No accountable owner
No named person can review the result, explain the decision, or intervene when the workflow fails.
Free resource / 12-point decision tool
Should AI carry
this task?
Score the work before you automate it. Six questions separate a useful delegation candidate from a task that still needs human context, judgment, or responsibility.
Print-friendly: use your browser’s Print command, circle one score per row, then total the final column.
Step 1 / Hard stops
Some tasks should not be scored yet.
If any statement below is true, stop. A high numerical score does not override missing authority or a consequence the team cannot responsibly absorb.
No permission
The workflow would use data, systems, identities, or communication rights the operator is not authorized to delegate.
Unbounded high stakes
An error could create serious health, legal, financial, security, employment, or safety harm without qualified review.
No evidence trail
The team cannot inspect the inputs, verify the important claims, or reconstruct what the system did.
Step 2 / Score the task
Choose 0, 1, or 2 for each row.
Score the actual workflow you can run today—not the ideal version you hope to build later.
| Criterion | Question | 0 points | 1 point | 2 points | Your score |
|---|---|---|---|---|---|
| Repetition | How often does substantially similar work recur? | One-off or unpredictable | Occasional pattern | Frequent, stable pattern | ___ |
| Outcome clarity | Can a reviewer tell whether the result is acceptable? | Success is subjective or disputed | Some standards exist | Clear result and quality bar | ___ |
| Input readiness | Are the permitted inputs available and understandable? | Missing, scattered, or unauthorized | Available with cleanup | Reliable, named, and permitted | ___ |
| Verifiability | Can claims and actions be checked against evidence? | Hard to inspect | Partly traceable | Evidence is visible and testable | ___ |
| Reversibility | Can a bad output or action be corrected without lasting harm? | Irreversible or costly | Recoverable with effort | Easy to review and reverse | ___ |
| Consequence | What happens if the system is confidently wrong? | High-stakes harm or obligation | Contained operational cost | Low-stakes and bounded | ___ |
| Total | ___ / 12 | ||||
Step 3 / Choose the operating boundary
The score recommends a starting level—not permanent permission.
Keep it human or redesign the task
The workflow is too unclear, consequential, or difficult to verify. Simplify it before introducing AI.
Use AI to draft
Let the system prepare work, but require a person to inspect every output and take every action.
Pilot with an approval gate
Test real cases, log corrections, and let AI prepare the next action without completing it.
Consider bounded action
After proven runs, allow narrow actions inside explicit limits with logging, monitoring, and escalation.
Hard stops always win.
If the task failed Step 1, resolve that issue before using the numerical band. When uncertainty sits between two scores, choose the lower one.
Step 4 / Brief the pilot
Turn the score into a testable job.
Recurring result: ______________________________________________
Permitted inputs: ______________________________________________
Quality standard: ______________________________________________
Must never: ___________________________________________________
Evidence required: _____________________________________________
Human approval before: ________________________________________
Find the hidden assumptions.
Mark missing context, unsupported claims, unnecessary corrections, and every place the system should have stopped.
Compare with the baseline.
Measure attention returned, correction time, exception rate, and whether the reviewer can explain the result.
Change one boundary at a time.
Broaden access or action only when the prior level works across ordinary cases and known edge cases.
Final approval check
Before AI takes the next step…
- □A named person owns the outcome.
- □The task passed every hard stop.
- □The score reflects today’s workflow.
- □The permitted data and tools are explicit.
- □Evidence and activity will be logged.
- □Exceptions have a clear escalation path.
- □The team can correct or reverse the action.
- □More autonomy requires another review.
Keep the useful signal
Score one task today.
Join Issue 001 next.
Join Issue 001 now. We’ll email you when the first five-minute Sunday brief goes live.