If AI writes the code, why do you still need a head of product?

A founder asked me this last month, more or less in those words. He’d read that AI would replace his developers, his CTO had told him the team was already using coding agents, and he wanted to know what, exactly, he’d be paying a product person for.

It’s a fair question. Here’s the honest answer.

Building got cheap. Deciding didn’t.

For twenty years the expensive part of software was building it. Engineers were scarce, so the whole discipline of product management grew up around not wasting them: prioritise carefully, write it down, argue about scope, because every week of engineering time was precious.

That constraint is loosening fast. A coding agent given a clear spec will produce working software in hours. Not perfect software, and not without review, but a real, running thing. The cost of a feature is falling toward the cost of describing it properly.

Which means the bottleneck has moved. It’s no longer “can we build this?” It’s “should we, and what exactly is it?” And most companies haven’t noticed the move, because their process still treats engineering time as the scarce thing and product thinking as overhead.

AI builds whatever the spec says

The uncomfortable corollary: a vague spec now gets you the wrong product faster.

When a human engineer got a one-line ticket, they’d come and ask. They’d push back, suggest a simpler version, notice that the request contradicted something shipped last month. That conversation was slow, and it was also where a lot of the product judgement in a small company actually happened.

An agent doesn’t come and ask. It builds the most plausible reading of what it was given, confidently and quickly. If the brief was a guess, you get a well-built guess. Then it comes back in review, and the “speed” you bought turns into rework.

I’ve watched this happen in teams that adopted the tools without changing anything else. Build time dropped. Review-and-redo time went up. Net, they were busier, not faster, and nobody could tell you which features had moved a business number.

So what’s the job now?

Three things, and none of them is “writing tickets”.

Deciding what matters. Someone has to turn forty things the business could do into the three it will do this quarter, say no to the rest in writing, and own that call when sales, customers and the board all push back. That was always the job. It’s just more visible now that the building part is fast.

Specifying well enough that the fast thing builds the right thing. The spec is now the product, in a very literal way: it’s the artefact everything downstream is generated from. Outcome, constraints, what’s out of scope, how we’ll know it worked. A page, not a novel. Good enough that an engineer, or an agent, can start without a meeting.

Judging the result. Did it move the number? Should we do more of it, or stop? AI can generate options all day. It can’t tell you which one your customers will pay for, or that the elegant solution is three times bigger than the business needs. That’s judgement, and judgement is what you’re hiring.

Where I’d push back on my own argument

If you’re pre-product with two freelancers, you don’t need a head of product. You need to talk to customers and build something small. Come back when there’s a team and some revenue to protect.

And if you have a CTO who genuinely enjoys making product calls and is good at it, you may not need one either. The problem isn’t engineers making product decisions; it’s product decisions being made by default, by whoever’s nearest, with nobody accountable for the outcome.

What to do this week

Three cheap checks:

1. Ask three people for the top three priorities. If you get three different lists, decisions aren’t being made; they’re being assumed.
2. Look at the last five things that shipped. Can anyone say what business number each one was meant to move, and whether it did?
3. Pick a recent feature that came back for rework. Read the original ticket. Would you have built the right thing from it?

If those come back badly, the fix isn’t more engineers and it isn’t more AI. It’s someone owning the decisions and the specs. Whether that’s a full-time hire, a day a week of someone senior, or your CTO with a clearer mandate depends on the company. But it’s a seat, and right now it’s probably empty.