The decision a company makes about its people is rarely the one in its press release. So the useful question about AI and work is not what the announcements say. It is what the incentives beneath them reward.
Three years of measurement have settled the argument. AI is not eliminating work at any scale visible in aggregate data; the most careful labour-market studies find no detectable displacement through 2026 (Anthropic, 2026; Yale Budget Lab, 2025). The change underway is quieter, and harder to manage. Capability keeps climbing: the length of task a model can finish unaided has doubled roughly every seven months for six years (METR, 2026). Adoption keeps broadening. What lags is everything that turns capability into delivered value. The fall from what a model does in a demo to what an organisation can depend on in production still runs about a third (deployment studies, 2025–2026). A token price that has dropped tenfold meets an agent that burns a hundred times more tokens to finish a real job, and the two settle at a number that is still high.
The capability question, can it, is closing. The operating question, can we trust it, wire it in, and afford it at scale, is not. The space between them is the subject of this piece. Call it the gap. Everything that follows is about who fills it.
The gap gets filled by someone.
Start from the assumption that companies act sensibly. Each one responds to the incentives it actually faces: cost of capital, quarterly expectations, the labour market it hires in, the regulation it sits under. Given those, the moves that follow are the rational ones. None of this requires bad faith.
The gap does not close because it is inconvenient. A cost that exists has to land somewhere. Economists have a precise word for this, taken from the study of taxes: incidence. The question that matters is never whether the cost exists. It is who actually bears it, which is rarely the party named on the invoice.
AI has an incidence. Every company will, in effect, decide who fills the distance between what the technology promises and what it delivers. The decisions will not present themselves as that. They arrive as a hiring plan, a tooling roll-out, a reorganisation, a new performance metric. Each is a sensible local choice. The difficulty is that sensible local choices aggregate into a structural outcome that no single decision-maker chose, and that no board paper is titled to describe.
The rational moves, and who fills the gap in each.
Don't replace the entry-level role. Stop hiring into it.
The cheapest way to adopt AI is to stop hiring. No severance, no notice period, no filing, no headline. The junior role goes unfilled and the team absorbs the work. It is already underway: the one clear signal in the labour data is a fall of about 14% in the job-finding rate for workers aged 22 to 25 in exposed fields, against no change for those over 25 (Anthropic, 2026). Nobody is fired. The bottom rung is simply not rebuilt.
The people never hired. They appear in no displacement figure, because they never entered the count.
It is the easiest of the seven to forecast, because it is the cheapest decision a CFO will make and it shows up in no statistic. The cost is deferred, not avoided. The mid-level talent a company needs in 2032 is the entry cohort it chooses not to hire in 2026, and a first rung that has been automated cannot later be climbed.
Keep the people to stand behind the work, not to do it.
Companies will keep people, increasingly to vouch for output rather than produce it. The tail of any real process is where models fail unpredictably and where being wrong is expensive, and a model cannot put its name to anything. Liability needs a person. Keeping one there is the right call.
The worker whose job shifts from making the thing to vouching for it.
The role is rational to keep, and harder to staff well than it looks. Checking is a different skill from doing, and it tends to erode when the underlying craft is never built. A layer of reviewers who never learned to make the work is a control that weakens over time, not by anyone's intent but by how the role is structured.
Decide adoption at the task, not the job, and re-decide it every quarter.
The choice is no longer whether to adopt AI for a role, but which tasks, priced one at a time and re-priced as the models get cheaper. For each task a manager routes to whatever produces acceptable work at lower cost, model or person, and runs the sum again the following quarter. The freelance market ran this experiment in public first: after ChatGPT, writing postings on Upwork fell by roughly a third and the more automatable coding work by about a fifth, while freelancers who folded AI into their offer commanded higher rates (Hui & Reshef, 2024). The floor fell and the ceiling rose. The same logic is now moving inside the firm.
Anyone whose work sits just above a cheap model's reach. Their wage now tracks the falling price of the next-best substitute.
The split this produces does not run where most people expect. Not young against old, not technical against non-technical. It runs between those who can do what the frontier cannot and those who cannot.
Announce broadly. Restructure slowly.
