The Cheapest Place To Be Wrong

Two weeks ago I put something in the newsletter and then admitted I had not solved it.

The middle of the organisation, the layer that has been thinning everywhere, was where people learned to judge. They learned it through repetition and correction, on work that mattered but was not fatal, and not only on the work itself but on running a workstream, managing people and holding a client relationship. If AI removes those repetitions as a by-product of doing the job well, then they have to be created deliberately, and as I said in the newsletter, I was not convinced any of us knew how.

A number of people replied, so thank you to all of you! Most agreed, and despite a few good efforts, none of them really had an answer either. So I've dug a bit deeper into it and ended up somewhere I did not expect, which is that I had filed the problem under the wrong heading entirely from the start.

The straight line decline everybody is drawing may not exist

The assumption underneath almost every conversation I have about this is that AI takes work, work becoming automated takes jobs, and the junior jobs go first. It is a tidy line and the evidence does not draw it.

In its June quarter, TCS reported annualised AI services revenue of $2.6bn, up 13.6% on the quarter before. In those same three months it added 9,300 people, its strongest quarterly hiring in more than 3 years, taking its workforce past 593,000. That is a company growing its AI business and its headcount at the same time, and it said openly that AI work is project-based and non-recurring, so the revenue is volatile. Whatever is happening there, it is not substitution.

Then in June, Bloomberg Economics published an analysis arguing that AI is being wrongly blamed for Britain's job losses. Vacancies in the roles most exposed to AI were already falling before ChatGPT was released at the end of 2022. They fell further afterwards, and they have been rising again since the summer of 2024. In the payroll data the relationship was weaker still, with little evidence that AI is driving employment decisions and the number of private-sector workers in the most exposed sectors actually higher than before ChatGPT arrived. What businesses themselves pointed at was a subdued economy and higher employment costs after the rise in payroll taxes and the minimum wage.

I want to be careful here, because that is not the same as saying nobody has lost a job to AI. It is saying the story ran a long way ahead of the data, and a lot of people absorbed the consequences of a narrative that was not carrying its own weight.

I nearly published a version of that narrative myself

In July I had a figure in front of me: about half of the companies that swapped people for AI end up rehiring, and at greater cost than keeping the original workforce would have been.

When I went back to the sources again, no study said it quite that clearly. Forrester had forecast that half of AI layoffs would be reversed in some form by the end of 2026, and somewhere between the forecast and the coverage of the forecast, a prediction became a finding. The real numbers are more modest and more useful: Robert Half found roughly a third of hiring managers who cut a role for AI later rehired for the same or a similar one, and Careerminds found around half of those who rehired did so inside six months.

So the argument was never really about headcount

So you can hold your headcount flat, or grow it as TCS did, and still stop making senior people. The repetitions do not disappear because the roles disappear, they disappear because the work that used to generate them, the first pass, the draft, the deliverable that was slightly wrong, the paragraph that got sent back, is now produced by something that does not need to learn from being corrected. You can keep every junior on the payroll and still remove the mechanism that turned them into the person you would put into the senior role.

That is a harder problem than a hiring plan. Judgement is not knowledge, so you cannot hand it over in a training module. It is a residue that builds up on someone who has been allowed to make calls and live with them.

The answer was already in the room

In the same newsletter I described how I have stopped planning one type of delivery and started planning two. There is a fast track, with lighter governance, speed as the default, a higher risk threshold and therefore controlled access to data. And there is a production track, with full review, high quality and a longer timeline. The client chooses which one they are buying, or buys both and knows what to expect from each. What makes the fast track safe is not harder review, it is blast radius: the work is bounded by design, which matters enormously when you are deploying agents.

I built that as a risk control. It is now somewhere in the business where a person can own a decision and be wrong without it being fatal. Bounded blast radius is not only how you deploy an agent safely, it is how you make a senior person, give them the space and the confidence follows.

A bounded track on its own is only half of it, though, and the other half is the part I think gets underestimated. Somebody experienced has to be close enough to say what was missed. Coaching matters more now, not less, and I would go further: it becomes one of the main ways people develop rather than something you get to when delivery allows. The exposure to detail that used to arrive on its own, from sitting near the work while it was being done, now has to be arranged deliberately by the people who already have it. I am also using the same bounded route to bring client teams into a programme, which is how you build their capability without carrying significant risk, and it is the honest answer to who owns the thing once we have gone.

The obvious risk is that the fast track is already carrying real delivery, and a training ground that is also a live commitment is not really a training ground. There are people going at this from other directions worth watching: I wrote recently about an AI-native accounting firm paying its junior staff to stop billing time and to go and integrate AI and map new processes instead, which is the same instinct applied to the incentive rather than the governance.

The question I would sit with

Forget the headcount for a moment, because it has been the loudest number in this debate and the least informative.

Where in your organisation can somebody own a decision and be wrong about it without it being fatal? Name the place. If you cannot name one, no amount of coaching will rescue it, because there is nothing for the coaching to work on. That is where your next generation of senior people was supposed to come from.

Sources

All figures as at 19 August 2026.

TCS, Q1 FY27 results, quarter ended 30 June 2026. Annualised AI services revenue of $2.6bn, up 13.6% quarter on quarter, and the company's own note that AI work is project-based and non-recurring. Results release. Headcount figures, a net addition of 9,300 and a workforce past 593,000, its strongest quarterly hiring in more than 3 years, reported by Business Today, 10 July 2026.

Bloomberg Economics on UK job losses. "AI Wrongly Blamed for Britain's Job Losses, Analysis Suggests", Bloomberg, 16 June 2026. Vacancy trends in AI-exposed roles before and after ChatGPT, the payroll data, and what businesses themselves pointed at.

Forrester, Predictions 2026. The forecast that half of AI-attributed layoffs would be reversed in some form by the end of 2026, reported by The Register, 29 October 2025.

Robert Half on rehiring. Roughly a third of US hiring managers who cut a role because of AI later filled the same or a similar position, reported by CNBC, 1 July 2026.

Careerminds, survey of 600 HR professionals, February 2026. Of the organisations that rehired for roles they had eliminated, around half did so within six months.

My own earlier pieces referenced above. The newsletter that set the question, and the two-track delivery model: Escape the Deck, Issue #4, 13 August 2026. The AI-native accounting firm paying its junior staff to stop billing time: Time Is a Weak Proxy for Value, 11 August 2026.