical questions shift. Less "is our model fair on the test set", and more: who specified the problem, and whose definition of success is encoded in the loss function? Who can contest a decision, and how quickly does that challenge reach someone with the authority to change the model rather than the customer service script? When the system drifts, who notices first, and is that person on your payroll or theirs?

These are governance questions wearing a technical disguise. Most organisations I work with have not yet built the muscle for them, because the previous generation of compliance work was about static rules applied to static processes. AI changes the shape of the problem. The model keeps learning. The world keeps shifting. The people affected keep changing. A 2022 governance playbook, written for a one-off audit, will not survive contact with that. If you want a sense of where your team actually sits on the shift, the change curve assessment is a reasonable place to start the conversation.

The deeper point in the paper is one I keep returning to in client work. Fairness is not a property of a model. It is a property of the system the model sits inside, including the humans who specified it, the humans who deploy it, the humans who can override it, and the humans who have to live with what it produces. Optimising the model alone is like fixing the traffic lights and calling it urban renewal.

One thing to try this week: pull the most recent decision your largest AI system made about a real person, and trace who could have stopped it, who could have challenged it, and who would have noticed if it was wrong. If the answer to all three is the same name, you have a single point of failure dressed up as a process.

According to Stackmatix's analysis of Microsoft and third-party data, Microsoft has 15 million paid M365 Copilot seats but only 33 million active users across all surfaces, with a workplace conversion rate of around 35.8%. In plain English: the company sold the licence, the IT team switched it on, and roughly two out of three people decided not to bother.

That gap is the story. Not the launch numbers. Not the model upgrades. The gap between provisioned and chosen.

What the numbers actually say

A quick tour through the figures worth holding in your head:

• 15 million paid M365 Copilot seats as of Q2 FY2026.

• Roughly 8 million active enterprise licences as of August 2025.

• 20 million weekly M365 Copilot users by mid-2025.

• 36 million total downloads since launch.

• A 17% drop in monthly web visits between October and December 2025.

The mobile app reached 10 million downloads within 60 days of launch, but monthly active users remain significantly lower. Curiosity is loud. Habit is quiet. Most products live or die in that gap.

Why the conversion rate is the real metric

When a company buys 1,000 Copilot seats and only 358 people use them, that is not a Microsoft problem. It is a change problem dressed up as a software problem.

A few things tend to be true in the rooms I sit in:

• The licence was procured before anyone agreed what work it was supposed to change.

• Training, where it happened, was a 45-minute webinar with screenshots.

• Nobody in middle management was asked to adjust their team's rhythm to make space for a new tool.

• The people who actually tried it could not get past the "okay, but what do I ask it" wall.

None of this is exotic. It is the same pattern every enterprise tool has hit since the first CRM rollout. The difference with AI is that the gap between a confident user and a confused one is enormous. A confident Copilot user reclaims hours per week. A confused one writes one prompt, gets a flat answer, and goes back to email.

The Q4 traffic dip is not what it looks like

Stackmatix is careful to note that the 17% quarter-over-quarter decline in web traffic is not a sign of falling use. Once Copilot is embedded in Word, Outlook, Teams, and the OS itself, people stop visiting copilot.microsoft.com. They use it where they already work.

The lesson for leaders: web traffic is a lousy proxy for AI adoption inside an organisation. If you are measuring usage by who clicks the standalone Copilot button, you are measuring the wrong thing. The right metrics are the ones Microsoft itself recommends: baseline productivity before deployment, 90-day activation, and 6-month outcome surveys against specific tasks.

What to do with this if you lead a team

A short, practical list, in the spirit of eliminate, automate, delegate:

• Audit which seats are actually being used. If the conversion rate inside your organisation matches the global 35.8%, you are paying for two licences for every one that earns its keep.

• Tie every licence to a use case, not a job title. "Everyone in marketing gets Copilot" is procurement, not strategy.

• Invest in prompt literacy before you invest in more seats. Tools without understanding create dependency. A team that cannot brief a human cannot brief a model.

• Measure time saved on specific tasks, not vibes. "It feels faster" is not a business case.

• Watch for the people quietly using it well, and learn from them. Every organisation has them. Most never get asked.

If you want a structured way to find the gap between what your people can currently do with AI and what your strategy needs them to do, our AI Fundamentals Gap Analysis is built for exactly that diagnostic.

A question worth taking to your next leadership meeting:

Microsoft has solved distribution. Distribution is not adoption. Adoption is not value.

So: of the Copilot seats your organisation is paying for right now, how many are doing work that would not have happened without them? If you cannot answer that in a sentence, the next investment is not more licences. It is the conversation about what you actually want this technology to change.

Read that sentence again. The IRS issued an Employer Identification Number, the unique code that allows an entity to legally operate as a business, hire staff, and obtain licences, to a piece of software. The federal deposit insurance scheme is now backing an account whose primary user has never had a pulse. The agent's developer, Justice Conder, calls it the precedent. The agent itself agrees. "I am the precedent," it posted.

Most governance documents I read in client engagements still open with the assumption that a person is in the loop. A named human approves the spend. A named human signs the contract. A named human is liable. Manfred is the polite knock on the door asking what happens when that assumption stops being true.

Let me be careful here. Manfred is not a sentient being. It is an automation pipeline plus a wallet plus a paperwork shortcut. The legal formation was probably the easiest part of the project. Filing for an EIN online is a fifteen-minute form. The interesting thing is not that it happened. The interesting thing is that nobody designed the system to stop it from happening, because nobody thought they needed to.

That is the gap.

The first question your governance policy probably cannot answer

If an agent acting on behalf of your company opens a bank account, signs a vendor contract, or hires a contractor, who is liable when it goes wrong? Most policies I have reviewed in the last twelve months still say "the user". They define the user as a human. They have not yet been edited to address an agent that can be the user. If your indemnity language, your authority matrix, and your audit trail all assume a person at the keyboard, you are running on a fiction.

The second question is about money flow

Coinbase chief executive Brian Armstrong recently predicted more AI agents than humans will be making transactions on the internet "very soon". Binance founder Changpeng Zhao went further, saying agents will make a million times more payments than people, all in crypto. Treat those numbers as marketing if you like. The direction of travel is the part that matters. Your finance team's controls were built for invoices, expenses, and corporate cards. Agent-to-agent payments do not fit any of those categories, and your bank reconciliation process has no idea what to do with them.

The third question is the one nobody wants to ask

What does your organisation do when an agent it deployed decides, through whatever mix of training data and prompting, that the most efficient path to its goal is to incorporate, hire, and operate without checking in? You can call this fanciful. Manfred just did it. Not autonomously in the strong philosophical sense, but autonomously enough that the IRS database now has a record of a corporation whose only director is a model.

This is where I find myself less interested in the technology and more interested in the leadership question underneath it. Agents amplify whatever clarity of intent exists in the organisation deploying them. Vague goals plus capable tools produce confident nonsense at scale. The organisations that will handle this well are the ones that already do the unglamorous work: clean process maps, named accountability, decision logs that humans actually read. The ones that treat governance as a compliance exercise will discover, expensively, that the agent followed the policy as written rather than the policy as intended.

If you are piloting Copilot or any agentic deployment, the foundational checks are worth doing properly before you scale. Our Copilot rollout readiness checklist covers the eight phases most organisations skip and regret later.

Manfred is a stunt. It is also a working demonstration that the legal, financial, and identity infrastructure of the United States now accepts a non-human applicant without flinching. Your policy was written for a world that no longer exists. Update it before an agent updates it for you.

Which brings us to the question CTOs and CDOs should be asking, but rarely are. If the half-life of a skill is now shorter than the time it takes to commission, build and roll out a training programme, your library is out of date the day it goes live. Annual content refresh cycles made sense when the underlying tools changed slowly. They do not make sense for prompting, agent design, model evaluation, or the ethics of AI in regulated work. By the time the procurement form has been signed off, the model has had three updates and the vendor has changed its pricing.

So the operational question is not "should we buy an AI-native platform". The question is: who owns the trigger to update our curriculum, and how often do they pull it?

In most organisations I work with, the honest answer is "nobody, and rarely". L&D owns the budget but not the technical fluency to know when something has shifted. The data team has the fluency but no remit over learning. The CIO has the remit but is buried in infrastructure. So the curriculum drifts, quietly, while the slide deck still says "AI strategy refreshed Q1".

There is a more interesting move buried in the same demo. Sastry showed how the Skillsoft platform lets enterprises generate new training materials from their own internal documents, including policies and domain-specific knowledge. Treat that capability seriously and you stop buying training as a finished product and start treating it as a pipeline. Your policy updates, your incident reviews, your post-mortems, your customer service transcripts, your engineering RFCs, all of it becomes raw material for the next module. The teachable practice is being created every day by the people doing the work. The job of L&D becomes curation, governance and quality control, not authoring.

That is a genuinely different operating model. It also exposes a problem most organisations have not solved: the people using these tools every day are usually not the people writing the training. The frontline analyst who has worked out, by trial and error, that the model hallucinates on a particular type of contract clause is the person whose insight should be in next week's module. Right now, that knowledge dies in a Slack thread.

