← Journal· AI · Jun 28, 2026 · 7 min read

How to add AI features to your product without it feeling bolted on

Most AI features fail the same way: they are added because AI is expected, not because they solve a real job. Here is how we decide what to build.

The pressure to ship an AI feature is real, and it is usually the wrong pressure. Teams add a chat box or a summarize button because the board asked about AI, then wonder why nobody uses it. The feature that lands is the one that removes a specific piece of work a user already hates doing. Everything else is decoration.

Start from the job, not the model

Before you pick a model, write down the exact task you want to shorten. Not the capability, the task. Draft a first reply to a support ticket. Turn a call transcript into a structured follow up. Pull the three risky clauses out of a contract. When the task is concrete, the design gets easy, because you can measure whether the feature actually saved time.

A good filter: if a competent new hire could do the task in a few minutes with clear instructions, a modern model can usually do a solid first pass of it. If the task needs judgment your business is legally on the hook for, keep a human in the loop and design for review, not autopilot.

Design for the wrong answer

Language models are confident when they are wrong, so the interface has to make the wrong answer cheap. Show the source. Make edits one click away. Never auto send. The products that feel trustworthy are not the ones with the best model, they are the ones where a bad output costs the user two seconds instead of an apology to a customer.

Keep the human in control of anything that leaves your walls: emails, payments, public posts.
Attach the evidence. If the feature makes a claim, show what it was based on.
Log inputs and outputs from day one so you can measure quality and catch regressions.
Set a fallback. When the model is unsure, say so and hand it back to the person.

Scope the first version to two weeks

The first AI feature should be small enough to ship in a fortnight and honest enough to measure. Pick one task, one screen, one clear success metric. If it helps, expand it. If it does not, you learned that cheaply. We would rather ship one feature people use every day than five that demo well and gather dust.

The best AI feature is usually the least ambitious one that removes a task somebody was dreading.

On models, the practical advice in 2026 is to build against the current frontier and keep your prompt and evaluation layer separate from the model itself, so you can move as capabilities change. Anthropic, OpenAI and Google all publish model and pricing updates worth tracking if you build on top of them.