AI and the Future of WorkWhat It Is Already Doing Inside Your Organisation

In April 2025, Tobi Lütke, the chief executive of Shopify, posted an internal memo on X, saying it was about to leak anyway. Its opening line set the terms: “Reflexive AI usage is now a baseline expectation at Shopify.” AI use would be written into performance and peer reviews. Teams asking for more people would first have to show why AI could not do the work instead.

Much of the commentary read the memo as a forecast, a glimpse of where every company was heading. It is more useful read as a description of the present. In one short memo it changed what it takes to be a member of the organisation in good standing and moved the burden of proof about who is needed. Then it tied both to the performance review, which is where people learn what an organisation wants from them.

AI and the future of work is one of the most discussed subjects in management, and nearly all of the discussion happens in the future tense: which jobs are exposed, which skills will be in demand, what the organisation of 2030 will look like. The present gets far less attention. Inside organisations, AI is already changing who holds power and what people can admit they don’t know, what it means to be good at a job and who feels they still belong. Very little of this gets named, and the future tense is part of the reason.

Why AI and the Future of Work Is Discussed in the Future Tense

Futures are comfortable places to put difficult things. A forecast carries none of the accountability of a decision made last Tuesday. When an organisation talks about AI as the future of work, it can talk at length about change without discussing what has already changed, or who changed it.

The future tense also flatters the people who use it. Preparing an organisation for the future is a leadership task, with its own strategy days and roadmaps. Dealing with the present consequences of decisions already taken is messier, and it tends to involve the people those decisions affected. Framing AI as the future of work keeps the conversation among the people doing the preparing.

A great deal already has been. In Microsoft and LinkedIn’s 2024 Work Trend Index, a survey of 31,000 people in 31 countries, three in four knowledge workers said they were using generative AI at work, and nearly four in five of them were bringing their own tools rather than waiting for their employer to provide any. In the same survey, 60 per cent of leaders worried that their organisation lacked a plan or vision for AI. The technology reached people’s working days before it reached the strategy. Much of the change happened below the level at which the organisation was looking, so its picture of its own AI transition was, from the start, a picture of the part it had sanctioned.

The part it can see is less reliable than it looks. In 2025 the research group METR ran a randomised trial with experienced open-source developers working on their own code. Before starting, the developers expected AI tools to make them about a quarter faster. With the tools, they took 19 per cent longer. Afterwards they still estimated they had been about 20 per cent faster. METR presents the result as a snapshot of one setting at one point in time, and the tools have improved since. The gap it exposed is the lasting finding. People using AI are poorly placed to judge what it is doing to their work, and a large share of what organisations believe about AI productivity rests on those same judgements.

This is where unknowing stops being abstract. An organisation that has committed money and reputation to AI, and made headcount decisions on the strength of it, needs the investment to be working. Evidence that it is working arrives easily and is welcomed. Evidence that it isn’t has to meet a higher standard of proof. Self-reported time savings make their way into board papers. Other signals travel less well: work that has to be redone, judgement thinning out, the junior analyst who can produce a report and cannot defend it.

The evidence is also skewed before anyone looks at it. The people invited into AI pilots are usually the people who volunteered, so much of what an organisation knows about how AI is working comes from its enthusiasts, reporting on tools they chose to leaders who were hoping for good news. Sceptics are consulted later, if at all, and by then their doubts arrive as objections to a decision instead of information about it. This happens without anyone being dishonest, through the ordinary habit organisations have of listening hardest to the people who already agree with them.

Someone has to carry the uncertainty the organisation won’t hold. Admitting that nobody knows what AI will do to a given role is costly for a leader whose authority rests on seeming to know, so the admission rarely comes. The uncertainty moves down the hierarchy instead and settles in individuals, who are left to wonder about their own prospects privately, often while being told the future is bright.

How AI Moves the Burden of Proof at Work

Every AI implementation is also a decision about power, though it seldom presents itself as one. The Shopify headcount rule shows what the shift looks like in practice. Under the old arrangement, a manager who wanted another person made a case for the work. Under the new one, the manager has to prove a negative: that a machine cannot do it. The people with most at stake in that argument, those whose work might be automated, are rarely the ones making it, and they are rarely in the room. Whatever the intention behind such a rule, its effect is to make the absence of people the default position and their presence the exception that needs justifying.

Once that default is in place, AI becomes useful for purposes that have little to do with AI. Restructures planned for other reasons acquire a justification that is hard to argue with, because arguing with it sounds like arguing with the future. Timelines follow what leadership has already announced to investors and boards rather than what the organisation can absorb. The technology becomes an argument that ends arguments. People lower down usually know when this is happening; they can see which changes were on the table long before anyone mentioned AI. What they lack is a way of saying so that won’t be heard as resistance.

Informal power moves as well. People fluent with the new tools become the people others consult, and some of that is healthy, expertise finding its level. Fluency is also a signal, though, and signals get read by whoever controls promotion. An organisation can sort its people into those who represent where it is going and those who represent where it has been, without anyone ever writing the distinction down.

And the watching grows. In April 2026 polling for the AFL-CIO, 94 per cent of working Americans said workers should be told if AI is being used to monitor their work. Seven per cent said their employer had told them. Foucault’s observation about discipline was that surveillance works best when it is possible rather than certain, because people start watching themselves on the watcher’s behalf. Monitoring that is suspected but never disclosed fits his account closely. It also sets up the next problem, because people who are watching themselves become very careful about what they let anyone else see.