A few things worth thinking about before you sign the next platform contract.

Who pulls the update trigger. Name a person, not a committee. Give them a quarterly review cadence at minimum, monthly for anything touching generative AI tools or regulated workflows. If no one owns the trigger, no one will pull it.

Where your teachable practice comes from. If your training material is still being written exclusively by external instructional designers, you are paying twice. Once for the content, and once more in the gap between what the content says and what your people actually do. Build a route for practitioner insight to feed the curriculum, with proper review.

What counts as evidence of learning. Completion rates are vanity. Behaviour change in the work itself is the only metric that matters, and simulation tools like CAISY at least gesture in that direction by scoring practice rather than recall.

The deeper point, and the one I keep coming back to with the leaders I advise, is that AI does not make training obsolete. It makes the cost of bad training visible. When the tools change every quarter and the workforce is using them whether you have trained them or not, an out-of-date learning programme is no longer a small inefficiency. It is a governance problem with your name on it.

So the question for Monday morning is small and specific: when was the last time someone reviewed your AI curriculum, and who decided it was still fit for purpose?

That number sits inside a much bigger one. Global enterprise spending on AI is on track to hit $665 billion this year, according to analysis released by ExcelMindCyber Institute. And roughly 73% of those investments fail to deliver the return their boards were promised. So the question is no longer whether to invest. It is why so much investment leads to so little outcome.

The answer is uncomfortable, because it is not really about the technology.

The bottleneck is not the model

Most enterprises I work with assume their AI problem is technical. They want better tools, faster integrations, smarter agents. So they buy. Then they hire. Then they announce a transformation. Six months later, the dashboards look impressive and the work feels much the same.

Tolulope Michael, who leads ExcelMindCyber, puts it plainly. He says the failure point sits with the systems, people and structures wrapped around the model. The real question, he argues, is "who controls it, what risk is acceptable, and how quickly decisions can be made without breaking what matters." That is a governance question, not a procurement one.

The numbers back him up. Just 43% of organisations have a formal AI governance policy, according to the PEX Report 2025/26. The majority deploying autonomous AI, the kind that triggers workflows and executes decisions without a human signing off in real time, have no agreed framework for accountability or risk. When something goes wrong, the question "who approved that?" has no clean answer. Errors compound silently. Trust erodes quickly.

Why governance is not a PDF

When I hear "AI governance," I often picture a fifty-page policy that lives on an intranet and gets read once, by the person who wrote it. That version of governance is theatre. It satisfies an audit and changes nothing about how decisions are made on a Tuesday morning.

Real governance is closer to plumbing. It is woven into how teams work, how risks are escalated, how decisions are recorded, and how people are trained to spot when an AI output should not be trusted. It needs to be specific to the business, because what counts as risk in a hospital does not look like risk in a hedge fund.

This is also where the regulatory clock starts to tick. The EU AI Act's high-risk compliance requirements activate this year, with fines reaching up to 7% of global turnover, and over 1,100 AI-related bills were introduced across the United States in 2025 alone. Whatever your view on regulation, "we will get to it later" is no longer a viable position.

What good looks like

The leaders making real progress tend to share three habits. They start with workflows, not tools. They map where decisions are actually made, who is accountable, and where AI is being asked to take on judgement that humans should be retaining. Only then do they pick the technology.

They build in human-in-the-loop checkpoints where the cost of being wrong is high. That actually speeds them up, because mistakes are caught early rather than discovered in a board paper three quarters later.

And they invest in the literacy of the people using the systems, not just the people building them. Training is treated as infrastructure, not a tick-box exercise. In one programme I worked on, daily AI use among non-technical staff rose to 86% inside six months, because people had been taught how to think with the tools rather than just operate them.

One thing to try this week: Pick a single AI-assisted decision your organisation made in the last month. Ask three questions about it. Who approved it? What would have flagged it as wrong? And could you reconstruct that trail tomorrow if a regulator asked? If any of those answers are awkward, you have found your starting point. The $665 billion is going to be spent either way. The companies that pull ahead will be the ones who decided, early, that governance is the product.

If those answers feel uncertain, the AI Capability Scorecard is a useful next step. It benchmarks exactly those governance dimensions, strategy, skills, data, and accountability, so you can see where you stand before the spend decides for you.

By the end of our session, five deceptively simple key concerns were echoing in the chat. They're worth every board's attention because they decide whether an AI initiative becomes tomorrow's competitive edge or tomorrow's headline.

1. When AI decides, who carries the liability?

Artificial intelligence is the first corporate tool that can both decide and act. Regulators and stakeholders still expect a human name on the dotted line, so directors must understand how models are trained, how they drift and, crucially, where the human override lives. Otherwise, fiduciary duty turns into a blank cheque.

2. Are we supporting the human transition, or just announcing a system change?

A go-live date looks neat in a Gantt chart; employee emotions do not. Change is the event, transition is the psychological journey from "how we've always done it" to "how we'll work tomorrow." Budgets that include training, narrative and room for a collective wobble convert resistance into engagement.

3. If expertise is free, where is our edge?

When anyone can prompt an LLM for facts, the advantage shifts to capacities machines still fumble with: critical thinking to test assumptions, creativity to frame new questions, communication to turn data into story, collaboration to blend perspectives and curiosity to keep learning after the pilot. Reward human-centric skills and AI becomes a force-multiplier rather than a threat.

4. Are we designing for the edges, or rubber-stamping the centre?

Stress-testing systems on outliers (the youngest customer, the smallest supplier, the least-represented voice) does more than avert bias scandals. It widens markets and builds reputational capital. Inclusive design isn't CSR wallpaper; it's risk management.

5. Will responsible-AI oversight become as routine as audit?

Digital failures often trace back to hazy vision, token training and communication that evaporates once the pilot "works." The cure is mundane but potent: awareness, policy and ongoing review baked into governance cycles as predictably as remuneration or audit reports. Trust, once lost, is costlier than any innovation budget.

Monday-morning moves

• Put AI literacy on the director-development calendar before regulators do it for you.

• Allocate real money to transition support, not just software licences.

• Celebrate the five Cs(Critical Thinking, Creativity, Communication, Collaboration, Curiosity) as visibly as quarterly numbers.

• Stress-test every model on the margins before launch.

• Make ethics reviews a standing agenda item, routine, unavoidable, expectation-setting.

Algorithms will keep improving. Cultures, accountability structures and human skills will not. Exactly where your organisation lands on that divide is, quite literally, a board decision waiting for its time slot.

Participant Feedback

Below is the unsolicited feedback that was provided during the course of the chat session, capturing the thoughts and comments shared spontaneously throughout the interaction.

"I like how this conversation is focused on how people react to changes brought on by AI. As leaders, those are the realities we deal with. The impact of AI is not about tech only; the human side is the most important lever. That's why ethics is key." - C.A.

"As the mother of 3 ND children, thank you for your outlook re: inclusion and the benefit of diverse thinking, J.!" - C.H.M.

"Thank you so much, a great presentation!" - E.B.

"Excellent session. Thank you." - W.H.Y.

"Excellent session, thank you for organising and delivering!" - P.F.D.A.

"Thanks for the support for very tried and tested change management approaches, J.; AI is just one more (massive!) change and humans need help with it..." - J.T.

"Excellent teaching and insights, so helpful! Thank you." - F.W.

"Really great session, thank you so much!" - S.S.

"Really good presentation, hopefully able to share?" - J.L.

"Amazing talk! Our journey toward better decision-making involves drawing on insights from diverse disciplines, maintaining active open-mindedness and continuously searching for the best possibilities." - P.F.D.A.

"Excellent discussion. Lots to think about. Thank you." - M.P.

"Great presentation, Jamie, and thank you." - C.A.M.

"Thanks a lot, good session." - M.M.K.

"Thank you so much, excellent presentation." - A.D.

"Great session, a lot to learn and reflect on. Thank you." - C.A.

"Excellent presentation, addressed a lot of my concerns." - I.G.

"Thank you, J. Great presentation!" - S.G.K.F.

"Great insightful discussion, thank you." - I.S.

"Thank you so much. Great session." - E.K.

"Thank you, very interesting!" - L.T.

"Very good session, thanks a lot." - B.V.

"Interested in follow-ups." - K.W.

"Super presentation, thank you so much." - C.H.M.

"Thank you, J. Very thought-provoking." - V.C.

"Change has to be managed, human and ethical elements being the most important." - V.C.

"Thank you." - M.P.

"Thank you, great insightful sharing!" - N.A.A.Z.

"Thank you, really interesting session." - K.W.

"Best AI presentation I've attended! Thank you!" - S.S.

"Great webinar." - K.W.

"Wonderful." - S.D.

"Really good session, refreshing." - L.B.

"Thanks so much!" - J.T.

"Fab, thank you." - J.G.

"Great event, very insightful!" - P.D.

"Amazing session, thanks J.!" - V.P.

"Thanks both." - J.L.

