Where AI belongs in your product (and where it doesn't)

Start with the job, not the model

The AI features that survive are the ones nobody calls AI features after six months. They are just the part of the product where a tedious job became a short one. The ones that get removed are the ones that existed to be announced.

So the first question we ask a client is never "where can we add AI?" It is "where do your users do slow, repetitive work with information?" That is the territory where the current generation of models is genuinely strong, and it is bigger than most teams think.

Good homes for AI

Patterns we have shipped or would happily ship:

• Summarising what the user does not have time to read. Long threads, reports, feedback, tickets. Reading a hundred things and producing an honest gist is the technology at its best.
• Extracting structure from mess. Turning an email, an invoice, or a pasted document into filled-in fields. Users hate transcription; machines are now good at it.
• First drafts inside a workflow. A reply, a description, an outline that the user edits and owns. The keyword is inside: the draft appears where the work happens, not in a separate chat window.
• Search that understands intent. Letting people find things by describing them, not by remembering the exact title.

The common shape: the model does volume, the person keeps judgement, and the feature is invisible in the marketing and indispensable in the workflow.

Bad homes for AI

• The chatbot as navigation. If a user has to interrogate your product to find things, the interface has failed and the bot is an apology for it. Fix the interface.
• Generated content presented as fact. Anywhere a wrong answer costs your user money or trust, a fluent guess is a liability, not a feature.
• AI as the reason for the product. "It's like X but with AI" is a sentence about you, not about a user's problem. Features built on that sentence do not get used; they get demoed.

The design questions AI actually raises

This is the part that gets skipped. A probabilistic feature needs design that deterministic software never did: What does the user see while the model thinks? What happens when the answer is wrong, and how does the user say so? How does the feature earn trust in the first week instead of demanding it? Where does the user's data go, and can you answer that question in one sentence on the screen where it matters?

Teams that answer these before building ship features people rely on. Teams that do not ship a demo.

Our position

We treat AI as a material, not a message. It is very good at the parts of the work that are already understood, and no help with the part that decides whether the product is any good. If you are weighing an AI feature and want an honest read on whether it is a workflow or a press release, that is a conversation we enjoy.