Right now, we ask AI to build things.

Build a button. Build a flow. Build a landing page with three sections and a CTA at the bottom. We sit there dictating output, screen by screen, component by component. We’ve gotten pretty good at it. The AI has gotten even better.

But we’re still the ones doing the translating.

We always start with an outcome in our heads, more conversions, faster onboarding, fewer support tickets, and then we spend half our energy guessing which output will get us there. Then we describe that output to the AI. And it builds it. Nicely, quickly, without complaint.

Imagine skipping that step.

Imagine opening Cursor or Claude Code and saying: “I want our new users to get through the onboarding flow 30 seconds faster than they do today.” And the AI gets to work. It looks at your current flow. It measures where people drop off. It tries three variants, A/B-tests them against each other, ships the winner to production and comes back with: “Done. We’re down to 2 minutes 14. You saved 38 seconds.”

You never touched the output. You worked with the outcome.

That’s a different relationship with the tool. Right now, the AI is an extremely capable intern waiting for your orders. In this scenario, it’s a colleague who understands what you’re trying to achieve, and figures out how on its own.

More examples, so this doesn’t stay abstract

“Increase MRR on our Pro plan by 15 percent next quarter.” The AI looks at churn, at upgrade flows, at the pricing page, at onboarding emails. It figures out that the weakest link is an upgrade prompt that fires at the wrong moment. It rewrites it. It moves it. It sends you a report.

“Cut payment-related support tickets in half.” The AI reads the last 500 tickets. It finds the pattern: people don’t understand the invoice. It redesigns the invoice. It writes a better error message in checkout. Tickets drop.

“Get more of our free users to invite a colleague.” The AI tries four different in-app prompts, two different email sequences, a new onboarding slide. It measures. It picks the winner. It tells you why.

You never sat around debating button colors or headlines. You sat around thinking about the business.

This requires something we don’t have yet

The AI needs to be able to measure. It needs access to your analytics, your product, your database, your users. It needs to be able to experiment without setting anything on fire. It needs to be able to roll back when something goes wrong. And it needs to explain itself afterwards, so you’re not left with a black box that just says “trust me bro”.

We’re not there yet. But we’re closer than most people think.

MCP servers are starting to give AI access to real systems. Browser agents can already navigate your tools. Evals mature every month. The technical building blocks are there. What’s missing is the orchestration, and it’s coming.

And when it arrives, it changes what a product manager does all day.

You don’t spend your time describing features. You spend your time deciding which outcomes matter. Which trade-offs you’ll accept. Which lines the AI must not cross. Which metrics actually measure the right thing, and which ones just look good in a dashboard.

It’s a job that looks a lot like the job PMs always claimed they had. Right up until they got it.