Pilot group. Create a group (for example AI
Pilot) and turn the AI module on only there (section 11.7.1).
Publish.
Teach. Create a few clearly distinct types. Upload
20+ varied examples per type, in every language used.
Train.
Check the figures.type right ≥ 90 %,
unrelated documents recognized ≥ 95 %, precision at 70 % high.
Fix weak types before going further.
Test documents. Run a handful of real, unseen files
of each type — and a few that look similar but are not — through
Test a document. Add misses as examples and false
matches as counter-examples; retrain.
Test mode. Create the rules from the AI page (they
start in test mode at 70 %), deliver them to the pilot group,
publish.
Observe for one to two weeks. Review AI rule in
test mode matched in the Audit Log and the AI
Classification report. Upload wrong matches as
counter-examples; retrain (expect a re-evaluation each time).
Recommend. Turn test mode off with action
Recommend. Watch kept / changed in the report
— many changed suggestions mean the type or its label is
wrong.
Widen. Turn the module on for more groups.
Warn or Enforce only for types whose precision at
the chosen threshold is ≥ 95 % in the editor
and confirmed in recommend mode. Consider a higher threshold
(80–90 %) for enforcement.