Markets reward visible AI motion and treat the appearance of falling behind as a risk, while real redesign is slow, costly and hard. So the rational move is to announce broadly and restructure carefully, which usually means slowly. The data shows how common this already is: two-thirds of enterprises report productivity gains (Deloitte, 2026), a third say they are deeply changing how the work is done, and about a quarter of AI projects clear their own return-on-investment bar (IBM, 2026). The incentive is not to fake transformation. It is that looking transformed is cheaper and faster than being transformed.
The workforce handed tools on top of a process no one redesigned.
The cost lands on the company too. Layering AI onto unchanged work does not remove effort; it adds a system to feed, and is felt on the floor as more work. Over a few years, the firms that redesign before they deploy tend to separate from the ones that deployed for the market, because redesign is where the return actually lives.
A small core that orchestrates, an automated middle, a thin edge that checks.
When returns concentrate in the people who can direct AI well, the efficient shape of an organisation becomes a barbell: weight at the orchestrating core, weight at the verifying edge, and as little as possible in between. Companies will build toward this shape because the economics point there.
The professional middle: most automatable, least protected.
The middle was also the career. It was the route by which the edge became the core. Thinned out, it leaves a company that runs efficiently for a decade and then finds the path that once produced its senior people is no longer there. That is a consequence of the shape, not a decision anyone makes deliberately.
Keep a human who can still do the work, as insurance.
This one runs the other way, and it is the most durable reason left to keep certain people. Frontier capability is now a geopolitical variable, not only a commercial one. In June 2026 a US export-control directive forced Anthropic to switch off its two most capable models for every customer three days after launch, because foreign nationals could not be excluded any other way (June 2026). Chips, sanctions and regulation all sit outside a company's control, and a firm that has fully removed a capability has no fallback when the model it depends on goes dark.
A few people kept as insurance against the model going dark.
"Keep someone who can still do this without the model" becomes a line in the risk register. It is the most defensible reason to retain people through this decade. It is also, honestly, conditional, defensive, and small.
Manage the people who remain against a metric built for the machine.
When the return on AI is hard to prove, the pressure to show it lands on measurement. A system whose payoff cannot be demonstrated gets justified by measuring the people who use it, more closely than before. The instinct is reasonable: you manage what you can see.
The worker, now measured on what is easy to count rather than what matters.
The hazard is old and well understood. What gets measured gets optimised, and what gets optimised is what survives. The tacit, uncounted, relational work that holds a team together depreciates by default rather than by intent, because nothing in the system records it.
Where this goes, within logic.
Quiet, and mostly invisible.
The freeze and the optics dominate. Little visible displacement, a good deal of quiet intensification, and a thinning at the entry level that no quarterly number captures. Most companies adopt without redesigning, and experience the cost as friction rather than gain.
The deferred cost arrives.
The mid-level cohort that was never hired is now conspicuously absent. The roles that were kept have changed shape through verification and measurement; the barbell has thinned the structure. Companies find they automated the rungs their own people needed to climb.
Two outcomes, and the choice is being made now.
In one, companies used the decade to redesign deliberately, and to protect the judgement, the relationships and the tacit craft the machine does not understand. In the other, the gap was filled by default, what was countable got optimised, and the result is an organisation that runs well until the people who could actually run it are no longer there. Which outcome arrives is not a forecast. It is a choice, being made in roadmaps now.
If you are the one deciding.
Most of these moves are defensible on their own terms. No one decides to hollow out a workforce. The danger is that a hundred rational decisions, each sound in isolation and made under pressure, aggregate into an outcome nobody would have chosen on purpose. Naming the aggregate is the first act of controlling it.
The choice is not whether to fill the gap. It is who fills it. Decide it deliberately, or it settles by drift, and drift settles on the people with the least standing to object.
The cheapest role to cut today is often the one that produced your senior people. Seniority is built by doing the work, being corrected, and absorbing the unwritten rules of a craft, on exactly the rungs these choices remove. Nobody apprentices into a step that no longer exists, and nobody verifies what they never learned to make. Protecting the bottom of the ladder is not sentiment. It is how an organisation renews itself.
The technology will arrive as it always does, and slower than the narrative says. The companies that come through this decade in good shape will be the ones honest enough to choose who fills the gap on purpose, and disciplined enough to protect the part of the work the machine still cannot reach.