They replaced the developers. Now they want them back.

The interesting part of this story is not that companies got it wrong. It is where they got it wrong — consistently, across industries, at roughly the same point in every job. That pattern is worth understanding before you make the same decision about your own website, your own support inbox, or your own development work.

The reversal, in the numbers

These are survey and projection figures from business research firms, not audited accounts. They describe what employers report about their own decisions, which is useful for the direction and shape of the trend and unreliable to the decimal.

The 60/40 pattern

The recurring shape is that AI absorbs most of a role and stalls on the rest. IBM automated 94% of routine HR requests and found the remaining 6% contained the ethical judgements, the exceptions and the cases where being wrong was expensive — then announced it would triple US entry-level hiring. Ford has been rehiring experienced engineers for quality problems automation could not resolve.

The last stretch of a job is not the same kind of work as the first stretch. It is where ambiguity lives, where the edge cases are, and where somebody has to decide what the right answer is rather than produce it. Cutting the people who handle it and keeping the tool that handles everything else leaves you with a process that is 94% finished and 0% complete.

Why it costs more to undo than it saved

Returning staff are coming back on 20–35% more than they left on, because the employer is now buying back a specific thing it already had: knowledge of a system nobody documented. The salary is the visible part. The invisible part is the months between the decision and its reversal, during which the work was either done badly or not done at all.

This is the same arithmetic as any cheap rebuild. The saving is immediate and legible; the cost arrives later, in a different budget line, and is rarely attributed to the decision that caused it. We wrote about the website version of exactly this in what an AI-generated website actually gets you.

What this means if you are buying, not hiring

Most small businesses are not deciding whether to lay off a development team. They are deciding whether to buy a €49 AI website builder or hire someone. The reversal is directly relevant, because it is the same bet at a different scale: the tool will get you most of the way, and the remaining stretch is the part that determines whether the thing works.

The honest version is that AI has genuinely changed what the first 60% costs — that part is faster and cheaper than it was, and pretending otherwise would be selling you something. What it has not changed is who can finish. If you are weighing this, choosing an agency covers what to actually look for.

The one question worth asking

Before automating anything, ask what happens to the cases the automation cannot handle — not whether they exist, but specifically who deals with them and how they find out. If the answer is "we will see", you have not automated the job, you have deferred the hard part of it into a queue nobody owns.

That question is also how you tell a serious AI proposal from an unserious one. Anyone quoting you automation work should be able to say what falls out of the process and where it goes. Ours is in AI automation, and agents versus chatbots explains which tool suits which job.

Sources

Checked September 2026. Survey and projection data — directional, and self-reported by employers.

• Inc. — 55 percent of leaders regret AI layoffs, and a major hiring reversal has begun ↗

  The 55% figure, from Forrester’s 2026 Future of Work research, plus the Gartner projection that about half of AI-related cuts will be rehired within a year.

• CNBC — Employers who laid off workers for AI are reversing their decisions ↗

  Coverage of the boomerang-hiring pattern and the pay premium returning workers are commanding.

• Forbes — Companies fired workers for AI. Now they want them back ↗

  The IBM 94%/6% split and the Ford engineering rehires quoted in the table and the first section.