Write a short opportunity decision record that selects one bounded experiment and explains what would change the choice.
Define the decision you are actually making
An opportunity list contains possible improvements. A portfolio decision chooses which of them receives scarce attention now. Your goal is to justify the next commitment, not to predict every future return.
A useful decision boundary names the team, time available, problem to improve, and actions outside scope. “Select one two-week discovery experiment for the support team” is easier to assess than “choose our AI strategy.”
Start one step earlier if you cannot describe a single task and its current process. Read Choose an AI use case, then return with comparable task descriptions.
Make the alternatives comparable
Write each candidate as a problem, current baseline, proposed change, affected people, and observable outcome. Include a simpler non-AI alternative such as a template, routing rule, or clearer source document.
Separate purpose from technique by asking what improves for the person doing or receiving the work. Faster drafting matters only if the completed task still meets its acceptance conditions.
Record evidence maturity as observed, estimated, or unknown. A measured baseline and an enthusiastic prediction are different kinds of evidence, even when both appear in the same spreadsheet.
Use gates before rankings
A decision gate is a necessary condition for the specific next step. In this fictional team, a discovery experiment needs approved test material, a responsible owner, reviewer capacity, and no automatic external action.
An unmet gate does not become acceptable because the idea scores well elsewhere. Mark it blocked, or change the scope and assess the new boundary explicitly. An unknown gate needs evidence before approval.
Keep the distinction practical because an interview or a test with fictional inputs may require a different boundary from production use. Passing the discovery gates does not authorize live customer processing.
Compare three fictional opportunities
Fictional teaching scenario. A support team has six hours of reviewer time available each week. The figures below are estimates for discovery planning, not observed savings. All three ideas still need evaluation.
| Candidate | Proposed work | Reviewer demand / week | Evidence and current boundary |
|---|---|---|---|
| A · Policy reply drafts | Draft replies from approved standard policies | 4 hours | Source owner identified; fictional test set ready; sending excluded |
| B · Meeting action summaries | Extract proposed actions from meeting notes | 2 hours | Approved synthetic notes ready; compare with a structured template |
| C · Tailored account replies | Draft replies using account-specific terms | 5 hours | Source permissions unresolved; source owner not assigned |
C remains blocked for this experiment. Its potential value cannot replace the missing permission and ownership evidence. A and B fit separately within the six-hour constraint; running both would use all available review time before unexpected rework.
Estimate value as a range, with a boundary
For candidate A assume 200 eligible tasks per week. The existing process takes 8 minutes per accepted task. The AI-assisted process is estimated at 5 to 7 minutes, including task-level review and corrections.
Shared upkeep is a separate estimated 120 minutes per week, covering source maintenance and workflow checks. Do not add task-level review again: it is already inside the 5-to-7-minute estimate.
The calculation is 200 × (8 − 7) − 120 = 80 minutes at the lower saving assumption, and 200 × (8 − 5) − 120 = 480 minutes at the higher assumption. This is a scenario range of 80–480 minutes per week, not a statistical confidence interval or an achieved result.
The boundary matters because setup, fees, and changes in task volume are excluded. Time potentially freed does not automatically become a cash saving. Equivalent accepted quality is an assumption to test, not a fact established by this calculation.
Test whether the choice is fragile
If task-level handling instead takes 7.5 minutes, the same model gives 200 × (8 − 7.5) − 120 = −20 minutes: a net increase in effort. The break-even time is 7.4 minutes because 200 × (8 − 7.4) equals 120.
This sensitivity check identifies what to measure next. A pilot that records only draft-generation time cannot resolve the uncertainty about accepted-task handling time.
| Uncertainty | Low-cost evidence to collect | Decision consequence |
|---|---|---|
| Review effort | Time a small set through acceptance, including rejected drafts | Reconsider A if realistic handling stays above break-even |
| Need for AI in B | Compare action extraction with a structured meeting template | Prefer the template if it solves the task adequately |
| Source readiness in C | Ask the designated owner to verify scope and access | Keep C blocked while permission or ownership remains unresolved |
Break-even applies only to the stated time model. A failed quality gate can block the work even when the estimated time saving is positive.
Check shared capacity and dependencies
Capacity is not interchangeable when two ideas need the same reviewer or source specialist. An available engineering hour cannot automatically replace the person authorized to review policy exceptions.
A shared dependency can change the sequence. Improving the policy register might support several future candidates, but record its cost once and name which assumptions depend on it. Do not count the same benefit separately in every business case.
Leave recovery capacity for unexpected cases. The fictional team chooses A alone, reserving two reviewer hours. That is a local planning choice, not a universal utilization rule. If B is more urgent or A lacks a suitable owner, a different decision can be defensible.
Use scores to expose preferences, not manufacture certainty
A weighted score can make priorities explicit, but it also hides judgments behind a precise-looking number. Write what each criterion means, where its rating came from, and why its weight fits this decision.
Test the ordering by changing a plausible rating or weight. If the preferred candidate changes, report that the ranking is fragile and identify the evidence that would resolve it. Avoid multiplying an unsupported confidence percentage into a speculative value estimate.
Record exclusions separately so a high weighted total never overrides a failed gate. Keep uncertain impact, implementation effort, and reviewer capacity visible rather than compressing everything into one score.
Fund a question with a stopping point
The selected experiment for A checks whether approved policy drafts preserve exceptions and reduce accepted-task effort. Use fictional material and keep sending disabled. Record ordinary, exception, and missing-source cases separately.
A bounded commitment names an owner, available review hours, a deadline, and the question the experiment must answer. A small discovery set can expose failures; it cannot establish broad production reliability.
A review record explains why A was selected, why B was deferred, why C was blocked, and what evidence could reverse those decisions. After discovery, use the pilot decision guide to evaluate the next boundary.
Practice: defend a changed priority
The reviewer assigned to A is now available for only two hours per week. A still needs four hours at the planned scope. B needs two hours and has approved synthetic inputs. C remains blocked. Write a decision memo of no more than one page.
- Choose a next experiment or defer all candidates. State the scope and owner.
- Explain which gate or capacity constraint changed.
- Reproduce A’s 80–480-minute scenario and its 7.4-minute break-even assumption.
- Describe a non-AI comparison and one observation that would reverse your choice.
- Give the next review a date, an evidence requirement, and a stopping rule.
Success criteria: another reader can trace your choice to the capacity constraint, distinguish forecasts from observations, and see why unresolved gates remain unresolved.
Compare with the worked answer
Worked answer: Do not proceed with A at its original scope: four required reviewer hours do not fit two available hours. You could narrow A and reassess its task volume, review plan, and value model, or choose a bounded B experiment comparing AI extraction with a structured template. B consumes all stated review capacity, so a smaller experiment or deferral may be preferable if rework cannot be covered. C remains blocked. Keep A’s 80–480-minute estimate labeled as a forecast for 200 tasks; it does not carry unchanged to a narrower scope. Stop the chosen experiment if its required review cannot be completed or its agreed evidence boundary fails.
Use the method on your own task
Editable worksheet: Download the Markdown template and open it in a text editor. Use fictional or approved information; no account or paid service is required.
Download the worksheet (.md) ↓What to carry forward
Write a short opportunity decision record that selects one bounded experiment and explains what would change the choice.
Show your reasoning with the exercise above. Keep the evidence, the decision, and the remaining uncertainty together.