There's a version of "add AI" that's all demo and no value — a chatbot bolted to a corner, a feature that impresses in a meeting and gets ignored in production. The AI worth building does the opposite. It disappears into the workflow and quietly saves people time.
Here's how we tell the two apart.
Good AI removes a step, it doesn't add one
The best AI features feel less like a new tool and more like the product getting smarter. They take something the user already does and make it faster or easier — without asking the user to learn a new way of working.
- Drafting a first version of text the user will edit anyway.
- Summarizing a long record into the part that matters right now.
- Catching the likely mistake before it's submitted.
If a feature makes users stop and think "how do I use the AI?", it's in the way. The goal is fewer steps, not a new one.
Keep a human in the loop where it counts
For anything consequential — money, health, legal, a message to a customer — AI should propose, and a person should decide. We design these features as accelerators with an off-ramp: the output is always reviewable and editable, never silently final.
This isn't just safer. It's what makes people trust the feature enough to actually use it.
Be honest about confidence
AI is wrong sometimes, and pretending otherwise erodes trust fast. We design for the wrong answer:
- Show the source, so users can verify.
- Make corrections easy and fast.
- Fail gracefully when the model isn't sure.
A feature that admits uncertainty is one users keep relying on. A feature that's confidently wrong gets switched off.
Where we leave AI out
Not every problem wants a model. If a simple rule, a good default, or a clearer UI solves it, that's almost always the better answer — cheaper, faster, and predictable. We reach for AI when the problem is genuinely fuzzy, the input is messy, or the value is in handling endless variation no rule could cover.
The test we apply
Before we ship an AI feature, we ask one question: would the user miss it if it were gone? If the honest answer is no, it was a demo, not a feature. If the answer is yes, it has earned its place.