01
Begin with one need
An assistant may help find a procedure, draft a reply or gather case information. Combining every objective from day one makes testing difficult and increases risk.
Choose a frequent task, define users and the required answer format, and list questions the assistant must not answer.
02
Govern knowledge and access
Documents need owners, review dates and approval status. A draft instruction must not carry the same authority as an approved procedure.
A user who cannot open a source document should not obtain it through the assistant. Enforce access at retrieval time, not only at sign-in.
03
Make answers verifiable
Show the source, passage or document version behind an answer. Missing evidence should lead to a refusal or human escalation.
- approved knowledge only;
- defence against instructions hidden in documents;
- sensitive-data filtering;
- useful logs without excessive conversation content;
- clear disclosure that the user is interacting with AI.
04
Test before release
Include valid, ambiguous, conflicting, out-of-scope and adversarial questions. Evaluate correctness, completeness, citations, refusals and latency.
Collect feedback after launch, but do not automatically train on every conversation without validating quality and lawful use.
SOURCES
Verified source material
Information was checked on the publication date. Legal obligations should always be assessed for the specific organisation and use case.