Do you train people to use AI agents, or build the agents?

I wrote this question in a notebook in Oxford in May 2025 and could not do anything with it on my own. That's where Saeed stepped in.

If an AI agent can do the job, why would you train a person to do it?

It is not a comfortable question for someone who sells training software. I left it alone for four months.

Then on 14 September 2025 I was in Doha with Saeed, my co-founder, for the kick-off days of the Google for Startups accelerator. We went at this for most of a day. What came out of that conversation became our roadmap, so it is worth writing down properly.

The version of the problem that keeps you up

Most corporate training exists because a person needs to perform a task to a standard. The policy has to be followed. The system has to be used properly. The customer has to be handled a particular way.

Now assume agents get genuinely good at a meaningful share of those tasks. Not all of them. A meaningful share.

In that world the training budget for those tasks does not shrink. It disappears. You do not train somebody to do a thing that is no longer done by a person.

Which leaves two businesses you could be in:

• Training the people who work alongside agents

• Building the agents

They look adjacent from the outside. They are not. The first assumes the human stays in the loop and needs to get better at their part. The second assumes the loop closes.

I went into that room expecting us to pick one.

Where the argument broke open

Saeed's objection was that the question is too clean, and he was right.

Work is not a list of tasks. It is judgment about which task matters, in a context where the rules conflict, the documentation is out of date, and the person who knew why it was done this way left eighteen months ago. Agents are getting good at execution. They are nowhere near good at knowing what an organisation actually meant.

So the thing that gets scarce is not people who can perform tasks. It is people who can direct work, check it, and be accountable for it when something else performed it. Reading an output and knowing it is subtly wrong. Deciding what to do when the model is confident and the situation is unusual.

And then the part that reorganised everything for us. If you write down what you would need in order to trust a new hire with a process, and separately write down what you would need in order to trust an agent with the same process, you get the same list.

What is the job. What does good look like. What are the edge cases. How would I know if it went wrong.

We started calling both of them executors, because a manager does not care which one is doing the work. They care whether it got done to the standard. The question underneath is identical either way: can this executor do this job to our standard, and can we prove it?

Two clocks, and I had been reading them as one

This is where I had been confusing myself, and it took the argument in Doha to see it.

The near clock. Right now, organisations cannot deploy agents, and the blocker is not capability. It is that nobody can demonstrate the thing works reliably enough to be accountable for it. At the same time their people are not ready either. Workera's 2026 enterprise benchmark, across more than 88,000 assessments at large enterprises and US federal agencies, found only 13% of employees rate as accomplished in agentic AI skills, the lowest of the fourteen capabilities they measured. Deloitte's enterprise survey has leaders naming insufficient worker skills as the single biggest barrier to getting AI into their workflows.

Those are two descriptions of one problem. Organisations cannot deploy agents because they cannot verify them, and because their people cannot yet work with them. Somebody has to learn to delegate to an executor, supervise it, and check its output. Nobody is born knowing how to do that. It is teachable, and it has a budget line today.

The far clock. If agents do get genuinely good at whole categories of work, the training for those categories does not shrink. It disappears. That is the question I actually wrote in the notebook, and I want to be honest that the near-clock evidence does not answer it. Skills gaps that are urgent in 2026 tell you nothing reliable about 2031.

I spent months treating the first as though it settled the second. It does not. It buys time to find out.

The objection I have to answer

Here is the thing I have to say out loud, because it occurred to me on the flight back and has not gone away.

This is the most convenient possible conclusion for a man who already sells training software. We asked whether we were in the wrong business, argued for a day, and decided the question was malformed. That is exactly what you would expect us to decide.

So the test cannot be whether the answer feels right. It has to be whether it makes a claim that could turn out to be wrong.

It does, and it is the one in the middle of this essay. If the list you need to trust a new hire and the list you need to trust an agent turn out to be different lists, we are wrong. Not partly wrong. Wrong in a way that means the two businesses really were separate and we should have picked one.

I think they are the same list. I have thought so for eleven months and looked for counterexamples. But I would rather write down the thing that would disprove me than pretend the argument was harder than it was.

What we decided

We did not pick one of the two businesses, because we came to think the choice was badly posed.

To train somebody well, you have to understand the work: what the company knows, how a task should be done, where people reliably get it wrong. To hand that same work to an agent and trust the result, you need exactly the same understanding.

The understanding is the asset. Whether it comes out as a lesson for a person or an instruction and a check for a system is downstream of what the situation calls for.

On the near clock that is a product people will pay for now. On the far clock it is the only position I can find that does not require us to guess correctly about a date. Both of those matter, and they are not the same argument.

There is more to say about what that actually looks like when you build it. I will write that one when we have shipped it, and not before.

The question was written in the Radcliffe Camera, Oxford, May 2025. The argument that answered it happened in Doha on 14 September 2025, with my co-founder Saeed.