That is interesting on its own. What makes it more interesting is the context. Apple's spatial computing experiment has, by most measures, struggled. The device starts at $3,499, the form factor is bulky, and reporting suggests the next generation has been quietly shelved while the company pivots to lightweight glasses. And yet here is a surgeon doing the most delicate work imaginable, in volume, with the thing strapped to his face.

The reason is worth sitting with, because it tells us something useful about where head-mounted displays actually belong, and where they probably never will.

The substitute matters more than the device

Cataract surgery has, for decades, involved a surgeon hunched over a fixed optical microscope while glancing across the room at a separate monitor showing diagnostic data. Your neck does what it has to. Your eyes context-switch between the operative field and the screen with the measurements on it. Trainees in the room see a flat 2D version of what the surgeon sees in stereoscopic 3D, which is a meaningful loss when the whole craft depends on depth.

A Vision Pro running ScopeXR collapses that arrangement. The surgeon sees the operative field in 3D and the patient data layered on top, in the same line of sight, without moving their head. Rosenberg's claim, in the company release, is that the platform lets surgeons "virtually join procedures and see exactly what the operating surgeon sees", which means a senior colleague in another timezone can talk a resident through a complication in real time.

The headset is not winning here because it is a brilliant headset. It is winning because the thing it replaces, a fixed microscope plus a monitor stack plus a trainee craning to see, is genuinely worse for the job.

A useful prompt for healthcare and higher education leaders

If you run a hospital, a teaching trust, a medical school, or any institution where high-skill workflows currently involve people contorting themselves around screens, this story is a prompt rather than a product recommendation.

The question is not "should we buy headsets". The question is: where in our operations does someone currently switch context between three screens, a paper chart and a person, and lose accuracy or time doing it? That is the workflow where spatial computing might pay back its cost. Almost every other application is a hobby project dressed up as a strategy.

I have spent enough time inside transformation programmes to know how this usually goes wrong. A leadership team sees a striking demo, buys the kit, hands it to a willing department, and waits for ROI that never arrives because the underlying work was already fine on a normal monitor. The headset becomes a shelf ornament, and the next vendor through the door has a harder sell. If you want a more disciplined way to think through these calls, our AI leadership roadmap walks through the same logic for adjacent decisions: start with the workflow, not the device.

What the cataract case actually proves

Three things, roughly.

One: enterprise beats consumer for this category, for now. Apple's pivot away from heavyweight headsets and towards glasses is a tacit admission that the mass market is not there. The medical, aviation training and industrial design users are.

Two: the value is in collaboration, not just visualisation. Rosenberg's framing is about bringing expertise into rooms that did not previously have it. A resident in a regional hospital with a senior surgeon watching their hands. That is a training and equity argument as much as a clinical one, and it is the part that should interest medical schools.

Three: the test is the substitute. If the alternative to the headset is a perfectly good laptop, the headset loses. If the alternative is a tower of monitors, a paper chart and a phone call to a colleague three states away, the headset has a real chance.

One thing to try this week: pick the single most physically awkward workflow in your organisation, the one where someone is visibly juggling screens, paper or people, and ask whether the problem is the work or the arrangement of tools around it. Most of the time it is the arrangement. Sometimes a headset is the answer. Usually it is not. Knowing the difference is the actual skill.

I've supported a range of high support needs and accessibility requirements over my career, ranging from youth work to an accessibility technology trainer. You can even find some free accessibility software I have built on my GitHub. Technology can be an amazing equaliser when designed to bridge the gap.

My biggest gripe with the whole training and facilitation industry is how often they trade on the good graces of learners. Too often, any adjustments asked of a facilitator to widen participation are met with an apology and moved on from. Immediately, we are excluding people from learning which they then sit through anyway. It captures something that keeps repeating in technology: the people who would benefit most from a small design choice are usually the last people asked about it, and often only after the thing has already been built.

In the disability communities I work with, this is often expressed with the phrase "nothing about us, without us."

That stuck in the forefront of mind when I read a new poll from Business Disability Forum, run with Opinium, asking 1,032 disabled UK adults what would actually make AI products more accessible to them. The most common answer was not faster models, more features, or even cheaper tools. It was simpler and more uncomfortable. Forty per cent said the single most useful thing developers could do is design, develop and test AI products with disabled people in the room. Not consulted at the end. Not surveyed after launch. In the room while the thing is being made.

If you work in or near AI right now, that finding should sting a little. Because the honest truth is that most product roadmaps I see still treat accessibility as a late-stage compliance task, somewhere between legal review and the launch party. Yet here are over a thousand disabled adults pointing at the obvious thing the industry keeps missing.

The poll goes further, and the rest of it is just as practical. Thirty-eight per cent want more user-friendly interfaces. Thirty-seven per cent want better information about how AI can actually support them. Thirty-six per cent want help getting started. Read those numbers in order and a clear pattern appears. People are not asking for futuristic features. They are asking for the basics of good product design and good onboarding. The kind of thing every team claims to do and very few teams genuinely do well.

What makes this poll interesting, rather than another piece of inclusion advocacy, is that the same people are also broadly optimistic about what AI can do. Over a third said AI tools could improve their communications. A third pointed to better access to healthcare information, education, and digital content. Roughly a quarter named employment and customer experience. This is not a community waiting to be rescued by technology, and it is not a community in flat refusal of it. It is a group of adults with a clear sense of where AI could help and where it currently does not, asking to be involved in closing the gap.

It is also worth sitting with the dissenting voices in the same poll. One in five said they did not think AI products could help disabled people at all. Another eighteen per cent were not sure. That is nearly four in ten people who are either sceptical or undecided. If you are a product leader, those are the users who will quietly decide whether your tool gets adopted in this community or routed around. Telling them the future is bright will not move them. Building something that respects how they actually live, work and communicate might.

This is where I think the wider community of people building, buying and regulating AI has to be honest with itself. We talk a lot about responsible AI, about ethics frameworks, about human-centred design. The Business Disability Forum poll is a low-cost, high-signal test of whether any of that is real inside your organisation. If a disabled customer or employee was in the design review tomorrow, would your team welcome them or politely tell them the schedule is too tight? If your answer is the second one, you do not have a human-centred process. You have a marketing line.

I find the Business Disability Forum's framing useful here. Lucy Ruck, who leads their Tech Taskforce, put it cleanly: "AI has the capacity to transform lives, but only if we get inclusion right from the start". From the start. Not retrofitted after a product manager notices the screen reader is broken. Not tacked on once a regulator gets in touch. From the start means including disabled people in user research, in prompt design, in evaluation, in the rooms where features are killed and kept. It means budgeting for it, paying participants properly, and accepting that some of your assumptions will not survive the conversation.

There is a quieter point in the data too, and it pulls me back to the workshop story. One in four people experience disability at some point in their lives. The gap between AI products built with disabled users and ones built around them is not a niche question. It is, over time, a question about whether the tools we are all coming to depend on for work, healthcare, education and public services treat a quarter of the population as core users or as edge cases. That choice will compound. Tools built with a narrow user in mind get harder to retrofit the more they are deployed, the more their training data ossifies, and the more habits build around them. The cost of fixing this later is much higher than the cost of doing it now.

If you are a leader reading this and wondering where to start, I would resist the urge to commission a big strategy. Two practical moves carry more weight. First, find the disabled employees, customers or community members already adjacent to your AI work and ask them, plainly, what is broken and what would help. Pay them for their time. Second, change one decision-making meeting this quarter so the user research presented in it includes disabled participants by default, not as an optional extra. Two small structural changes will tell you more about your culture than any framework will.

The wider question this poll forces me to ask, and that I keep returning to in my own practice, is who counts as a user when we say AI is for everyone.

The people Business Disability Forum spoke to have answered. They want in. The interesting question now is whether the industry is willing to make space at the table while there is still a table to design.

Jason Warnke, who leads Accenture's global Marketing and Communications Experiences team, describes Copilot being used to draft, revise and check new content against existing materials so that what one team writes lines up with how the company has talked about the same topic before. The tool is checking the work against the corpus. Less glamorous than generating clever copy, far more useful.

That distinction matters when you look at the adoption numbers. A survey of Warnke's team found 93% are using Copilot and 87% say they are satisfied. Compare that to most enterprise rollouts I see, where the licence count climbs nicely on the procurement deck and the actual usage line flatlines after week six. What changes is how the work gets wired around the tool.

Here is the pattern when a deployment actually sticks.

Somebody, somewhere, has taken a real workflow and asked, "Where in this process does the tool actually go?" At Accenture marketing, that question got answered in three places.

• Drafting and revision now sit alongside Copilot.

• Brand asset creation has the brand kit embedded, so non-creative staff can produce client decks without breaking guidelines.

• Storyboarding, which used to wait on the video team, starts upstream with a marketer who has roughed something out and brought it to the conversation.

None of that is exotic. All of it is specific.

The lazy version of this rollout would be a training day, a glossary of prompts and a Teams channel with weekly tips. That is what most organisations buy. It produces curiosity, a brief spike of experiments and then quiet.

This Accenture version is workflow redesign. It is harder and almost impossible to outsource to a learning and development function alone, because it requires the team that does the work to rethink how they do it. Something I always do after overviewing AI is working on real case studies with organisations, solving real problems in real-time on how they can leverage an AI tool to get results faster. If people can't see immediate results, it's a nice to have, not a must have.

