Abstract
A four-page worksheet for deciding what must be demonstrated before one enterprise AI workflow moves into production or expands its authority. It is written to be completed by three people in the same room: the business owner, the technical lead and the control owner.
Each of the ten questions names what a yes has to mean, the evidence an outsider could inspect, and four honest answers: yes, partial, no, unknown. The questions run from ownership and certified information through business meaning, knowledge, identity, tool authority, evaluation, reconstruction, recovery and economics.
It is a discussion aid, not a certification. The most useful outcome is usually the question where the three of you disagree, and the worksheet closes by turning the three most consequential gaps into a release decision with named owners and closure evidence.
Decision it supports
Which gaps to close before one workflow goes live.
What you will produce
A completed worksheet agreed by the business owner, technical lead and control owner.
See it worked through in a guide
Claims package 1.1.0 · 26 October 2026. Synthetic worked example: team A, group disability, region 2. Read-only preparation of the assigned case; examiners retain every case decision. Hold candidate 1.0.0 after an unrelated employee’s claim is returned. In the synthetic decision, accept 1.1.0 for a monitored team A rollout only after prior-claims history is disabled and the remaining scope passes a fresh evaluation. All figures, evidence and sign-offs in that guide are invented; the paper provides the underlying method.
In the paper
Three passages, as published
Evidence, not opinion
Every answer should point to evidence another person can inspect.
Do not average it
Don’t turn the count into a percentage readiness score. A critical control failure can outweigh nine demonstrated capabilities.
Honest scoring
Never mark Yes for an untested or excluded control.
The worksheet
The ten questions
Each answer is yes, partial, no or unknown. Answer from the implemented path, not from the roadmap. For a quicker, nine-question version, take the online diagnostic.
Q01. Is there an owner and a measurable outcome?
A named business owner can change the workflow and accept results. Measures define the baseline, target, eligible population, human review and exceptions.
Evidence to inspect. Show the owner’s decision rights and outcome measure. A sponsor without workflow authority is insufficient.
Q02. Are the information dependencies certified?
Approved data products have owners, interfaces, quality and freshness expectations, and lineage. Current evidence shows that dependencies meet this workflow’s requirements.
Evidence to inspect. Show approved contracts and current quality and freshness results. A catalog entry alone is insufficient.
Q03. Is business meaning approved and tested?
An accountable owner approves entities, metrics, relationships and exceptions. Versioned definitions specify grain and context, and known-answer tests check that the intended meaning is applied.
Evidence to inspect. Show definition approval and test results, including context selection where several definitions are valid.
Q04. Does the knowledge path preserve source controls?
Retrieved knowledge is applicable, current, authorized and traceable to sources. Controls handle permission revocation, document replacement and deletion, and stale indexes.
Evidence to inspect. Show source-control tests. If retrieval is outside scope, record the boundary and the control owner’s acceptance.
Q05. Is identity enforced through the complete path?
Caller and agent authority is bounded at every enforcement point, including tools, APIs, databases, retrieval and caches. Tenant boundaries remain intact.
Evidence to inspect. Show denied-access, revoked-user and isolation tests across the complete request path.
Q06. Is the agent limited to approved capabilities?
Every tool has an owner, a defined interface and an approved action scope. A read request cannot become an unauthorized write or bypass policy through another tool.
Evidence to inspect. Show the tool inventory, side effects, approval rules, and tests of actions the agent must refuse.
Q07. Does evaluation block an unsafe release?
Versioned tests set pass thresholds for answer quality, correct refusals and resistance to attacks such as prompt injection. A failed control blocks release, and material changes trigger re-evaluation.
Evidence to inspect. Show test results, the release gate, failure cases, and the person authorized to stop a release.
Q08. Can one request be reconstructed?
A reviewer can trace authority, sources, component versions, policy decisions, result and outcome. Evidence has defined retention and reviewer access without unnecessary sensitive logging.
Evidence to inspect. Walk through a real request trace. Show what is retained, who can inspect it, and any evidence gaps.
Q09. Can the service fail safely and recover?
Operating limits, support ownership, service and freshness objectives, rollback and recovery are defined. Source, policy or tool failure produces a safe response.
Evidence to inspect. Show recovery and escalation tests. For writes, include duplicate protection and remediation of incorrect actions.
Q10. Can the internal team operate it economically?
The team tracks cost per completed useful task, adoption and business results. Internal owners support the service, review changes and fund recurring maintenance.
Evidence to inspect. Show costs, operating limits, support responsibilities, handover evidence, and the criteria to expand, change or stop.
Next step
Discuss this paper against your own program.
A free sixty-minute working session with Richard on your top gaps. You leave with a one-page summary within three business days, and three next questions.