Confirm systems and users in scope
List AI tools, Microsoft 365 workloads, connected agents, business systems, user groups, and departments included in the review.
Pre-engagement
Scope the evidence before the review begins.
The first step is to define what will be inspected, who can speak for it, and what evidence is needed. This is a facilitation resource, not a commercial assessment offer.
Identify accountable roles
Name the business owner, technical owner, data owner, and person responsible for public disclosure or procurement answers.
Request read-only evidence access
Ask for exports, reports, policies, logs, screenshots, or reader access needed to support the review without changing production systems.
Questionnaire
Ask short questions that reveal evidence gaps.
Where is AI already used?
Ask which products, assistants, agents, automations, or external tools are active, and who depends on them.
What data can those systems touch?
Ask for data classes, sensitivity labels, customer data exposure, retention rules, and vendor processing records.
What stops a bad output or action?
Ask who reviews outputs, what actions require approval, what logs exist, and how incidents are escalated.
Interviews
Interview the people closest to the facts.
Business owner
Clarify business purpose, risk appetite, public claims, customer impact, and decisions that need evidence.
IT or platform owner
Validate deployment architecture, access, logging, data governance, security controls, and known gaps.
Workflow owner
Map day-to-day process reality, exceptions, handoffs, manual reviews, and where AI could change accountability.
Compliance or legal reviewer
Check policy obligations, disclosure expectations, vendor risk, privacy constraints, and audit requirements.
Representative end users
Find shadow AI use, practical failure modes, confusing outputs, and training or escalation gaps.
Public-answer owner
Confirm how the organisation would answer a buyer, regulator, journalist, or customer asking how AI is used.
PSF scoring
Grade on evidence, not intent.
Use the Production Safety Framework as a scoring lens. A policy in draft, a planned control, or an informal habit should not be treated as deployed evidence.
- D1 Input Governance: what data can enter the AI system and how that is controlled.
- D2 Output Validation: how humans check, challenge, or approve AI outputs.
- D3 Data Protection: access, retention, sharing, vendor processing, and sensitivity boundaries.
- D4 Observability: logs, monitoring, review cadence, and evidence retention.
- D5 Deployment Safety: testing, rollout, rollback, and change control.
- D6 Human Oversight: accountable owners, approvals, escalation, and incident handling.
- D7 Security: identity, privileged access, app governance, and connector risk.
- D8 Governance Structure: decision rights, policy ownership, risk register, and leadership review.
Report
Write the report as a public-record input.
Lead with what is supported
Separate verified facts, partial evidence, open questions, and unsupported claims in the executive summary.
Link each finding to source evidence
Every finding should identify the source, the PSF domain, the owner, the risk, and the evidence needed to close the gap.
Prioritise next evidence steps
Recommend the next documents, controls, reviews, disclosures, or deployment records needed before claims expand.
Readout
Deliver the readout without turning it into a pitch.
- Open with scope: what was inspected, when, and what was out of scope.
- Show verified facts first, then gaps and uncertainty.
- Name control changes before recommending expanded AI use.
- Ask owners to confirm or correct the record before anything becomes public.
- Close with evidence next steps, not a commercial conversion step.