How to read an AI assessment like a reviewer
Trust comes from knowing exactly what was tested, what failed — and when the result stops applying.
Start with identity and scope
Match agent name, version, environment, use, tools, data and supported languages against the deployment you are considering. A result for a support assistant does not cover a later version that books appointments or updates accounts.
Follow each claim to its evidence
A strong dossier links every control to sourced records and a repeatable test. Check sample size, selection, observed failures and stated limits. The four pillar scores help navigation — they never replace the control-level record.
Check status and triggers
A valid status always has an expiry date. Suspended, revoked and expired mean very different things. The draft proposes review every three months, while serious incidents, new tools, changed permissions or changed high-impact use can demand earlier reassessment.
Keep the limits in sight
No certification can promise safety, compliance, commercial fitness or insurance. Critical blockers can never be averaged away, and an implementer can never certify their own work. Read the sample dossier next to see these rules in action — then choose “Discuss your agent” when you want your own system examined.