When the work gets cheap

I have been working with the machinery underneath AI for most of twenty years: parsing and extraction, anomaly detection, natural-language generation, screening for fraud, long before any of it was called AI. For most of that time the hard part of my job was judging the lag, the time between a technology becoming possible and it changing a company's profits. The technology usually arrived when I expected. It almost always took years longer than it looked to reach the accounts.

That lag has split in two. The technology lag has all but gone. The length of task a leading model can finish on its own has been doubling roughly every four to seven months, and the cost of finding out whether an idea works has fallen from a project to a few days. The organisational lag has barely moved. Most people now say AI has made them more productive. Far fewer businesses can find it in their profits. In McKinsey's August survey, about 80 per cent of respondents said AI had improved their productivity, and only 37 per cent put any effect on profit down to it.

That gap is where value gets created or lost, whether you are a start-up deciding who to hire, a corporate deciding what to automate, or an investor deciding what to pay for.

I should say where I stand. I'm optimistic. Every big wave of technology I know of has ended up creating more work than it removed, though rarely for the same people or on the same timescale, and I expect this one to do the same. Done well, it's the biggest opportunity I've seen in twenty years. What follows is mostly about the one cost that's easy to miss, because avoiding it is what turns the opportunity into value.

A business is worth its judgement

When doing the work gets cheap, what a business is worth moves to its judgement. I mean three plain things by that: people who know what good looks like, a written record of what good looks like, and a way of producing more people who know.

AI can give you an answer. It can't own one. The machines are getting better at judgement too, and will keep getting better. But someone still has to decide what good looks like, and be on the hook when it isn't.

What has changed is that AI touches all three at once, and the effects are hard to see from inside. None of that is a reason to slow down, but it does need a plan.

What to plan for

It removes the work people learned from. Most organisations are a pyramid, with lots of people doing the repeatable work at the bottom and a few making the calls at the top. AI takes the bottom out first. Plenty of people have drawn that. What I keep coming back to is that the people in the middle got there by doing the work at the bottom: the first-pass reviews, the month-end checks, the draft someone senior tore up. It was never designed as a training programme, but it worked as one.

If AI just makes people faster, junior hiring dips for a while and the shape comes back. If it changes how people learn, the shape doesn't go back. Friebel, Huang, Li, Shukla and Zhang draw exactly that line in an April paper, between a productivity shock and a learning shock. They don't say which one AI is. I think it's the second.

A start-up meets the same question from the other end. It has no pyramid yet, so it gets to design one. If agents do the junior work from day one, the question is who the first ten hires learn from.

It makes supervision harder the better it gets. The plan in most businesses is that people move up to supervising the machines. Two well-established findings say that plan is optimistic. Once a person is past about seventy per cent busy, the exceptions waiting for them pile up faster than they can clear them, and most plans staff supervisors towards a hundred. And the more reliable the automation, the worse people get at stepping in when it fails, because they no longer do the work often enough to stay sharp. So better agents can make worse supervisors, and most plans assume one person can run far more agents than they can.

It tempts you to let software act before anyone has written down the standard. This year the big vendors shipped the machinery for software that acts on a business's behalf. Each agent gets an identity, every action gets checked against a policy, and there are approval steps, records of what happened and tools for testing. All of it needs filling in. Someone has to name the person accountable, set the rule for what the agent may touch, and pick the cases from your own files that show what a right answer looks like. The vendors will enforce whatever standard you give them. They can't write it, because it's made of your cases and your people's judgement about them.

Why it's easy to miss

The saving from automating junior work shows up this year. The cost shows up in five, as a missing layer of experienced people, in a different budget and often under a different chief executive. No single report shows both.

A Chicago model of automation and careers (Afrouzi, Blanco, Drenik and Hurst, 2026) finds that the same technology can lead to two quite different places. In one, people keep learning. In the other, which the authors call a human capital trap, they don't, and cheaper technology pushes further into it. Asking businesses to behave better won't be enough on its own. It has to be designed into how the work is done.