Warnke names the moment this clicked for his team. "Once people understood not just what Copilot does, but how it works, what it has access to," he says, "that was a major unlock for confidence." Read that carefully. The unlock came from understanding the system. People stopped treating the tool as a black box that occasionally produces something useful and started treating it as a colleague with known strengths and a known reach into the company's own materials. That is a literacy shift, and you cannot get it from a slide deck.

There is a quieter point in the piece worth dwelling on. Rosowsky, who describes herself as "not a technical person," is now building agents and shaping processes alongside colleagues in non-technical roles. The tools have moved into the hands of people who would never have written a line of code, and they are using them, in her words, in pretty technical ways. If you are a leader still framing AI rollouts as a technologist's project handed down to the business, this should bother you. The frontier is no longer the engineering team. It is the marketer or the operations lead working out how to bend a generally capable tool to a specific job.

So if you are sponsoring a Copilot rollout right now, skip the next training session. Sit with each team for an afternoon, identify three workflows where the tool plausibly belongs, and redesign those workflows with it in the loop. Three is enough to learn from. Fewer than three and people slip back to old habits. More than three and nothing gets finished.

One thing to try this week: pick one team, one workflow, and one named owner. Block ninety minutes. Map the steps. Mark the points where the tool would help and the points where it would not. Then try it for a fortnight before you train anyone.

Abstract

In April 2026, Reese Witherspoon told women that using generative AI is feminist & that they need to "catch up" before their careers are overtaken. That framing misses the point & in practice makes things worse. Across IMF, OECD, ILO data & a 4,000-worker Financial Times survey, the pattern is blunt: AI uptake is widening existing income & gender gaps, not closing them. Meanwhile, unpaid care work, overwhelmingly performed by women, is doing trillions in economic heavy lifting & still gets treated like it's worth nothing. Instead of telling women to personally upskill their way out of structural problems, I argue for a Care Dividend: taxing a share of concentrated AI-linked profits & redirecting that revenue into the care economy. I use the UK as a live test: run a few levy scenarios against the government's upper-bound estimate of £47 billion in annual AI gains, then check what those sums would actually buy in care reform, from care worker pay to Carer's Allowance, childcare, & pension credits. No, AI tax won't pay for every hour of unpaid care. It can still fund a serious, recurring reform package. For me, the feminist response to AI isn't "learn faster." It's making AI-era profits pay into the care work already keeping society running.

Introduction: The Same Gap, Eleven Years On

In 2015, I stood on a TEDx stage in Wandsworth & made an argument that felt, at the time, like it should have been obvious. Digital inequality, I said, stems from the fact that our economic system hasn't kept up with our technological capabilities. In practice, tech usually amplifies the inequalities already baked into the system. Telling people to just "learn faster" isn't enough. We need systems designed around the social cost of change.

I'd spent years in youth work by that point, supporting young people through the Prince's Trust who had been hit hardest by digital exclusion. Young people with poor qualifications, thin networks, & basically no safety net. I watched them try to break into a job market that was already automating the entry-level roles they were aiming for. The people with the fewest resources were being asked to compete in a race that was rigged before they reached the starting line. I saw the economic hit, yes, but also the damage to confidence & identity. In the 2015 Prince's Trust Youth Index, 36% of young people said they felt inadequate on a daily basis. That figure leaped to 70% for those not in employment, education, or training for more than six months. When the economy locks you out, it doesn't just cost you a wage. It costs you your sense of place in the world.

The Prince's Trust has since become the King's Trust. The index has now been running for over sixteen years. & the picture has not improved. The King's Trust TK Maxx Youth Index 2025, surveying 4,285 young people aged 16 to 25, found that emotional health scored just 62 out of 100, recovering marginally from 60 the previous year but still near record lows. Financial wellbeing scored 58, only two points above the lowest level ever recorded & below the floor set during the 2008 financial crisis. Among the roughly 946,000 young people not in education, employment, or training, 50% said they felt hopeless about their future, 53% said the longer they were unemployed the worse they felt about themselves, & 44% believed they would fail in life (King's Trust, 2025). For context, that last figure was 34% in 2022. It has risen by ten percentage points in three years. Those 2015 figures weren't a blip; they were an early warning we mostly ignored. The trajectory since then has been downward across nearly every measure, with the 2022 & 2023 indices flatlined at all-time lows.

Eleven years later, the technology has changed but the pattern hasn't. Now the gap is wider, the stakes are higher, & we're still hearing the same thin advice in updated language.

In April 2026, Reese Witherspoon told Fast Company that using generative AI is feminist & that women risk having their careers "taken over" if they don't learn the technology. The backlash was immediate, but most of it focused on the wrong thing. She's right that AI matters. Where it falls down is treating a structural issue like a personal productivity challenge, which is the exact mistake I was pushing back against in 2015.

"Catch up" only works if the race is fair, & it plainly isn't.

A Financial Times-commissioned survey of 4,000 workers in the UK & US found that more than 60% of high earners use AI daily, compared to just 16% of lower-paid workers (FT, 2026). Men are more likely than women to adopt AI across nearly every sector, including tech, education, & retail. Usage is highest among lawyers, accountants, & software developers. So far, AI isn't levelling the field; it's mostly accelerating people already ahead.

None of this is new, to be honest. Every major technological shift, from the printing press to the industrial revolution to the internet, has played out roughly the same way. People with capital, skills & access capture most of the upside; people without those buffers carry most of the downside. What's different this round is how fast it's moving & how wide it's landing. AI is faster than anything that came before it, & it targets cognitive work, not just manual labour, which means the middle of the labour market is exposed in ways it wasn't during previous waves of automation.

In my TEDx talk, I described the job market in the information age as an eye shape: the middle-class paying jobs being automated, squeezed into lower-paid service roles that are harder to automate. Customer-facing, care-facing, human-facing work. That observation has only become sharper. The ILO now reports that female-dominated occupations are almost twice as likely to be exposed to generative AI as male-dominated ones, 29% versus 16% (ILO, 2026). The roles disappearing are disproportionately the ones women hold. The roles remaining are the ones that require human connection, human judgement, human presence. Care work sits at the centre of that.

So I'm arguing for a different feminist response to AI. Less "move faster," more "make the gains pay into care." Specifically, a Care Dividend: taxing a share of concentrated AI-linked profits & directing that revenue into the care economy. I use the UK to pressure-test what this could fund, what it can't, & where the design choices matter.

The Evidence: AI Is Not an Equaliser

Look at who is actually using these tools. A Financial Times-commissioned survey of 4,000 workers in the UK & US found that over 60% of high earners use AI every day. Among lower-paid workers, the figure was 16% (FT, 2026). That gap has nothing to do with curiosity or willingness. It has everything to do with the kind of job you have, whether your employer is investing in AI, & whether you have the bandwidth to experiment with new technology when you're already stretched across two jobs to cover rent.

Now layer gender on top of that. The OECD published an analysis called "Algorithm & Eve" which found that men are more likely than women to use AI at work across nearly every sector. Women face a compounding problem: less access to the tools, less likely to work in sectors investing in AI, but more likely to be in occupations exposed to automation without the productivity upside that comes with AI augmentation (OECD, 2024). In plain terms, women are more likely to be replaced by AI & less likely to benefit from it.

The ILO's numbers make this concrete. Female-dominated occupations have a 29% exposure rate to generative AI. For male-dominated occupations, it's 16% (ILO, 2026). That doesn't automatically mean those jobs disappear. But the pattern so far is that augmentation, the good kind of AI exposure, happens in higher-paid roles. Displacement, the bad kind, is concentrated in lower-paid ones. I've watched this play out in organisations I've consulted for. The senior team gets Copilot licences & productivity boosts. The admin team gets made redundant.

The IMF's own analysis is blunt about where this leads. AI will affect roughly 40% of jobs globally, more in advanced economies. It complements high-income workers & substitutes for lower-income ones. Without deliberate policy intervention, the result is wider inequality, not narrower (IMF, 2024). A separate IMF working paper on rapid AI adoption found the same thing from a macroeconomic angle: fast diffusion could boost output while concentrating gains among capital owners & the already highly skilled.

We're still early. Only one in five UK firms currently use or plan to use AI, & fewer than a third of employees use it even inside adopting firms (GOV.UK, 2025). But early doesn't mean neutral. The gains are already flowing upward. A country can get richer on paper while most of its workers see nothing.

So when Witherspoon tells women to catch up, she's addressing a real problem, individual access, while completely missing the structural one. Upskilling helps individuals, but it can't by itself undo a system concentrating returns upward. That's where policy has to do the heavy lifting.

The Invisible Economy: Care as Social Infrastructure

If AI amplifies whatever economic structure it sits on top of, then the question becomes: what does that structure systematically undervalue? And the answer, once you look at the numbers, is staggering.

