AI automates the easy half of the PM job

You open Linear on Monday, the tracker your team runs on, and the project update is already written. The status is pulled together from a week of messy activity. The blockers are flagged. The requirements are restated in clean prose out of a scatter of Slack threads and meeting notes. You wrote none of it. And what you feel is relief, because that writing was never the part of the job you wanted. You wanted the thinking. Now a tool does the drafting and leaves you the thinking. That’s the promise every PM has been sold about AI this year: it clears the clarity grunt-work off your plate and frees you for the valuable part.

Linear shipped this on June 18, and it isn’t a demo. The feature drafts status updates, restates requirements from unstructured input, and surfaces changed assumptions as risks. If you read no product news last month, here’s the whole of it: the routine “here’s where the project stands” documents a PM used to write by hand are now something an agent writes for you, inside a tool teams already pay for. The industry read is that this hands the PM back their time.

Producing clarity and judging it feel like the same job. As economic activities they’re opposites, and everything turns on which one agents make cheaper.

Producing clarity scales with agents. Point more agent capacity at status synthesis and you get more of it, faster and cheaper. That half of the promise is true, and it’s why the team really does move quicker.

Judging clarity scales the other way. Every artifact an agent produces is a plausible draft that might be subtly wrong. Someone has to decide whether it frames the right question and stays faithful to what the team actually agreed. Add an agent and you have not added judging capacity. You’ve added things that need judging. And that judgment funnels to one person: the PM who owns the decision and answers for it when the update misleads someone. You can spread the reading across a team. You can’t spread the accountability. So the faster the agents write, the more the judging piles onto the one person who can’t hand it off.

The work AI takes off your plate gets cheaper with every agent you add; the work it hands back gets heavier.

The code world already ran this experiment, and PMs can read the results. For two years, software teams have let AI write code while keeping humans to check it. The lesson came back blunt: the checking is the expensive part, and you can’t buy your way out of it by adding agents. Meta tried shipping AI-written code with no human review and it went wrong enough to become a warning story. PostHog, an analytics company that builds much of its own software with AI agents, published its rule in July: review as little code as possible. Only changes that touch security, architecture, or genuine complexity reach a human. The rest ships unread, on purpose.

That cut is the move the PM story is missing. The advice going around says the PM’s new job is to govern the scope and quality of what the agents produce. In plain terms: read the stream and judge it. Nobody has added the second half the code world had to learn, which is to judge less of it. There’s no risk filter, no rule that most agent-written clarity should ship without a human reading it. The PM version imported the part where agents make you faster and left behind the part that keeps the leftover human job from drowning you.

The fair objection is that checking is easier than creating. Reading a draft is lighter than facing a blank page, so surely AI makes the work lighter overall. For one document, that holds. Volume is where it breaks. Checking stays cheap only while there’s little of it and you can read all of it. At agent speed there’s more clarity produced than one accountable person can actually read, and each item is no easier to judge than before. There’s just more of it than fits. That’s where careful reading turns into rubber-stamping, and a wrong update, a clean status note that misstates what the team actually decided, ships because no one had time to catch it.

I’ll be honest that this is a prediction, not a measured result. The Linear ship is real and documented. The code-review lesson is documented too. The claim doing the real work here, that clarity behaves like code, cheap to generate and expensive to trust, is my inference, and it’s the one that could be wrong. Maybe judging a status update stays light enough that volume never bites. That is exactly the thing to measure, and the right measurement isn’t the tempting one. How much clarity your team now produces tells you nothing. Agents flatter that number by design. Two other numbers tell the truth: how long it takes you to approve agent-written clarity as its volume climbs, and how often a wrong one slips through once you can no longer read them all. If those rise faster than the volume, then “AI makes the PM faster” is the wrong sentence for what’s happening at your desk. The team got faster. Your own job got heavier and harder to share.

The encouraging part is that the fix is already written down, one domain over. Decide ahead of time which clarity earns your judgment, the decision-critical slice, and let the low-stakes stream ship unread. You can start building that discipline now, before the volume arrives to force it. The PMs who stay fast will be the ones who chose, on purpose, what to stop reading.