Lately the question has shifted from "should we use AI?" to "where do we put the chat?". And that's where a lot of products go wrong: they treat AI as a feature to show off, when it should be a means for the user to get something done that used to be hard.
Start with the job, not the model
The useful question is the same as for any feature: what job is the user trying to get done, and where do they get stuck? AI comes in when it removes a blocker that other solutions don't handle well.
The patterns where I see AI creating real value:
| Pattern | The user gets stuck because... | AI helps to... |
|---|---|---|
| Blank page | They don't know where to start | Generate a first draft to edit |
| Specialist knowledge | They don't master the method | Guide step by step, explaining along the way |
| Volume | There's too much information to read | Summarize and highlight what matters |
| Translating intent | They know what they want, not how to ask the system | Turn natural language into action |
A concrete example: a store owner who "never managed to put together a cash flow" doesn't need a generic chat. They need an assistant that talks about the context of their business and, at the end, leaves the categories and projections created in the system, not just explained in the conversation. The difference between "the AI explained it to me" and "the AI helped me do it" is the difference between curiosity and value.
Designing the experience
A few experience decisions that make a difference:
Draft, not decision
The AI proposes, the person decides. Everything the AI generates shows up as an editable draft, with the final action always in the user's hands. That lowers the cost of mistakes (which will happen) and keeps responsibility where it belongs.
Context is everything
A generic model with the right context beats a better model without it. The value of an in-product assistant is precisely knowing who the user is, what the business is, what has already been done. But context is also a responsibility: only what's needed goes in.
Your own method
What sets your product's assistant apart from any chat is the method it follows: the questions it asks, the order, the criteria. That method is product knowledge and should be written, tested and versioned like any other part of the system.
Fail gracefully
The model will get things wrong, be slow or go down. The experience has to plan for that: honest loading states, clear messages and a manual path that's always available.
Privacy by design
AI in the product means sending user data to a model. That calls for explicit decisions, ideally public ones:
- What the assistant receives and what it never receives. Sensitive data that isn't needed for the task stays out;
- Vendors that don't train on the data and don't keep it longer than necessary;
- Defined retention: how long the history and logs are kept;
- Transparency in the privacy policy, in plain language;
- Tests with fictional data during development, never with real conversations without authorization.
Engineering: keep it simple
On the technical side, the temptation is to build a complex architecture right away. What's usually enough at first:
- A service layer that isolates the model vendor (switching vendors can't require touching the whole product);
- Prompts and methods versioned in the repository, with regression tests using fictional conversations;
- Tools (tool use) so the model acts on the system in a controlled way: create the draft, fetch the context, never free access;
- Usage and cost logging per account, to understand the value and the margin.
How to tell it worked
The metric isn't "messages sent to the assistant". It's the same as always: did the user get the job done? A few questions for the dashboard:
- How many interactions end in a completed action in the product?
- Do people who use the assistant activate faster or retain better?
- How much of what the AI generates is accepted, edited or discarded?
Summary
AI in the product is a product decision before it's a technical one. Find the user's real blocker, design the AI as the one who proposes (not the one who decides), turn the result into something concrete in the system, treat privacy as part of the feature and measure by the job done. Without the hype, AI becomes what it should be: a way for your product to be more useful.
Murilo Anzini
Hands-on writing on engineering, product, data and marketing for people who want to understand, and build, the whole journey of a product.
Culture isn't a wall of values: it's ritual
The values on the wall say what a company would like to be. The rituals (the conversations, the reviews, what gets celebrated) show what it is. How to turn culture into a recurring practice.
A/B testing with little traffic: what to do when statistics won't help you
With a few visits a week, a classic A/B test takes months to tell you anything. How to experiment honestly in small products, without false winners.