FIELD NOTE 27  ·  JULY 2026  ·  7 MIN

What to keep in-house now that AI does the first draft

AI has taken over most of the work that used to justify a junior hire or a small consulting contract. Some of what is left cannot be bought at any price. Some of it should never have been done inside the building anyway.

I should declare an interest before I start. I run a consultancy. A fair part of what follows argues against hiring one, including this one.

The work that used to fill a junior's first two years has largely gone. The first draft of a report. The deck for Thursday. The summary of what four competitors are doing. The section of code that is nearly the same in every project. A machine now produces all of that in minutes, at a standard somewhere between mediocre and quite good, which is roughly the standard it was produced at before.

What survives is harder to name, because it never had its own line on an invoice.

The deciding part did not get automated

A machine will give you five options with a fair summary of each, and it will do that well. What it will not do is pick one and stand behind it in front of your board, knowing the two things about your business that quietly rule out option three. Analysing and deciding were always separate jobs. Only one of them has become cheap.

This is why the volume of strategy material keeps rising while the quality of decisions stays where it was. When producing a document costs nothing, more documents get produced. The scarce thing was never the document.

Ownership means somebody carries it when it goes wrong

Ask who takes the phone call when a decision turns out badly. If the answer is a supplier, you have bought a defence rather than an outcome. If the answer is one of your own people, that work belongs inside, whatever it costs, and no software changes that.

I have watched organisations outsource a decision and keep the consequence. It is the worst arrangement available and it is extremely common, because on a spreadsheet it looks like efficiency.

Context is the part nobody has ever written down

The reason you stopped selling to that customer. The reason invoices go out on the twenty-fourth rather than at month end. The two suppliers who cannot be in the same meeting. None of that exists in a document, which means no tool can reach it, and an outsider learns it slowly and then takes it with them when the contract ends.

Anything that depends on that knowledge stays with your own people. It is also why our first two weeks on any job are spent asking questions rather than building. We are buying context we do not have, and we are honest that we are starting behind your own staff on it.

What is safe to buy in

  • Work you can describe precisely enough to judge when it comes back. If you cannot tell good from bad on delivery, do not buy it from anyone.
  • Work you need for eight weeks and never again, where hiring would leave you with a person and no role.
  • Specialist judgement that is genuinely rare, where being wrong is expensive and nobody on your team has done it before.
  • An outside opinion on something your own people cannot say out loud in the building. That is often worth more than the analysis attached to it.

Notice what is missing from that list. Producing material is missing. If you are paying an outside firm mainly for documents, work out what those documents now cost to make, then look at the invoice again.

The junior problem, which nobody has solved

Here is the part I find genuinely difficult. If a machine does the work juniors used to learn on, where do the people with judgement come from in ten years? Judgement is not taught in a course. It is accumulated by being wrong in situations that do not matter much, under somebody's supervision, and we have just automated the situations that do not matter much.

My own answer, held loosely, is to keep hiring them and change what you ask of them. Hand them the machine's output and make them responsible for finding what is wrong with it. Pulling apart plausible work teaches faster than producing mediocre work ever did, and it builds the habit of not trusting an answer because it reads well. I am aware that is easier to write than to run, and that the pressure on most managers is to skip the hire and pocket the saving.

If you are trying to place a specific piece of work, the test I use is short. Could you tell, on receipt, that it had been done badly? Would you be the one answering for it if it was? If the first answer is no and the second is yes, that work is not safe to buy from anybody, us included, and I would rather say so now than after you have signed something.

Working through this on a live programme?

A 45-minute call with the engineer who would run the work. We will tell you whether AI is the answer, including when it is not.