The AI Playbook / 4 Frontier Operator / 4.1
STAGE 4 · FRONTIER OPERATOR · SECTION 4.1

4.1 Automate the proven work.

Builds: Automation decision map, including documented remain-assisted decisions; Agent registry.

Source: Interview 9. The interview numbers are the playbook's reference labels; this page is the section.

Personal

A confirmed appointment change produces a draft follow-up and waits for your approval.

Professional

A confirmed incoming request produces a draft follow-up, waits for approval, and never sends twice.

Artifact · Automation decision map, or a documented decision to remain human-assisted. Both are pass states.

Why it exists

Automation should reduce total effort without hiding failures, bypassing approval, or turning a stable process into a faster source of mistakes. This interview exists as much to say no as to say yes: a written, evidence-backed decision to stay human-assisted is a successful outcome, not a failed one, the honest-broker artifact.

PLATE 12 · MUSTARDAutomation decision flowFour questions in order: proven, measured, justified, enforceable. Any no keeps the workflow assisted; four yeses earn a bounded operator.A CANDIDATE WORKFLOWProven by hand,at least three times?1YESMeasured: baseline and currentin the evidence log?2YESDo the numbers justifythe build and the upkeep?3YESEvery boundary enforceable by atool or a human gate?4NOStays assisteddocumented, and revisited at thenext operating reviewYESAutomateas a bounded operator, with a run logand a drilled kill switchINTERVIEW 9 · FRONTIER OPERATORWHAT THIS FLOW REFUSESAutomating the unprovena workflow you have not done by hand cannot be specifiedAutomating the unmeasuredno baseline means no claim of improvement, everAutomating the unenforceablea boundary the tool cannot hold is a promise, not a control
Fig. 12Automation decision flow, proven, measured, justified, enforceable

Questions

  1. Which proven recurring workflow is a candidate for automation?
  2. What real evidence shows the workflow works? (Point to the run log and the observed trial, not impressions.)
  3. From the intake baseline and the run log: minutes per run before assistance, minutes per run now, review minutes, rework or error rate before and now.
  4. How much review and correction does the assisted workflow require?
  5. What maintenance burden would automation create?
  6. What benefit justifies automation, if any? (Total operating benefit must exceed review, correction, and maintenance cost, in minutes, not adjectives.)
  7. What trigger should start the workflow?
  8. What context should be gathered?
  9. What reasoning or classification should occur?
  10. What action should be proposed or executed?
  11. Which actions require approval?
  12. How should approval expiry work?
  13. How will we verify the result in the destination system?
  14. How should duplicates be prevented?
  15. What should happen on timeout?
  16. What should happen when context is missing?
  17. What retry rule should apply?
  18. How should rejected approval be handled?
  19. How should the workflow be stopped?
  20. What breaks if we change the tool or model? (Name the portability cost per automation.)
  21. Who owns maintenance?
  22. When will the automation decision be reviewed?
  23. What evidence would justify keeping the workflow human-assisted instead?

Build prompt

Using the interview answers, create an automation decision map for the proven
workflow.

Include: evidence the workflow works (run-log references, trial dates);
baseline vs. current (minutes per run, review minutes, error/rework rate, before assistance and now); expected benefit in minutes per month; trigger;
context; reasoning; proposed action; approval requirement and expiry; execution
boundary; verification method in the destination system; duplicate, timeout,
missing-context, retry, and rejection handling; stop control; portability note
(what breaks if the tool or model changes); owner and review date; assisted or
human-handoff alternative.

Conclude with one of three verdicts: automate, stay assisted, or stop.
Recommend automation only if total operating benefit exceeds review,
correction, and maintenance cost. A documented decision to stay assisted is a
valid pass state, write it with the same rigor, including the evidence.
Register any live automation in the agent registry (one line: name, owner,
permissions, last review).

Separate approved facts / proposed rules / assumptions / unanswered questions /
evidence needed; mark assumptions unconfirmed. Never invent baseline numbers;
a missing number is marked "unconfirmed, measure first."

Test

For a workflow being automated, run the failure-mode suite:

  1. Normal event, executes and verifies in the destination system.
  2. Duplicate event, does not repeat the action.
  3. Timeout, fails visibly to the failure queue, not silently.
  4. Missing context, stops and escalates.
  5. Rejected or expired approval, blocks the protected action.
  6. Retry, follows the retry rule without duplicating side effects.
  7. Stop, the stop control halts the workflow; verified at the destination.
  8. Destination verification, the result matches intent, checked by the reviewer.

For a workflow staying human-assisted: the documented decision passes if it cites run-log evidence, states the benefit-vs-cost comparison in minutes, names the evidence that would change the decision, and carries a review date.

Done

Definition of done

Either the failure-mode suite passes, an authorized real run is verified in the destination system, the benefit justifies its costs in minutes, and the automation is registered, or the documented decision to remain assisted is written, evidence-backed, dated for review, and logged. Both count. A decision to automate against the numbers fails.

Minimum viable artifact

Trigger, action, approval, verification, failure queue, stop rule, owner, review date, or the one-page documented decision to stay assisted.

Implementation intention"When the automation's review date arrives, I will re-measure minutes per run and error rate against this map before renewing it."

Depth by stage

Frontier Operator only; Expert participants may read it as the standard their operator must meet before automation is discussed. Every live automation enters the agent registry, reviewed at each operating review.