What the public numbers already show

It is starting to show, if you know where to look. In June the government published a snapshot of UK entry-level hiring, built from LinkedIn data. Overall, entry-level hiring is falling in step with the wider market, where hiring is down about 14 per cent on a year earlier. But the weakness is concentrated in the junior tier of information-processing roles, where AI has improved most: entry-level hiring of accountants was down 29 per cent, and of software engineers 27 per cent.

The authors are careful to say this is not yet causal evidence of AI's effect. They are right to be careful. But the cost of removing the work people learn from will not be provable in any dataset for years, and by then it will have happened.

The big firms are already adapting, which is encouraging. In September, KPMG's UK chief people officer told City AM that graduates "will become more involved in reviewing and exercising judgement rather than simply producing outputs". I understand why. But you learn to review work by having done it. Moving graduates straight to judgement is the right aim, but it takes away the lower rungs of the ladder unless someone decides, on purpose, which work they still do by hand.

How I'd go about it

Start with the job, not the technology. I look at every piece of work the way a product person would: what is the job, what pain does it remove, what gain does it create, and what happens five minutes before and five minutes after it. That last question is the one people skip, and it changes the answer more than the others. Take a contract renewal: the work might be four days and the whole thing seventy. The prize is the waiting, not the work. In programmes I've run, judgement-heavy work barely moved while repeatable work improved several times as much. Point AI at the wrong part of the job and you get the small number.

One workflow at a time. Take each one through the stages, with a decision to carry on or stop at every stage, and the stop criteria written down before the work starts. I have used the same five stages for years (ideation, design, execution, acceleration, scaling). AI makes every stage faster. It doesn't remove a gate, and it doesn't mean jumping straight to scale.

Keep the work that teaches. Before automating a junior task, ask what doing it taught as well as what it was worth. Automate what taught nothing. Where the task was how people learned, let the software draft and keep a person deciding, or keep a share of the volume for people and treat it as a training budget. Keeping one routine task in ten done by hand is a cost you can put a number on. A missing layer of managers in 2031 is not.

Write down what good looks like. Use your own cases, with the right answers marked by people qualified to judge them. Then let software earn the right to act against that standard: watching first, then drafting, then preparing, then acting. Each step up is granted on criteria set before anyone started measuring, with a named person accountable for what it does, and it steps back down automatically when its results drift. I call this earned autonomy. Versions of the ladder are common now (AWS publishes one, and Microsoft lets you give every agent a human sponsor). And someone who holds an agent's identity isn't the same as someone who is on the hook for what it does.

Keep supervision honest, and count only what reaches the accounts. Put the number of agents one person can responsibly run into the plan, and keep supervisors doing some of the work. Time saved isn't money until someone has redeployed the people to work that earns, or taken the cost out.

Own the layer underneath. Buy the models; they are the part that gets cheaper and more alike every year. Route each step to the cheapest one that passes your tests. Own the connected picture of your clients, your decisions and your standards, because that is the part no vendor can sell you.

What it's worth

Done this way, judgement becomes something you can show. That covers the record of decisions, the tests of what good looks like, and the evidence that software acting in your name is under control. It also covers a pipeline still producing experienced people. A buyer discounts what was claimed and pays for what was banked. When someone checks a business properly before buying it, that's the difference between a story and evidence.

It isn't only defence. Once you can prove the quality of your work, the proof itself is worth something. First you get more capacity per person. Then you get quality you can prove. Then you get a service your clients will pay more for. That's the order I'd expect it to arrive in, in a start-up, a corporate or a portfolio company alike. Across several businesses at once, the approach is the same, with a few things shared at the centre; I've set that out in Across a portfolio.

Three questions I'd put to any leadership team

1. Which of your work is the bottom of the pyramid, what does it cost, and what did it teach?
2. Who decides what good looks like once the machines are doing the work, and is it written down?
3. Where will your next generation of experienced people come from?

In my experience most boards can answer the first. Very few can answer the third.

Where I could be wrong

Here is one way to check me. Thomson Reuters tracks how many associates US law firms have for every partner. If I'm right about how people learn, that number should peak and start to fall between 2029 and 2031, as the smaller 2026 and 2027 intakes fail to come through to mid-level. It reached a record at the end of 2025. If it's still climbing in 2031, I was wrong, and I'll say so here.