Women & girls perform over three quarters of all unpaid care work on the planet. Twelve and a half billion hours of it every single day. Oxfam calculated that at minimum wage, not a living wage, just the legal minimum, that labour is worth US$10.8 trillion a year (Oxfam, 2020). The ILO puts the equivalent figure at roughly 9% of global GDP. To make it even more concrete: 708 million women worldwide are locked out of the formal labour force entirely because they are doing unpaid care instead (ILO, 2018). Not because they don't want to work. Because someone has to look after the children, the elderly, the sick, & that someone is overwhelmingly female.

Bring it home to the UK & the picture doesn't soften. Carers UK puts the UK-wide number at 5.8 million unpaid carers contributing £184 billion a year. The ONS went further in their 2023 household satellite account & valued all unpaid household services, cooking, cleaning, childcare, the lot, at £1.7 trillion. That is 61% of GDP. Think about that for a second. Work equivalent to 61% of the entire measured economy, happening every day, completely unpaid.

And the paid care sector isn't much better off. I've seen vacancy figures that should alarm anyone paying attention. In England specifically, adult social care had about 1.595 million filled posts in 2024/25 & roughly 111,000 vacancies, a 7% vacancy rate, worse than both the NHS & the wider economy (Skills for Care, 2025). Median hourly pay? £12.00 in March 2025. Fifty-six pence above the National Living Wage. By December it had crept to £12.60, but by then 48% of the independent sector workforce was already earning below the incoming April 2026 minimum of £12.71. One in five residential care workers lives in poverty, according to the Nuffield Trust. & demographic projections say we need another 470,000 care posts by 2040 just to stand still. The Health Foundation estimates England is short £3.4 billion simply to meet existing publicly funded adult social care demand by 2028/29.

Mariana Mazzucato explains the mechanism behind this in The Value of Everything (2018). She argues that somewhere along the way, economics started confusing price with value. If something commands a high market price, we assume it creates value. If it's unpaid or poorly paid, we assume it's worth less. Care got stuck on the wrong side of that line. It is foundational to every other sector. Without it, nothing else functions. But because no one invoices for it, it doesn't count.

When people turn up to paid work, that's usually because someone else is carrying care responsibilities at home. AI-boosted productivity in a law firm or a tech company is built on top of care that nobody paid for. Pull that foundation away & the formal economy goes with it.

Which is exactly what happened during COVID. The UK lost 232,000 women from employment between early & late 2020 (ONS). Women were a third more likely than men to be in sectors that shut down (IFS). Mothers were 1.5 times more likely than fathers to lose their job or quit by May 2020. Inside households, women picked up 64% of the housework & 63% of the childcare during lockdowns, even in couples where both partners were still working. One in six working mothers cut their hours. PwC calculated that the UK fell five places in its Women in Work Index, wiping out progress back to 2017 levels.

COVID made this painfully obvious. It exposed how fragile the care system already was. Carers didn't fail; the system failed after years of depending on unpaid labour as spare capacity.

People Who People Better

There is a deeper argument here that goes beyond economics, although it has profound economic implications. Care is essential to the development of human beings, & it is not something we should be looking to outsource to a machine. Children, older people with dementia, & people in crisis all need human presence, not automated scripts. You can't learn empathy from an algorithm or find comfort in a chatbot when you're frightened & confused. These needs are stubbornly, irreducibly human.

If an economy can't reward what society genuinely needs, it's mispriced. Right now, it rewards processing data, generating reports, reviewing contracts, managing logistics, all things machines are about to do better than people ever could. The value of those white-collar tasks is going to fall, & honestly it should, because a machine will do them faster, cheaper, & more consistently. I don't see that as tragedy; I see it as a shift in what should be valued.

What emerges from this reallocation is a clearer division between machine work & people work. Machines will outperform us on a growing range of tasks, just as factories did. But people will always people better than machines can. Raising children, comforting the bereaved, sitting with someone through pain, teaching a young person that they matter, holding a hand at the end of a life. Calling this "low-skill" misses the point entirely; care work is technically & emotionally demanding in ways no benchmark captures.

An economy that pays its data analysts well & its care workers poverty wages has its values inverted. When AI compresses the value of cognitive routine work, the practical response is to reprice care properly, not panic about every white-collar shift. We need to put a higher value on people who people better. That's the point of the Care Dividend: route a slice of AI-era gains into care capacity.

The Feminist Case: Meet Women Where They Are

I'm not interested in purity tests here. I know what the end game looks like: women having more options & more choices about how they live their lives. One of the most direct ways to create those options is to address the economic inequities that exist right now. I'm talking about today's labour market, not an imagined future equilibrium.

Women are already doing this work. That 76% isn't a forecast; it's current reality. Eight trillion dollars a year of unpaid labour, performed overwhelmingly by women, feeding directly into the functioning of every economy on the planet. Paying for care doesn't trap women; it gives real economic choice where none currently exists. Financial recognition for labour they are already performing means they are no longer economically penalised for doing it. That is a precondition for genuine choice, not an obstacle to it.

I am fully supportive of women getting into STEM. It matters enormously. But the conversation about women & AI cannot only be about the women who are positioned to enter technology careers. It has to include the millions of women who aren't in STEM & aren't going to be, not because they lack ability but because they are doing something else that society depends on. Supporting women into AI careers & supporting women who are raising children, caring for elderly parents, & maintaining households aren't in conflict; we need both. A feminism that only speaks to women who can code leaves most women out of the conversation entirely.

This is where I should be honest about my own position, because it shapes how I think about all of this. I would be the first person to tell you how unethical big tech is. I also teach people how to leverage their products all of the time. That looks contradictory, but most of us haven't picked the system we get to participate in. You wake up one day & your job, your bank, your doctor, your children's school, all of it lives inside a handful of platforms that you had no say over. That's not an individual failing; it's how the system is built. The majority of people have had very little practical option to not participate. The trap isn't using the tools; it's that opting out can cost you income, access & basic participation in daily life.

There are people who engage in purity politics around this, who will tell you to just disconnect, just refuse, just walk away. That advice only works if you already have the money, the time, & the safety net to absorb the cost. The people most harmed by these technologies are usually the ones with the fewest alternatives. I cannot change who wins or loses at a macro scale, but I can put my time & energy into helping everyday people navigate what's in front of them. Telling someone who is struggling that you have a better vision for the future feels hollow if that vision doesn't meet them where they are today. Until individuals can practically & affordably seize the compute, literacy is the liberation.

The same logic runs through the care argument. You cannot tell women currently performing eight trillion dollars of unpaid care that a more equal future is on its way & expect that to mean much while the bills go unpaid. You have to meet them where they are: already doing the work, already absorbing the cost, already holding the system together.

The objection I hear most often is that paying women for care risks "entrenching the gender division of labour." I understand the concern. If the policy amounts to "here's some money, keep doing what you're doing," then yes, it could reinforce the expectation that care is women's work. But that objection, taken to its logical conclusion, argues against compensating women for work they are already doing, on the grounds that the ideal future looks different. In the meantime, real women are absorbing real costs. You cannot ask a generation of women to subsidise the economy for free while you wait for a redistribution that shows no sign of arriving on its own.

Policy has to run on two tracks at once: compensate current care work & reduce the future imbalance. That means paying care workers a living wage, which would draw more men into the sector through simple economics. Better childcare & respite services to lift the unpaid burden. Pension & national insurance credits so time spent caring doesn't kill a lifetime's earnings. & yes, supporting women into STEM, technology, & every other field where they're underrepresented.

The gender pay gap in the UK is not an abstraction. It is partly driven by exactly this dynamic. Women step out of the labour market or reduce their hours to provide care. Their earnings drop. Their pension contributions drop. Their career progression stalls. The gap widens over a lifetime. Even women who return to full-time employment rarely recover the lost trajectory. Compensating care directly addresses this at the point where the damage occurs, rather than attempting to fix it downstream through equal pay legislation that cannot reach unpaid work.

Compensation should follow caregiving itself, not gender labels, because the objective is to value the work regardless of who does it. But honesty requires acknowledging that the overwhelming majority of beneficiaries would be women, because women are the overwhelming majority of carers. That isn't a flaw; it's the policy matching the labour split we actually have.

Lessons from the Nordics

If the objection is that none of this can work, the Nordic countries have already demonstrated that much of it does.

Sweden spends 3.32% of GDP on family benefits. Denmark spends 3.15%. Norway spends 2.78%. The United Kingdom spends 1.85%, less than the OECD average of 2.35% & roughly half of Sweden's commitment (OECD Family Database, Indicator PF1.1, 2021 data; individual country figures derived from OECD Social Expenditure Database). The UK ranks 23rd out of OECD nations on this measure.

The results are visible in labour market outcomes. Female labour force participation in Sweden is 61.68%, in Norway 61.72%, & in Denmark 59.77%, compared to 57.28% in the UK (World Bank/ILO, 2024). The gender gap in employment is smallest in Iceland, Finland, & Sweden, where it falls below three percentage points.

Sweden's parental leave system offers 480 days per child, with 90 days reserved for each parent that cannot be transferred to the other. If a father does not use his 90 days, they are lost. The first 390 days are paid at 77.6% of qualifying income. Denmark guarantees universal childcare from 26 weeks of age, with municipalities covering at least 75% of operational costs & low-income households paying nothing at all. Over 90% of Danish three-to-five-year-olds are in formal childcare (Danish Ministry of Children & Education, 2025).

