Canvas and completion guide
Use the blank sheet with its full completion guide and identified supporting schedules.
Print the blank Canvas Customer-service case Investment-committee case
Appendix A. Blank Agent Delegation Canvas
A Mandate Who authorizes what, and for how long?
B Actions and human support What is permitted and feasible?
Proposed permissions: authorized / prohibited / pilot only. Omission grants none.
| Action / object / destination | Permission / limits | Required human support |
|---|---|---|
Support: prerequisites, affected parties, consequences, recourse access / owner / response time.
| Human duty / dependency | Qualified person / effort | Available / window / deadline |
|---|---|---|
Show demand/effort bounds, adequate performance, competing work, shared reserves and continuing duties. Name pause thresholds/delays. A cap alone does not prove feasibility.
C Capabilities and controls
Inspect independently of Canvas Section B. Include unmatched capabilities and alternate/unknown routes.
Disposition: authorized / blocked / contained / unresolved
D Decision and reconsideration Assess the request first; record alternatives and justify reused findings.
Outcome (select one): authorize the stated scope / pilot only / remediate / do not delegate
Permission needs all four checks met, pilot requirements if applicable, and signed, effective scope with prerequisites satisfied. Missing support is not met. Unsigned, remediate and do not delegate grant no authority. Expiry requires fresh checks and ratification. Include necessary attachments in the signed scope.
Appendix B. Completion guide
Put the decision on the sheet
Complete sections A–D in order, revisiting earlier entries when inspection reveals another capability or human duty. Use descriptive names. The sheet should explain the scope, required support, findings and decision without a codebook. Add an identified, dated attachment only when the necessary detail will not fit; include it in the signed scope. Keep inspection evidence available. Missing material support leaves the affected check not met. The Canvas records a decision; people and systems must implement it.
A. Mandate and resources
State the purpose, agent configuration, permitted scope, exclusions and affected people. Name the principal delegating work, the ratifier entitled to approve it and the operating owner. Record consequence tolerances and reasons, qualified resources, protected time and other commitments. Obtain confirmation from whoever owns shared resources. Give permission a start, expiry and timezone; justify its duration by consequences, evidence freshness and review capacity. Fictional relative periods need real dates before use.
B. Actions and human support
List consequential actions with their objects, destinations, limits and proposed permission: authorized, prohibited or pilot only. Related actions may share an entry when their boundaries remain explicit. Approval of a workflow and omission of an action grant no permission.
Name the human work needed for service, controls and correction: who performs it, with what expertise, how much effort, in which window and before which deadline. Include preparation, handoffs, rework, monitoring and duties continuing after stopping. Explain unnecessary duties or supported substitutes. For analytical work, assess evidence selection and omissions as well as statements in the draft.
Show that the duties fit together, including competing work, dependencies and exclusive attention. Use coincident upper bounds unless a supported joint restriction rules them out. Count each reserve once and specify its amount, expertise, contingencies and window. Aggregate hours alone cannot establish timely support. Name who owns the demand and effort estimates, their basis and uncertainty. A known shortage cannot become an exception by renaming it. Check accepted responsibilities, competence, actual availability and arrangements for adequate performance; an allocation or signature cannot prove engagement.
C. Capabilities and controls
Inspect what the system can actually do, including direct calls, alternate routes and unknowns, independently of the permission list. Record what was checked, when, by whom and within which boundary. Describe controls, their human dependencies and remaining exposure. Use these dispositions:
- Authorized: capability and limits match explicit permission.
- Blocked: checked prevention stops the prohibited effect, including known bypasses.
- Contained: the effect remains possible, but checked measures keep consequences within explicitly accepted bounds. Its prohibition remains.
- Unresolved: the necessary support is missing; identify the check and owner.
Prompts and absent observations do not prove prevention. A release gate supplies no assessment time, and isolation alone supplies no bound on cost or lost work.
D. Findings, decision and reconsideration
For each candidate, record met or not met, with a reason, for legitimacy and specification; consequence and recourse; capability-control; and operational responsibility and reconsideration. Capacity can affect several checks; findings are not an additive score. All four must be met for permission. Compare value only among eligible arrangements. Choose authorize the stated scope, pilot only, remediate or do not delegate. A pilot also needs a decision-relevant learning question, observations, exposure limits, supported learning duties, stopping criteria, expiry and fresh continuation criteria.
Record the selected scope, reasons, accepted exposure, implementation owner and ratifier's dated signature. An unsigned assessment or signed negative decision grants no permission. Name how affected people seek correction, who responds and when. Withhold unsupported admission; suspend affected authority when staffing, demand, evidence, capabilities or controls invalidate its basis. State detection and response delays, pause owner, notice, handoff and continuing obligations, including losses that stopping cannot repair. Check restoration before resumption; changed or expired scope needs fresh ratification.
The cases stipulate quantities, inspections, qualifications, accepted duties and adequate performance. They provide no empirical validation. AI assisted discovery, drafting and internal review; some separate agent contexts were used, but checks could share model errors. These were neither human expert review nor operational tests.