Trust is the asset you are protecting
An AI feature that is confidently wrong in its first week is remembered long after it is fixed. Users who stop trusting a system route around it permanently, and adoption never recovers to where it would have been.
So we launch in stages, and each stage is gated on evidence rather than on a date.
The four stages
- Shadow. The system produces output, nobody acts on it, humans do the work as usual and the two are compared. Two to four weeks.
- Assisted. Output is presented as a suggestion; a person accepts, edits or rejects. Speed improves, quality stays owned by people.
- Autonomous on a narrow class. The boring, safe, high-confidence cases only.
- Widen by evidence, one category at a time, watching re-contact and correction rates rather than volume.
Shadow mode is where you learn the real accuracy
Every AI project we have run has found something in shadow mode that testing missed. Real traffic has a distribution that curated test cases do not, and that is precisely the point of the stage.
What to watch after each stage
- Correction rate — how often humans change the output, by category
- Re-contact or rework rate, which catches confidently wrong output that looked fine
- Time per item, which should fall or the feature is not helping
- Refusal rate — a jump usually means retrieval has broken
- Cost per completed task, before it becomes a surprise
Tell users what it is
Disclose that AI is involved, say what a human still checks, and make the route to a person obvious. Users are markedly more tolerant of a disclosed AI making an error than of discovering one afterwards.
Also give them a one-click way to report a bad answer. It is the highest-quality feedback you will get and it costs almost nothing to build.
Have a way to turn it off
A switch that disables the feature in seconds, in the hands of someone in the business rather than requiring a developer.
When something is going wrong at volume, that matters more than any capability.