What is AI rebounding?

AI rebounding is the name SHL, a talent assessment company, gave on 29 October 2026 to a pattern its survey found in the United States and Britain: a role is cut because AI will do it, and months later the same employer asks the same people back. The term is SHL’s, and the Associated Press calls the same thing the AI boomerang. It matters because it is the moment the cost of automating a role before anyone assessed the human capability inside it comes due, usually at a higher price. This page sets out what the survey says and does not, the cases on the record, why the pattern is predictable, and what an organisation can decide before it has to buy a capability back.

What SHL found#

SHL’s release of 29 October rests on a survey by Propeller Insights of 2,036 employed adults, 1,029 in the United States and 1,007 in the United Kingdom, taken in June and July 2026. Just over half, 52 per cent, said they had encountered employees replaced by AI and then asked back; 19 per cent of the American respondents and 13 per cent of the British said it had happened to them. Asked what it would take to return, 51 per cent said a substantial pay increase, 42 per cent a guarantee of job security and 40 per cent “full transparency” about how AI would be used in the role. Lucy Beaumont, SHL’s senior vice-president of product and design, defined the term as what happens when a company automates a role before it has assessed the human capability inside it. Ginger Christ at HR Dive reported the release on 5 November alongside Gartner’s prediction that by 2029, 30 per cent of workers terminated or replaced by AI will be rehired, often at higher cost.

Three cautions attach. The survey records workers’ beliefs, not employers’ records; the company that commissioned it sells the assessments it recommends; and “encountered” includes hearsay, so the 19 and 13 per cent are the figures to carry.

The cases on the record#

Cathy Bussewitz at the Associated Press reported on 5 November that the reversals happen “often enough to acquire a nickname: the ‘AI boomerang’”. Her first case is Ford, which said in June that it had hired or promoted 350 veteran engineers after AI tools deployed for quality control “did not perform at the expected level”; a Ford spokesperson said that “AI utilization has been an important part of our quality gains” alongside the technical expertise of its teams. Forrester’s Betsy Summers gave the diagnosis in one sentence: “We don’t know all the work that’s required to do things well within the organization.” Each case is one company’s account of itself, and none has published the comparison that would show the rehiring was caused by the AI falling short rather than by demand returning.

Why it is predictable#

The pattern follows from the first principle of a review published on 28 October by Giuseppe Riva, Stefania La Rocca, Brenda Wiederhold and Andrea Gaggioli: “assisted performance differs from retained competence”. A role automated on the evidence of a pilot is automated on assisted performance, measured while the people who knew the work were still in the building to catch what the tool missed. Remove them and the assistance runs unassisted. Riva and colleagues call the stock that was removed cognitive capital, “the accumulated, renewable and depletable stock of knowledge, cognitive and metacognitive capabilities, and practices” that lets a person or organisation judge and adapt, and list “substitution of practice” as the first of four ways AI draws it down. A model posted on 4 November by Moritz Diener and colleagues at KIT and IBM Research shows the same thing from inside one worker: “experts may lose skill, even as AI-assisted outcome accuracy remains high”. The accuracy on the dashboard holds while the capability beneath it goes. Rebounding is what happens when the dashboard is finally wrong about something that matters.

The checking was never free either. Founder Reports’ survey of 2,052 American workers who use AI weekly, published on 4 November, found that among those who had trained a new employee, 60 per cent said overseeing and fixing AI’s work takes at least as much effort as training a person and one in three said more; Christiane Schroeter of Cal Poly put it as “You’re not saving any cost; you’re simply delaying it.” The business case counted the salary and not the supervision.

How it sits beside capability debt and experience debt#

Capability debt is the gap between what an organisation’s people can do with their tools and what they can still do without them, a term others also use and nobody on this site claims. Experience debt, Dayforce’s term of 26 October, is the same gap seen from the person who never got the repetitions. AI rebounding is the debt being called: the point at which an organisation discovers the gap by needing the capability and finding it has to buy it back from the people it let go, on their terms. SHL’s survey puts those terms at a substantial rise, a security guarantee and transparency about the AI, from half or more of the people asked. The three words describe one mechanism at three moments: the repetitions stop, the capability thins, the bill arrives.

What an organisation can decide beforehand#

Beaumont’s definition contains the remedy: assess the human capability inside the role before automating it. In practice that means writing down, for each role to be cut, what the person did that the pilot never measured, usually the exceptions, the judgement about when the tool is wrong and the relationships, then naming who does those things afterwards and how anyone would know if they stopped. The Productivity Institute’s report of 30 October, as Dan Robinson at The Register reported it, found coding activity up about 240 per cent where software releases rose about 30 per cent, “due to humans still needing to check and test their work before release”, which is the supervision line the business case should carry. A board asked to approve savings from a role can ask three things: what the measured performance was measured with the people still present; who holds the exceptions once they have gone; and what it would cost to hire the capability back at the terms SHL’s respondents named. The position of this research is that the risk that can be measured sits in the handover of decisions to machines and in what happens to human judgement afterwards, and that the decision an organisation controls is which decisions a machine may make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. Those questions are set out at Rules Before Tools.

What this does not show#

It does not show how common rebounding is. The only figures are self-reported by workers, in a survey commissioned by a company that sells the remedy, and the AP says plainly that it is unclear how common the reversals are. It does not show that the rehired capability was the capability the AI lacked: a company may rehire because demand returned, because a tool was badly deployed rather than inadequate, or because of regulation. Ford’s case is Ford’s account. Gartner’s 30 per cent is a forecast. The reviews and models cited here explain why the pattern should occur; none of them measured it in a firm that cut a role and bought it back. And the term is a fortnight old. If it survives, it will be because organisations find it describes something they recognise, not because a survey counted it.