Worked example
Atlas: a bounded service-desk assistant
The completed Canvas for the running example in the book. It shows the decision and boundary in each section, without reproducing every supporting artefact.
- User
- Service-desk colleagues
- Scope
- Two global policy domains: travel and expenses
- Ambition
- Augmentation, not automation
- Requested route
- A controlled 30-60-90 day pilot
| 1Business problem and outcome | Service-desk colleagues repeatedly locate policy evidence and draft routine answers. Improve handling time and consistency without transferring policy authority. |
|---|---|
| 2User and workflow | The primary user is a service-desk colleague. Atlas retrieves and drafts; the colleague inspects, edits, sends or escalates. |
| 3AI task, pattern and ambition | Finding, comparing and grounded drafting through a filtered RAG assistant. Augmentation, not automation. No entitlement decision or write action. |
| 4Value and success | Capacity release and service improvement measured against handling time, rework, claim support, review effort, escalation, cost and guardrails. |
| 5Knowledge, data and access | Two approved global policy collections with status, jurisdiction, replacement and permission metadata. Regional synthesis remains excluded. |
| 6Required behaviour and boundaries | Cited drafts, visible passages, clarification, conflict and no-evidence routes, deterministic prohibited topics and safe service fallback. |
| 7Human oversight and exceptions | Service desk reviews routine evidence; policy owner resolves conflict and interpretation. Review behaviour and workload are measured. |
| 8Risk, governance and ownership | Controls cover eligibility, permissions, claim support, stale sources, incidents and change. Named product, policy, engineering, security and service owners. |
| 9Evaluation and monitoring | Versioned routine, difficult, restricted, prohibited, stale-source and failure cases across component, response, task, workflow and outcome layers. |
| 10Pilot and delivery roadmap | Controlled 30-60-90 day service-desk pilot. Pre-pilot hard gates, supervised evidence, operating review and an explicit decision route. |
Illustrative teaching example from the book, not a deployed customer result.
Deliberately incomplete
The gaps stay visible.
A credible Canvas does not force every box to look green. Atlas still has open work, and it is written down because it affects the investment decision.
- Permission and source-change tests still have to pass.
- Review capacity needs supervised evidence from real service-desk work.
- Sustained value and operating cost need pilot observation.
The five decision routes
Every gate ends in one of five routes.
Atlas asks for a bounded pilot once the pre-pilot gates pass. Record the evidence, owner, unresolved risk and the condition for reconsidering.
Material gates pass and the next investment is justified.
The opportunity remains credible, but named readiness, control or evidence work must happen first.
The problem remains valuable, but the user, workflow, scope, pattern or control model must change.
A dependency blocks useful progress and a clear event could reopen the opportunity.
The value does not survive cost or consequence, or a material control cannot be met.