Find the decision before choosing AI
Auteur’s five stages—Frame, Direct, Deliver, Verify and Learn—move from defining a problem to bounded work, a reversible slice, behavioral checks and lessons for the next decision. Start with Frame: what needs to change? See The Auteur framework (https://www.gabrielfenton.com/auteur-framework.html). Source: The Auteur framework.
Fictional worked example: a small software company is considering AI suggestions for support-ticket categories and priorities. One ticket reads: “I cannot sign in after changing my email. I also received a password-reset message I did not request.”
The fictional team’s policy requires designated human review of unexpected-reset reports. The message does not establish an account compromise. The first question is whether the existing workflow already sends this ticket to the right reviewer.
Before choosing a build, ask who handles triage, where exceptions go and which mistakes matter most. Treat slow or inconsistent handling as a question to investigate. Gather examples of the current process, including corrections and escalations, rather than assuming AI is needed. Source: The Auteur framework.
Keep three options open. Retain manual handling if the workload is manageable. Test a simple rule if a documented condition covers the need and its exceptions. Compare those options with AI suggestions on the same representative cases, recording handling effort and unresolved gaps. A simpler workflow is a valid outcome. Source: The Auteur framework.
Turn the boundary into one usable workflow
Auteur’s internal-tool example combines category and priority suggestions, supporting reasons and human approval. Use that starting boundary for the fictional team’s trial. The Auteur framework (https://www.gabrielfenton.com/auteur-framework.html). Source: The Auteur framework.
For Direct, adapt the Auteur direction brief (https://www.gabrielfenton.com/auteur-direction-brief.md). Define the outcome as a suggestion a support agent can inspect, correct and approve in one eligible queue. Document current routing and exceptions; leave unknowns as questions. Name the approval and escalation owners, permitted ticket data and allowed categories. Exclude automatic replies, ticket closure and autonomous high-risk routing. Specify acceptance evidence before building: the required escalation remains enforceable, corrections are preserved and manual completion works when suggestions are unavailable. Source: AI Product Direction Brief.
For Deliver, build that complete path. Place the original ticket beside the proposed category, priority and a reason tied to its text. In this fictional example, keep the unexpected password-reset detail visible. A routine suggestion must not bypass the team’s designated-review requirement; the approver needs a clear way to correct it and escalate.
Send only permitted content for processing. Preserve the original suggestion and final decision under defined access and retention rules. Include manual completion and a way to disable suggestions without stopping triage. These are proposed behaviors to implement and check; the example supplies no working interface or test results.
Use failure cases to decide whether the slice is ready
Auteur calls for checking correctness, product intent, risk, maintenance and the user journey, including failures and recovery. See The Auteur framework (https://www.gabrielfenton.com/auteur-framework.html). Source: The Auteur framework.
For this fictional triage workflow, establish expected handling for ordinary, ambiguous, escalation and unavailable-service cases before evaluation. Compare manual triage, a simple rule and AI suggestions on the same permitted cases.
Adapt the table: assign owners, replace the fictional policy and set acceptable limits for consequential errors, review effort, correction burden and operating cost beforehand. These are starting checks; broader coverage depends on your product’s risks.
| Check | Failure to look for | Correction | How to verify | Priority |
|---|---|---|---|---|
| Would simpler handling suffice? | No defined gap. | Keep manual triage or test a rule. | Compare cases, exceptions and effort. | Before build |
| Does escalation survive? | Unexpected-reset detail missed. | Prevent routine approval bypassing designated review. | Exercise the running ticket, paraphrases and mixed-intent variants. | Before trial |
| Is priority supported? | Invented outage or deadline. | Correct or handle manually. | Trace reasons to ticket text or approved policy. | Before trial |
| Can triage continue? | Timeout blocks work; edits disappear or stale output persists. | Preserve edits; allow manual completion. | Interrupt requests; try recovery and retry. | Before trial |
| Can approval be challenged? | Ticket hidden; correction impractical. | Expose original text and correction controls. | Reject a wrong suggestion; inspect saved decision. | Before trial |
| Is release justified? | Demonstration replaces evidence. | Defer or choose simpler handling. | Assess errors, effort, cost and unresolved risks against agreed limits. | Before release |
Record the decision and the next bounded step
Use Learn to leave the next task with a clear decision record. The Auteur direction brief (https://www.gabrielfenton.com/auteur-direction-brief.md) separates changes, reasons, verified evidence, uncertainty and the next action. Add your chosen scope, corrections and an accountable owner. Mark checks you have not run as planned; record what you actually observed separately. Source: AI Product Direction Brief.
If a later trial misses the unexpected-reset detail, investigate why before widening the scope. Turn the cause into a targeted check: can a routine suggestion bypass required escalation? This applies the learning principle in The Auteur framework (https://www.gabrielfenton.com/auteur-framework.html). It describes a possible future failure, not an incident that has occurred. Source: The Auteur framework.
For the fictional team, the proposed decision is to defer the AI trial. The support lead first documents the unexpected-reset escalation path and expected handling. The product lead confirms permitted data and comparison criteria. Those unanswered questions explain the deferral; no completed trial or measured benefit is assumed.
Reconsider the decision when that work is complete. Later evidence might justify keeping manual triage, choosing an explicit rule, revising the suggested workflow or approving a bounded release. Record which option the evidence supports, what remains uncertain and who owns the next step.
AI-assisted research and drafting.