AI removes work. It doesn't remove judgment.

The honest answer about AI in product work is that it is useful for the parts of the job that are mechanical, and useless for the parts that are not.

It is great at the third draft, the boilerplate, the test fixture, the obvious refactor. It is bad at the call you make on day one, the call you make on day six, and the call you make at handoff. Those are judgment calls. They are what you are actually paying for.

Most of the bad output I have seen in AI-assisted product work falls into two patterns.

The first: AI doing what looks like work but is not. A junior engineer asks the model to draft a database schema for a booking app and gets six tables, two of which have no purpose, one of which subtly contradicts the others. Looks productive. Costs three days to unwind during the first real test. The AI did not know the actual requirements. The engineer did not interrogate the AI to discover that. The output looked like progress, so it shipped.

The second: AI substituting for taste. The agency promises Figma magic and gets it - thirty variants, all aesthetically plausible, none of which fit the audience. AI is excellent at producing things that look like a brand. It is bad at producing things that are a brand. Brands resolve a thousand small calls about audience, voice, history, and the calls that AI has not been told about - which is most of them - it makes by averaging across its training distribution. Result: a brand that is generic in a polished way.

The pattern is the same: AI confidently fills in the shape of work that requires judgment, and the judgment-free version looks correct enough that it gets shipped.

Where AI earns its keep.

• Generating boilerplate the operator was going to write anyway. Form scaffolds, type definitions, test fixtures, error states.
• Refactoring known code in known ways. Renames, extractions, mechanical migrations.
• Drafting copy for a human to edit. Three drafts in two minutes is a real speedup; using draft one as final is the trap.
• Accelerating research with a bounded answer. What does the Stripe webhook payload for X look like. Useful. Concrete.

Where it pays for the next rebuild.

• Picking the data model.
• Picking the user flow.
• Picking what to cut from the v1.
• Writing the brand voice.
• Deciding what done means.

Those are not productivity problems. They are taste problems. AI in 2026 does not have taste. It has an average, and the average is exactly the wrong place to land for a brand or a v1.

The line we hold at Pointline: AI removes work. It does not remove judgment. The judgment is what you are paying for. The work is what we are removing for you.

If an agency tells you they ship faster because AI does the design, ask which decisions AI is making and which ones a human checked. The honest answer narrows what the AI is actually doing. The dishonest answer ships you a brand that looks like every other brand the same model wrote that week.