Main idea

A model doesn't understand a market the way someone living in it does, doesn't anticipate a need, doesn't defend a vision in front of a sponsor, doesn't hold a NO GO and doesn't answer for its calls six months later. It works very well as a personal assistant: it picks up thinking already under way, swallows the sources, does the grinding work.

What can't be delegated is the commitment. When a startup has to decide whether it goes all in or dies, the choice belongs to the people who founded the company and who will carry what follows. A machine can work the option up, produce the scenarios and cost them out; it cannot stand in front of others and answer for the result.

Analyzing and owning are two distinct acts. The second commits someone over time, in front of identifiable people.

Layer added by "Why You Should Audit Your Skills: Master the Hidden Token Limits" (2026-03-25). The same line reappears on a far more modest scale than an executive decision. A token audit report establishes that one step rereads a hundred and forty analyses to use only two, and costs out what stopping would save. It doesn't say whether revalidating everything every month was waste or precaution: that answer depends on what you accept missing, and it belongs to whoever will answer for it. Working up a cost and deciding whether it is justified remain two distinct acts, even when the first is fully quantified.

Layer added by "PM, Developers, and AI: Roles Are Blurring, Responsibilities Remain" (2026-08-01). The line reappears at the heart of a tooled-up Product Manager's work, and takes its sharpest form there. The role becomes decisive exactly where no option is good: two or three bad solutions, different costs, different risks, different political effects and different customer impacts. A model can work up each of those options, compare the scenarios and cost out the consequences; what it doesn't do is settle the matter and carry the choice in front of the team, the committee and the customers.

Why it matters

This draws a usable line inside an organization that automates: what can be worked up by a system, and what has to stay signed by a person.

And it avoids a widespread misunderstanding in discussions about automating product roles, where a critique of a division of labor is read as an announcement of replacement.

Nuances and limits

A person's commitment can be an organizational fiction: an executive who approves a recommendation produced by a system without being able to contest it only owns it in appearance.

And the boundary moves with the stakes: for a reversible, low-cost decision, demanding a human commitment adds slowness for nothing.

Open questions

• Above what level of stakes does a person's signature add something the worked-up case doesn't already provide?