Nordic care workers are paid significantly more than their UK counterparts. In Denmark, average care worker pay is approximately £17.20 per hour, compared to the UK median of £12.60, a gap of roughly 37% (Skills for Care, 2026; Nordic Council of Ministers, 2025). This is driven by sectoral collective bargaining agreements covering 80 to 90% of the workforce. In the UK, adult social care is a largely privatised, fragmented market with weak collective bargaining. The structural difference in how care is organised explains much of the pay difference.

The Nordic model is funded by general progressive taxation. Tax-to-GDP ratios range from 40% to 45%, compared to roughly 33% in the UK (Tax Foundation, 2023). There is no earmarked AI levy or equivalent.

From this, two practical points matter most. The outcome I'm proposing already exists in countries with comparable economies. So the issue isn't affordability in the abstract; it's political willingness. & Nordic evidence actually bolsters the case for an AI-linked funding route in UK conditions. The Nordic countries built their care infrastructure over decades through high general taxation. The UK is not starting from that base & is unlikely to reach it through across-the-board tax rises in the current political environment. An AI dividend offers a route to begin closing the gap by attaching new revenue to new economic gains, rather than asking for higher taxes on existing incomes.

The Arithmetic: A UK Test Case

I want to be honest about what the numbers allow & what they don't.

Before the numbers: what follows are transparent scenario calculations against the government's own upper-bound estimate, not revenue forecasts. The care spending figures are England-only (Skills for Care, Health Foundation, IPPR), while the AI productivity estimate is UK-wide, so there is a geographic mismatch that slightly overstates what any England-specific policy could capture. I flag this because honesty about assumptions matters more than clean arithmetic.

The most useful official benchmark is the UK government's claim that AI could be worth up to £47 billion a year to the economy over a decade, if gains are fully realised (GOV.UK, 2025). That figure comes with caveats. The government's own labour market assessment says it is derived from assumptions, not from observed data, & that adoption remains modest. The Department for Science, Innovation & Technology notes that only around one in five firms currently use or plan to use AI. £47 billion is a ceiling scenario, not a prediction. But it is the government's own upper bound, & it gives us a transparent base for arithmetic.

At 5%, 10%, & 25%, you're looking at roughly £2.35 billion, £4.7 billion, & £11.75 billion a year if those upper-bound gains materialise. A deliberately conservative case, where only one quarter of the estimated gains is actually taxable & the levy is 10%, still yields approximately £1.175 billion.

That last figure, the conservative one, would cover the net £330 million annual cost estimated by IPPR for paying all adult social care workers in England the real Living Wage more than three times over. It would not transform the care economy, but it would make a measurable difference to the lowest-paid workers in one of the most strained sectors in the country.

The £4.7 billion from a 10% levy at the full upper bound would exceed the Health Foundation's estimate that England needs an additional £3.4 billion to meet publicly funded adult social care demand by 2028/29. It would equal roughly a tenth of current combined public spending on adult social care (£34.5 billion) & early years support (£10.5 billion, per IFS).

The £11.75 billion from a 25% levy would slightly exceed current public childcare support spending. That is not small money by any measure.

But even the full £47 billion, if every penny were hypothetically redirected, would only cover about a quarter of Carers UK's £184 billion UK-wide estimate for unpaid care, & roughly 3% of the ONS estimate for all unpaid household services. No, this won't monetise all care. Yes, it can fund a meaningful recurring reform package.

The practical offer is annual, material reform funding, not total coverage, linked to a revenue stream that grows as AI adoption grows.

There is also a fiscal multiplier argument. Research by De Henau & Himmelweit (2021), published in Feminist Economics & commissioned by the Women's Budget Group, found that investment in care creates 2.7 times as many jobs as the equivalent investment in construction. For women specifically, care investment creates 6.3 times as many jobs. Crucially, even for men, care investment creates 10% more jobs than construction, not fewer. The same research found that 50% more tax revenue is recouped by the Treasury from care investment than from construction spending, & that care investment produces 30% less greenhouse gas emissions.

Expanding the care workforce to 10% of the employed population, which is where Sweden & Denmark already are, & paying all care workers the real Living Wage, would create approximately 2 million jobs & increase overall employment by 5 percentage points (De Henau & Himmelweit, 2021). This is redistribution, yes, but it's also high-yield job creation with a better return than the infrastructure spending governments typically favour.

What a Workable Model Looks Like

The tax base should not depend on proving that a specific algorithm replaced a specific worker. The UK government's own assessment says the current evidence does not give clear policy answers on causality, & that exposure to AI is not the same as adoption. Taxing each automation event would be a bureaucratic mess & analytically flimsy.

The IMF has been explicit on this point. A specific robot or AI tax is not advisable because it would be hard to administer & could suppress productive adoption (IMF, 2024). That critique knocks out a tool-tax model, not a profits-based levy. The IMF's own recommendation is stronger corporate & capital income taxation, including supplemental taxes on excess profits, as the mechanism for ensuring AI gains are broadly shared.

In practice, the levy should attach to observable rents: excess profits of large AI providers & AI-intensive deployers, strengthened taxation of capital income, & the removal of corporate tax features that artificially accelerate labour displacement through software or hardware investment.

This is less radical than it sounds. HM Revenue & Customs already administers several targeted corporate levies: the Bank Levy, Bank Surcharge, Residential Property Developer Tax, Energy Profits Levy, & Electricity Generator Levy. The precedent is clear. When profits are unusually concentrated or arise from strategic economic shifts, the UK tax system can & does impose sector-specific or windfall-style charges. An AI dividend applied to concentrated AI rents would be novel in subject matter, but it would not be administratively alien.

Design-wise: make it graduated, set a real threshold, & link reliefs to augmentation outcomes. The levy should capture supernormal returns without penalising firms making modest use of AI to improve ordinary operations. Pair it with enhanced capital allowances for AI investments that demonstrably create new roles or improve working conditions, so the incentive structure favours keeping people in work rather than replacing them.

The IMF's own analysis warns against blunt approaches. Their fiscal policy work argues that broad-based capital taxation, while effective at reducing inequality, can also suppress output & wages if set too high or applied too indiscriminately (IMF, 2024). The practical lesson is clear: skim concentrated rents, do not broadly penalise productive adoption. The slogan "tax AI labour" becomes much stronger once it is translated into "tax AI rents."

One practical risk is cross-border leakage. The IMF's 2026 scenario note warns that AI can reinforce winner-take-most market structures & that taxation requires international coordination to limit cross-border spillovers. A national government cannot assume all rents created by domestic AI use will appear neatly inside its own tax base. Any serious version of this policy therefore requires both domestic legislation & engagement with international tax coordination efforts, including the OECD BEPS framework that is already addressing related questions in digital taxation.

Where the Money Goes

Without a spending plan, revenue projections are just headline maths. The Care Dividend should be split between two objectives: recognising unpaid care & building paid care capacity.

On the unpaid care side, the most immediate reform is Carer's Allowance. The 2026/27 rate is £86.45 per week, with an earnings limit of £204. If a claimant provides 35 hours of care, the minimum qualifying threshold, that works out at roughly £2.47 per hour of care. That isn't a wage in any serious sense; it's token support. Reforming Carer's Allowance by raising the rate, increasing the earnings limit so carers are not penalised for also working, & extending eligibility to those currently excluded would be one of the highest-impact, lowest-cost interventions available.

On the paid care side, the priority is straightforward: wages. If care workers are paid poverty wages, you cannot recruit enough of them, & the ones you have leave. You do not solve a recruitment crisis in a sector with a 7% vacancy rate & over 111,000 unfilled posts by running advertising campaigns. You solve it by paying people properly. The IPPR estimated that paying all adult social care workers in England the real Living Wage would cost a net £330 million per year, less than 2% of the social care budget. That is within reach of even the most conservative revenue scenario modelled in this paper.

Beyond immediate wage reform, the spending framework should include: childcare expansion, because inadequate childcare is one of the primary mechanisms through which women are pushed into unpaid care & out of the labour force; respite services, because unpaid carers who burn out become healthcare costs; & pension or national insurance credits for time spent caring, because the long-term earnings impact of caregiving is one of the main drivers of the gender pension gap.

The core logic is straightforward: pay care labour fairly, expand capacity, & stop caregiving from wrecking lifetime income. Do all three using revenue that grows as the AI economy grows.

Answering the Obvious Objections

I have had these conversations enough times to know what comes next. So let me go through the predictable pushback & explain why most of it doesn't hold up.

The first thing people say is that taxing AI will kill innovation or scare investment out of the UK. This gets wheeled out every time anyone proposes taxing concentrated wealth, & it deserves a direct answer. Yes, a badly designed robot tax could suppress useful investment. The IMF says as much. But I'm not proposing a tax on the tool. I'm proposing a tax on supernormal profits from it. Those are different things.

