1. Business problem and outcome

Assumption: service-desk colleagues spend excessive time locating policy evidence and drafting routine replies. Gap: measure current handling time and rework before promising benefit. Outcome sought: faster, accurate responses within two eligible global policy domains.

2. User and workflow

Decision: a service-desk colleague receives a routine question, confirms context, asks Atlas for a cited draft, checks the evidence, then edits, sends or escalates through the existing workflow. Employee self-service is excluded.

3. AI task, pattern and ambition

Decision: retrieval-augmented drafting; ambition is Augmentation. Compare against improved search plus response templates. Atlas cannot determine entitlement, resolve policy conflicts or send replies.

4. Value and success

Assumption: retrieval and drafting save time after review effort is included. Proposed measure: median end-to-end handling time, correction effort and supported-claim rate. Gap: establish baseline and agree thresholds before any pilot. Capacity released is not a cashable saving.

5. Knowledge, data and access

Decision: only approved, current documents in two global policy domains are eligible. Regional synthesis is excluded. Gap: knowledge owner must verify authority, dates, permissions and updates; access changes must propagate before continued use.

6. Required behaviour and boundaries

Decision: return a draft with supporting passages, source references and scope limitations. Missing context triggers a question. Missing or conflicting evidence triggers escalation. Permission enforcement is deterministic. Service failure returns the colleague to manual handling.

7. Human oversight and exceptions

Decision: the colleague reviews every draft against visible sources and can edit, reject or escalate. Policy owners retain interpretive authority. If a reviewer is unavailable, no reply is sent by Atlas. Measure corrections, escalations and review effort.

8. Risk, governance and ownership

Risks: unsupported advice, stale policy and unauthorised disclosure. Proposed controls: eligible corpus, permission checks, cited review, access revocation tests and a disable route. Gap: assign named product, knowledge, security and incident owners before approval.

9. Evaluation and monitoring

Proposed evaluation: routine, ambiguous, conflicting, restricted, stale-source and service-failure cases. Any observed unauthorised disclosure or autonomous send fails the release gate. Agree quality and workflow thresholds in advance; passing a sample does not prove zero future failures.

10. Pilot and delivery roadmap

Request: Sequence. First establish the baseline, approve eligible sources, assign owners and pass permission and behaviour tests. Then consider a supervised service-desk pilot. Stop use after material control failure; return to manual handling. Self-service and write actions remain deferred.