Four functions, 175 people, one transformation. I ran it, and I built in it.

I ran it across four functions.

I owned the transformation for the whole company, across 175 people and the vendors we worked through.

I built the prototypes myself and wrote the dashboards in section three. You cannot set the boundary on what a model is allowed to author unless you have watched it get things wrong in your own hands.

What it meant in each function.

Two things came out of it. Every function's read of the same week landed in one place, which is section three. And we went from one production build every two weeks to two a week.

Product is where it went deepest. The rest of this page is that.

The prototype replaced the PRD, and became the thing we argued against.

Generative tools inflate a requirements document. Every model I pointed at a PRD came back with more rules, more edge cases, more requirements, and no ranking. The document grew and the decisions inside it got harder to find.

So I moved the source of truth to something bounded. One HTML file, real logic underneath, built against components pulled from our production code. If a thing has a state, it goes in. If it cannot be seen, it does not get built. That constraint does the scoping a PRD was supposed to do, and it does it visually, so a disagreement gets settled in the room.

Here is where it sits in the flow a product manager actually runs.

Too thin and it is a picture. Too thick and you have built the product before deciding whether to build it. We stopped writing PRDs.

The premium subscription is the test I use on all of it. I got direction on the 11th. It was live on mobile and web on the 31st, and it held. Twenty days, both platforms. The distance between a decision and a shipped product collapsed, and I still own the number that says whether shipping it was right.

I built the read, so the schedule stopped living in a Slack reply.

Before this, the answer to whether we were going to make it lived in four places. A ticket nobody had updated. A meeting you had to be in. A Slack reply you had to have seen. And somebody's read of the room, which is the one that gets reported upward.

Four subjective accounts of the same schedule, and the most confident one wins.

So I built the read. It runs on its own and it answers four questions without anyone being asked.

Everyone gets the same read, at the same time, before the conversation.

Four decisions never left my desk.

One question sorts everything: if this comes back wrong, how long until I find out? What matters is whether I catch it today or next quarter.

Anything checked the same day got handed over. Prototyping, converging the week, rebuilding the backlog, reading the market.

Mine · stays wrong quietly

• Whether we should build it at all.
• Whether what worked somewhere else will work here.
• Which number would settle the argument.
• Who answers for it when it is wrong.

Those four come back plausible, get agreed to in a room, and the bill arrives two quarters later. A Harvard trial of 758 BCG consultants measured exactly this. On the one task built to need hard data read against a soft signal, which is the shape of every product decision, the people using AI landed 19 points further from the right answer, and graders scored their work higher while it was wrong.

The second boundary is written down. What can be agent-authored, what can be agent-reviewed, and what needs a human on both ends is set by a risk tier I own. Payments, auth and anything touching session state sit at the top. Section six is why.

It designed onboarding for a brand we had not earned.

Asked to rebuild onboarding, it came back with an interest picker and suggested follows, both up front, so the feed could personalise from the first session. The instinct is right. Ask people what they like and the first moment of value arrives sooner.

It was reasoning from TikTok and Instagram. Those apps can open with a questionnaire because you already decided to be there. We were emerging. Every screen before a new user saw anything worth seeing was a place to lose them, and losing them there looks like nothing. No error, no complaint, a smaller number the following month.

So I kept the idea and moved it. Into the product in under four clicks, then ask about interests once they are inside and have a reason to answer. 60% of active users had told us what they liked by their second session, inside 72 hours. The answers were better too, because by then people knew what they were answering for.

Nine days to find one bug, and no test could have caught it.

We shipped a change that passed every test and broke session management. It issued duplicate tokens. Users got logged out and had to sign back in.

The failure only appears when a session expires, and ours expire after seven days. The first report came in on day seven. Two more days to establish it was not user error and find the cause. Then a hardening sprint.

When an agent writes the code and the tests, the tests verify what was written. Nobody has checked them against what was required. That circular verification is invisible on every dashboard you have. And a test suite has a time horizon. Ours could not see past a single run, and the bug lived seven days out.

Time-dependent state moved to the top risk tier that same week. Human-authored, human-reviewed, with tests derived from the acceptance criteria rather than from the implementation.

It is the best asset I have ever managed.

AI took the work that does not need me and handed back the work that does. A decision reaches a shipped product in twenty days instead of a quarter, and nobody spends Monday assembling what happened last week.

What it did not do is take anything off my desk that I answer for. A model has no stake in the outcome, cannot be held to the number, and will not be in the room in two quarters when the bill arrives.

One thing I have not solved. The path from junior to senior used to run straight through writing the boilerplate an agent now writes. My juniors are getting better at judgment faster than any cohort I have managed, and I do not yet know where they get the depth that only comes from having built the thing yourself.