When we started building AI release notes, the first prototype was impressive in demos and nerve-wracking in practice. It was confidently wrong about one change in ten. For a feature that speaks on our customers’ behalf, that was nowhere near good enough.
Keep a human in the loop
Every AI draft in Liftoff is a suggestion. Nothing is published until a person reviews it, and the editor highlights exactly which sentences came from the model.
Make the source visible
Each generated line links back to the pull request or task it was written from. If something looks off, you can check the source in one click.
Clear labels on anything AI-generated
Citations back to the original work
An easy way to report a bad suggestion
No training on customer data without opt-in
Measure trust, not just usage
We track how often drafts are accepted unchanged, edited or discarded. Usage tells you people tried the feature; acceptance tells you they trust it.