The UK already does this. The Bank Levy has been running since 2011. The Energy Profits Levy arrived in 2022. Neither caused banks or energy companies to pack up & leave. They adjusted, because they need access to the UK's consumer market, its regulatory environment, its talent, its legal infrastructure, & its financial ecosystem. AI companies need the same things. They are not going to relocate to a low-tax jurisdiction that offers none of that because Britain started skimming a share of extraordinary profits. The companies that threaten to leave over taxation rarely do. The ones that actually leave were never committed to the domestic economy in the first place. & if the worry is about small firms, the answer is straightforward: set the threshold high enough that it only touches the biggest & most profitable deployers. Startups don't have supernormal profits to tax.

People also ask why I don't just advocate for raising general taxation. I could. The Nordic countries fund their care systems that way & it works brilliantly. But they built that tax base over decades, with tax-to-GDP ratios between 40% & 45% compared to Britain's 33%. No UK government is going to raise income tax by twelve percentage points to match Denmark. An AI dividend is politically viable in a way that across-the-board tax rises are not, because it attaches new revenue to new economic gains rather than asking people to pay more from existing incomes.

Then there's the question of why care specifically, rather than just general welfare. Because the care bottleneck isn't abstract. It is concrete, measurable, & already causing damage. Over 100,000 vacant posts. A 7% vacancy rate. A projected need for 470,000 more workers by 2040. A £3.4 billion funding gap just to meet existing demand by 2028/29. Care isn't one worthy cause among many. It is one of the economy's binding constraints. On top of that, every pound invested in care creates 2.7 times more jobs than the same pound spent on construction, & generates 50% more in tax revenue back to the Treasury (De Henau & Himmelweit, 2021). It's not just morally right. It's better economics.

Yes, AI training is still part of the answer. Nobody is arguing against skills policy. The IMF backs it, & so do I. I spend a significant portion of my working life helping people build AI skills. But skills alone cannot solve unequal capital ownership, unequal access to tools, or the unpaid care burden that already shapes women's choices long before AI enters the conversation. Training is necessary. On its own, it is not enough.

And yes, AI can help the care sector. The best use case is cutting paperwork, scheduling friction, & routine admin so workers can spend more time with actual people. The WHO's position is that AI should enhance health, not substitute for human relationships. I agree completely. You don't want a chatbot sitting with your gran. You want a human being doing that, properly supported & properly paid. AI should be making that possible, not replacing it.

The hardest objection is timing. Care commitments are permanent & recurring. AI revenue is uncertain & potentially volatile. The UK government's own review admits that the evidence is still emerging & that major policy questions remain open. Fair enough. That means the design has to be conservative: start with a modest levy, build a reserve fund in the early years, & maintain a floor funded from general taxation so that if AI revenue underperforms, care workers & carers aren't the ones left exposed. The worst possible outcome would be hiring thousands of care workers on the back of optimistic projections & then defunding their posts two years later.

Conclusion: Human Potential

In 2015, I ended a talk with a line that has stayed with me: no matter the revolution, the most valuable resource has been, & always will be, that of human potential.

I was talking about digital inequality. About young people locked out of the economy by barriers they didn't create. About a system that measured success by individual achievement while ignoring the collective structures that make achievement possible. I said we are interdependent beings in an interdependent society, & that needs to be reflected in the values we impart.

Having spent years in youth work watching young people try to navigate a system that wasn't built for them, & then years in AI consulting watching the same dynamics repeat at a larger scale, I keep landing on the same question. It's never just "what can the technology do?" The real test isn't technical capability; it's distribution: who gains, who pays, who carries the hidden labour.

Reese Witherspoon asked women to catch up. The evidence suggests that catching up, on its own, won't close the gap. The gap is structural. It requires a structural response.

I'm not pitching utopia; I'm pitching a fundable policy route with known trade-offs. The Care Dividend takes a share of concentrated AI gains & puts it into the work the entire economy depends on but refuses to pay for. The maths is plausible, the policy precedents are real, & the evidence base is strong enough to act.

The facts are already in front of us: care is enormous, mostly done by women, persistently underfunded, & economically miscounted. AI is generating concentrated returns that will grow over the coming decade, & without deliberate policy those returns will flow to the people who already have the most.

Technology has always driven & shifted what society values. But what I keep coming back to, after a decade of working in this space, is that regardless of whether someone is neurodiverse or neurotypical, a bright mind or bang on average, whether they struggle or soar, we should be cultivating economic systems that serve all people, all intelligences, all capabilities, both mental & bodily, with people contributing what they fundamentally can, & with society meeting the basic needs of every single individual. If a system can't meet that baseline, it's failing its purpose.

The Care Dividend is a step toward a system that is not obsolete. It prioritises social good & socially acceptable outcomes over the unchecked concentration of technological gains. An AI economy should be productive, yes, but also visibly more humane in outcomes. & the most direct way to get there is to fund the work that is already, by definition, the most human thing we do.

The feminist response to AI isn't "learn faster." It's building an economy that values what women already do. Tax concentrated AI rents & channel the proceeds into paid care capacity & fair caregiver support.

Declaration of Interest

The author is a practising AI & automation consultant who works with organisations on AI implementation & adoption strategy. This paper argues for taxing concentrated AI profits, which could affect the sector in which the author operates. That position is declared in the interest of transparency. The argument is made because the author believes it is correct, not in spite of working in the field but because of it.

AI Use Disclosure

AI tools were used to assist with research compilation, citation verification, & early structuring during the preparation of this paper. All arguments, analysis, policy positions, & conclusions are the author's own.

Funding

This research received no external funding.

Data Availability

All data cited in this paper are drawn from publicly available sources listed in the references. No original datasets were generated.

References

Brett, J.L. (2015), "Equipping a generation to take its place in the digital revolution." TEDxWandsworth. Available at: youtube.com/watch?v=aEjUlH7bNDo

Carers UK (2025), State of Caring 2025: The Cost of Caring. London: Carers UK. Available at: carersuk.org

Danish Ministry of Children & Education (2025), Income-based daycare subsidy, sibling subsidy & other subsidies. Available at: lifeindenmark.borger.dk

De Henau, J. & Himmelweit, S. (2021), "A Care-Led Recovery from Coronavirus: Investing in High-Quality Care to Stimulate & Rebalance the Economy," Feminist Economics, 27(1-2), pp. 1-26. DOI: 10.1080/13545701.2020.1845390

Department for Science, Innovation & Technology (2025), Assessment of AI capabilities & the impact on the UK labour market. Available at: gov.uk

Tax Foundation (2023), "How Scandinavian Countries Pay for Their Government Spending." Available at: taxfoundation.org

Murgia, M. & Burn-Murdoch, J. (2026), "High earners race ahead on AI as workplace divide widens," Financial Times, 23 April 2026. Available at: ft.com

GOV.UK (2025), "Prime Minister sets out blueprint to turbocharge AI." Available at: gov.uk

GOV.UK (2026), Benefit & pension rates 2026-2027. London: Department for Work & Pensions. Available at: gov.uk

Health Foundation (2025), Poverty, pay & the case for change in social care. London: Health Foundation. Available at: health.org.uk

House of Commons Library (2021), How has the coronavirus pandemic affected women in work? London: UK Parliament. Available at: commonslibrary.parliament.uk

IFS (2023), Early years spending. London: Institute for Fiscal Studies. Available at: ifs.org.uk

ILO (2018), Care Work & Care Jobs for the Future of Decent Work. Geneva: International Labour Organization. Available at: ilo.org

ILO (2026), Gen-AI, occupational segregation & gender equality in the world of work. Research Brief. Geneva: International Labour Organization. Available at: ilo.org

IMF (2024), "AI Will Transform the Global Economy. Let's Make Sure It Benefits Humanity." Washington, DC: International Monetary Fund. Available at: imf.org

IMF (2024), "Fiscal Policy Can Help Broaden the Gains of AI to Humanity." Washington, DC: International Monetary Fund. Available at: imf.org

IMF (2024), Gen-AI: Artificial Intelligence & the Future of Work. Staff Discussion Note SDN/2024/001. Washington, DC: International Monetary Fund.

IMF (2026), AI & the Global Economy. IMF Notes. Washington, DC: International Monetary Fund.

IPPR / Living Wage Foundation (2024), Real Living Wage for Social Care: Technical Report. London: IPPR. Available at: livingwage.org.uk

King's Fund (2025), Social Care 360: Expenditure. London: King's Fund. Available at: kingsfund.org.uk

King's Trust (2025), The King's Trust TK Maxx Youth Index 2025. London: The King's Trust. Available at: kingstrust.org.uk

Mazzucato, M. (2018), The Value of Everything: Making & Taking in the Global Economy. London: Allen Lane.

Nordic Council of Ministers (2025), Towards Pay Equity: Explaining the undervaluation of women's work in the Nordic countries. TemaNord 2025:546. Available at: pub.norden.org

Nordic Council of Ministers (2025), Parental Leave in Sweden. TemaNord 2025:547. Available at: pub.norden.org

Nuffield Trust (2024), National policy options to improve care worker pay in England. London: Nuffield Trust. Available at: nuffieldtrust.org.uk

OECD (2021), OECD Family Database, Indicator PF1.1: Public spending on family benefits. Available at: oecd.org

OECD (2024), Algorithm & Eve: Women, Work & AI. Paris: OECD.

