Adding AI to a product is now easy. Getting people to rely on it is not. The features people come back to tend to share a few design decisions, and almost none of them are about how clever the model is.
This article walks through those decisions in the order we think about them, then ends with a short checklist you can use on your own product.
Start from the task, not the chat box
The fastest way to add AI to a product is a chat box. It's also the easiest way to make people unsure what to do next. An empty text field asks the user to know what the tool can do, phrase it well, and judge the answer, all at once.
Start instead from the task someone is already doing. A person reading a contract wants the key dates pulled out. A support agent wants a draft reply to the message in front of them. A shop owner wants last month's invoices sorted by what's unpaid. Each of these can be a button, a pre-filled field or a suggestion in the place where the work already happens.
A chat box can still be useful as a second way in. It just shouldn't be the only one.
Show what the AI used
People trust an answer more when they can see where it came from. If a feature summarises a document, show which parts it drew on. If it answers a question from your own knowledge base, link to the pages it used. If it filled in a form from an email, highlight which fields came from which line.
This does two things. It lets people check the answer quickly, and it teaches them what the feature is good at, so they use it for the right jobs.
It also exposes problems early. When a source is missing or wrong, the user sees it straight away instead of finding out later from a customer.
Make results easy to check, edit and undo
An AI result should arrive as a draft, not a decision. Design for the moment after the answer appears:
- Check. Put the result next to the input, so comparing the two is one glance, not three clicks.
- Edit. Let people change the result in place. A draft they can fix is more useful than a perfect answer they can't touch.
- Undo. Make every AI action reversible, and make the undo obvious. People try new features more readily when they know they can take them back.
If a result can't be undone (an email that has been sent, a payment that has been made), it needs a confirmation step that shows exactly what will happen.
Say plainly what it can't do
Every AI feature has limits: documents it can't read, questions it can't answer, languages it handles poorly. Say so in the interface, in plain words, near the feature itself. "Works with typed PDFs, not scanned images" is more useful than a general disclaimer in the footer.
When the feature isn't confident, it should say that too, and offer the manual path instead of guessing.
Keep a person in the loop where mistakes are costly
Not every task needs the same level of checking. Sorting emails into folders can run on its own; a wrong folder is easy to fix. Sending a quote to a client, changing a customer's record or replying on the company's behalf should wait for a person to approve.
A useful way to decide: ask what happens if the feature is wrong, and who finds out first. If the answer is "a customer, after the fact", a person approves before anything goes out.
A short checklist
Before you ship an AI feature, check that:
- It starts from a task the user is already doing, in the place they already do it.
- It shows what it used to produce the result.
- The result appears as an editable draft.
- Every action can be undone, or needs a clear confirmation first.
- The interface says what the feature can't do, in plain words.
- Anything costly to get wrong waits for a person to approve.
- You can see, after launch, how often people edit or undo its results, so you know where it's falling short.
If you're thinking about adding AI to one of your own workflows, our AI and automation page describes how we start: one workflow, tested on your real examples, then expanded.