The Double Bind of Using AI at Work

The most revealing finding in Microsoft’s report had nothing to do with productivity. Just over half of the people using AI at work said they were reluctant to admit using it for their most important tasks, and slightly more worried that using it on important work made them look replaceable. These were the same people their organisations were urging to use AI more.

Gregory Bateson called a situation like this a double bind: a contradiction the person cannot easily leave and is not permitted to name. AI at work can meet both conditions. The explicit message says use AI: it is the future, and your review will ask about it. The implicit logic, which the headcount rule makes official, says that work AI can do is work the organisation may not need people for. Put the two together and visible, skilful use of AI becomes evidence that your work can be done without you, while reluctance marks you as someone the organisation is leaving behind. Naming the bind is risky in its own right. The person who says in a meeting that these tools make them feel replaceable has just told their manager they feel replaceable.

So people do what people in double binds usually do. They hide. They use the tools and say less about it. They present assisted work as unassisted, or keep their best uses to themselves. The organisation then complains that adoption is patchy and invisible, and answers with more pressure to use AI visibly, which tightens the bind.

Managers sit inside a version of the bind with an extra layer. They are asked to deliver the gains AI promises and to keep their teams engaged through the change, and under a rule like Shopify’s they are also the ones who must set out, in a headcount request, why the people they manage cannot be replaced by software. A manager who argues hard for the team’s irreplaceability is arguing against the organisation’s stated direction. A manager who doesn’t has, in effect, agreed with it. Most find a way of wording the request that commits them to neither, and learn to live in the vagueness.

This is a belonging problem as much as a productivity one. The bind tells people that their membership depends on demonstrating a capacity that may make them unnecessary, and it tells them this in a way they cannot safely discuss. The concealment also runs both ways. Workers keep their AI use from their employers; employers, on the AFL-CIO figures, keep their AI monitoring from their workers. An organisation in which both sides hide their relationship with the same technology has two conversations about AI going on at once, and neither of them is the one on the intranet.

That distance between the official account and the lived one is what the AI Transition Diagnostic was designed to surface. One of its contradiction pairs puts two statements side by side: that the AI transition is being managed collaboratively, with input from across the organisation, and that publicly disagreeing with how AI is being rolled out would be noted. Someone who agrees with both is describing an organisation in which collaboration is the story and caution is the experience. Neither answer is false. The space between them is where the organisation lives.

What Happens to Expertise When Junior Work Disappears

For most people, work is part of who they are. The thing they spent years learning to do well, and the judgement they have come to trust in themselves, account for much of what makes it meaningful. AI brings something organisations have very little language for: the possibility that the thing you are good at can be approximated by a machine, quickly and cheaply, well enough for most purposes.

The usual response is to treat this as a skills gap. People need to learn to work with AI, so the organisation provides training. The training is useful, and it also misses what is happening. A prompting course has nothing to say to someone wondering what two decades of judgement are worth when a newer colleague with a subscription can produce something that looks the same. That is a question about meaning, and answering it as a question about skills tells people their concern has been heard and filed.

The pressure is sharpest at the start of careers. In August 2025, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford published an analysis of payroll data covering millions of American workers. In the occupations most exposed to AI, employment of 22 to 25 year olds had fallen by double digits relative to other groups, while more experienced people in the same occupations held steady or grew. The decline was concentrated in work where AI automates tasks rather than assisting with them. The authors titled the paper “Canaries in the Coal Mine?” and kept the question mark. A year later, with data through June 2026, the same authors found the gap had widened to 19 per cent, and that it came mostly from young people not being hired rather than from anyone being let go.

The tasks leaving junior roles are the tasks through which people used to become senior: first drafts, reconciliations, literature searches, the summary nobody wanted to write. Nobody designed them as training, and nobody enjoyed them much. They trained people anyway, because doing something many times, badly at first, with someone more experienced correcting you, is how judgement forms. An organisation that automates the bottom rungs of its ladder saves money now and borrows against expertise it will need in ten years. The debt appears on no balance sheet, and the people who incur it will mostly have moved on by the time it falls due.

Further up, the meaning of good work is being revised without announcement. When the output of the careful expert and the output of the fast, assisted generalist are hard to tell apart on the page, organisations tend to reward the faster one, because speed is what their measures can see. The expert is left choosing between a standard nobody seems to measure any more and a new one with no place for what they were proud of. Many make that choice alone, which is how most of the changes described here get made.

There is an uncertainty inside this that rarely gets spoken. Experienced people often cannot tell whether their judgement is still better than the assisted version, because the difference seldom shows in the moment. It shows later and elsewhere: in the client relationship that holds through a difficult year, or in the error caught because someone had seen its cousin a decade earlier. Those outcomes are real, but they take time to appear and are hard to attribute to anyone. Speed shows up in this quarter’s figures. An organisation that rewards what it can see will, over time, lose the ability to see what it is losing.

Karl Weick argued that organisations make sense of what they do retrospectively. They act first and assemble an account of what the action meant afterwards. The account of this period will be written later, by whoever is in a position to write it, and it will probably say the organisation embraced AI and invested in its people. Some of that may even be true. What it is unlikely to hold is what people said to each other while it was happening, and what they decided not to say to anyone.

In that sense the future of work is something organisations will discover they have already had, described after the fact by people who were not always the ones living it. Whether anyone is describing it now, while it can still be argued with, gets decided in small ways every week: in who speaks at the town hall, and in which tools people admit to using.