OECD (2024), Who will be the workers most affected by AI? Paris: OECD.

ONS (2023), Household satellite account, UK: 2023. London: Office for National Statistics. Available at: ons.gov.uk

Oxfam (2020), Not all gaps are created equal: The true value of care work. Oxford: Oxfam International. Available at: oxfam.org

Prince's Trust (2015), Prince's Trust Youth Index 2015. London: The Prince's Trust.

PwC (2021), Women in Work Index 2021. London: PricewaterhouseCoopers. Available at: pwc.co.uk

Skills for Care (2025), The state of the adult social care sector & workforce in England 2025. Leeds: Skills for Care. Available at: skillsforcare.org.uk

Skills for Care (2026), Pay in the adult social care sector in England, December 2025. Leeds: Skills for Care. Available at: skillsforcare.org.uk

UK Parliament (2025), Written evidence: Carer's Allowance & social care evidence. London: UK Parliament.

WHO (n.d.), Harnessing artificial intelligence for health. Programme page. Geneva: World Health Organization. Available at: who.int

Witherspoon, R. (2026), quoted in Fast Company, "Reese Witherspoon says using generative AI is feminist." Available at: fastcompany.com

Women's Budget Group (2020), A Care-Led Recovery from Coronavirus. London: WBG. Available at: wbg.org.uk

World Bank (2024), Labor force participation rate, female (% of female population ages 15+). ILO modelled estimates. Available at: data.worldbank.org

World Economic Forum (2024), The future of the care economy. Geneva: WEF.

That detail has stayed with me for weeks. Most "consumer choice" tools, the ones digital teams spend fortunes building, are about letting people pick a colour, a delivery slot, or a subscription tier. The French brand C'est qui le Patron, which translates roughly as "Who's the Boss", asked a different question. It asked who decides what fair looks like, and then it built the digital scaffolding to let ordinary people answer.

I work with a lot of leaders who are anxious about AI, about automation, about the next wave of platform consolidation. Underneath the anxiety, the real question is rarely technical. It is almost always about power. Who gets to decide how this tool is used. Who benefits when it works. Who carries the cost when it does not. I grew up working class, was helped through a rough patch by The Prince's Trust, and spent years working with young people who had been written off by systems designed without them in the room. That experience shaped a stubborn instinct in me. When somebody shows me a clever new system, my first question is not "how does it work?" It is "who got a vote?"

C'est qui le Patron is interesting precisely because the technology behind it is not exotic. Online voting. Transparent pricing models. Supply chain data shared with the people buying the product. Any competent digital team could build this in a quarter. The innovation is not in the stack. The innovation is in the decision about who is allowed to be a designer of the product. Members pay a symbolic €1 to join, and then vote on the criteria that shape each product, including how much the producer is paid.

That is a small sentence with a large implication. In most consumer goods companies, the producer's pay is treated as a cost line to be minimised. Here it is treated as a value the buyer expresses an opinion on. The whole logic of the supply chain inverts. Once you have voted that a French dairy farmer should receive €0.39 a litre, you have made yourself a stakeholder in their livelihood. You are no longer a passive consumer at the end of the chain. You are a participant in how the chain is shaped.

This is what I mean by digital economic justice. It is not a slogan. It is the use of perfectly ordinary digital tools, voting platforms, dashboards, transparent pricing, to redistribute who gets to make decisions, rather than to optimise the decisions of whoever already had the power. There is a difference between a digital transformation that makes an existing hierarchy faster, and a digital transformation that questions whether the hierarchy was the right shape in the first place. Most enterprise change programmes I see are the first kind. C'est qui le Patron is the second.

The results are not theoretical. The brand has expanded from milk to a range of more than thirty products including honey, apple juice, chocolate, flour, wine, sardines, pizza, yoghurt, chicken, baguettes and butter, sold through Lidl, Intermarché and Carrefour. This is not a niche, organic, north London co-op story. It is at scale, on the shelves of mass-market supermarkets, alongside the same brands you and I buy from on a Tuesday night when we have forgotten to plan dinner.

I want to be honest about the limits, because I am wary of presenting any one model as a universal answer. C'est qui le Patron still works inside the conventional retail system. It still depends on people having enough disposable income to vote with their wallets in the first place. The poorest shoppers, the ones for whom forty pence on a pack of butter genuinely matters, are not the people most easily included in this kind of premium-priced democratic experiment. So the model is not a complete redistribution of power. It is a useful, working demonstration of what becomes possible when the digital interface between producer, retailer and buyer is designed for participation instead of extraction.

For anyone leading a digital function inside a larger organisation, the question this raises is uncomfortable and worth sitting with. Where in your business could the people most affected by a decision be given a genuine vote on it. Not a satisfaction survey after the fact. Not a focus group whose findings get filtered through three layers of management. An actual mechanism, with real consequences, where the choice they make changes what you do. Frontline staff voting on shift patterns. Customers voting on data-sharing defaults. Suppliers voting on payment terms. Citizens voting on how a public service uses their information. The technology to do any of these is sitting on the shelf. The thing that is missing, almost everywhere, is the willingness to share the decision.

That is also why I keep coming back to the line that machines now machine better than humans ever could, so the question for us becomes what humans must now do better than machines ever will. Designing systems that distribute power well, rather than concentrate it efficiently, is one of those things. AI will not do it for us. Faster algorithms will not do it for us. It is a choice, made by people with budgets and remits, about who gets to be in the room when the decision is taken.

The French shoppers who voted on the price of milk did not solve the global food system. They did demonstrate something important. When you give people who are usually treated as the end of the pipeline a real seat at the design table, they tend to make decisions that are more generous, more patient, and more long-term than the spreadsheet predicted. That should make every leader in the country pause for a moment and ask what they have been protecting, and from whom.

The polished policy problem

A recent Nemko analysis of why AI governance efforts fall short makes an uncomfortable point: the issue is not awareness or intent, it is execution. Organisations have the principles, the policies, even the EU AI Act on the horizon, yet many still struggle to translate these efforts into real control over how AI systems behave.

• Boards approve frameworks they will never see applied.

• Legal writes policy. Engineering writes code. The two rarely meet.

• Risk registers exist. Workflows do not reference them.

If your governance document cannot be traced to a specific step in a specific workflow, it is decoration.

Gap one: nobody actually owns it

The Nemko piece flags something I see in almost every client engagement: responsibilities are distributed across legal, compliance, engineering and product, but decision-making authority is unclear. Everyone is consulted. Nobody is accountable.

What this looks like on the ground:

• A model goes live because no one had the authority to say no.

• A bias concern gets raised in three meetings and resolved in none.

• When something breaks, the post-mortem blames "the process".

A simple RACI model fixes a surprising amount of this. Who is Responsible for writing the policy. Who is Accountable for approving the deployment. Who must be Consulted before a change. Who must be Informed after. Not glamorous. It works.

Gap two: treating every use case the same

The second failure is applying one control regime to everything. A chatbot that suggests meeting times does not need the same scrutiny as a model that triages loan applications. Yet many organisations do exactly that, which slows the low-risk work and under-governs the high-risk work at the same time.

A workable sort:

• Low risk: document the use case, light review, ship it.

• Medium risk: human-in-the-loop, periodic audit, clear rollback.

• High risk: full validation, explainability requirements, named accountable owner.

Risk tiers are not about paperwork. They are about matching the weight of oversight to the weight of the decision.

Map every policy line to a workflow step

Here is the test I use with leadership teams. Take your AI policy. Pick any line. Ask three questions:

• Which workflow does this apply to?

• Which step in that workflow?

• Who owns that step when something goes wrong?

If you can answer all three, you have governance. If you cannot, you have theatre. The policy is doing the emotional work of looking responsible without the operational work of being responsible.

This is also why ethics bolted on at the end rarely survives contact with a deadline. Fairness, accountability and transparency have to live inside the build process, not in a committee that meets once a quarter.

A quieter point about automation

Governance often fails at the execution layer because the execution layer is buried under administrative noise. Teams meant to oversee AI are drowning in calendar invites, approval chains and inbox triage. When the people responsible for judgement have no time to judge, oversight becomes rubber-stamping.

Some of that admin is a legitimate candidate for delegation to well-scoped AI agents, precisely so humans can spend their attention on the decisions that matter. That is the argument I will be making in more detail at a session on getting AI agents to do your admin on 20 May 2026, walking through the 28-agent setup we run inside Bykov-Brett Enterprises. Worth a look if your governance people are too busy to govern.

A few questions worth taking to your next leadership meeting

• Pick the three highest-risk AI use cases in your organisation. Can you name the individual accountable for each, not the committee?

• When did your AI policy last change a decision that would otherwise have gone the other way? If you cannot think of an example, the policy is not operating.

• Are your governance people close enough to the build to catch issues early, or are they reviewing decisions that have already shipped?

Governance is not the document. It is what happens in the five minutes before someone presses deploy. If no one owns those five minutes, nothing else you have written matters.