Skip to content
NEWY
Book a Call

AI features that survive contact with users

The demo is the easy part. What matters is what happens when the model is wrong, and whether the user can tell.

Building an impressive AI demo has become straightforward. Building one that people keep using has not, and the difference is almost entirely in how the feature behaves when it is wrong.

Assume it will be wrong in public

Design the failure case first. Can the user see what the system did? Can they correct it? Does the correction stick, or will it make the same mistake tomorrow?

Put it where the work already happens

A separate chat panel asks people to change their habits. The features that stick tend to sit inside an existing workflow — drafting the reply they were about to write, filling the form they were about to complete.

Show your working

  • Cite the source when the answer came from a document

  • Say when confidence is low rather than guessing fluently

  • Make the edit path as fast as accepting the suggestion

  • Never present a generated figure as though it were retrieved

Decide what it must never do

Write down the actions the system may not take without a human — sending, paying, deleting, promising. That list is a product decision, not a technical one, and it is much cheaper to make before launch than after.