1. The situation and the decision
Imagine a professional-services team preparing 80 recurring internal reports each month. Each combines records from several systems, a short narrative, and checks by a domain reviewer. Leadership has seen an AI draft appear in seconds and is considering a broader implementation.
The proposed business outcome is less human effort per accepted report, with factual accuracy and accountable approval preserved. The decision concerns the next development commitment. It does not yet concern autonomous publication or replacing the reviewer.
The available capacity is one bounded investigation before the next planning decision. A useful output must distinguish a promising intervention from an attractive demonstration and leave the team with an actionable next step.
2. What the available observations establish
For this synthetic scenario, assume the team has mapped ten recent routine reports and tried a general drafting assistant on three of them. The table uses rounded illustrative minutes for one comparable accepted report. The assisted column is a working estimate drawn from that tiny exploratory set, not a reliable estimate of the production average.
| Work | Current workflow | Drafting assistant |
|---|---|---|
| Prepare and reconcile source records | 12 min | 14 min |
| Write the initial narrative | 18 min | 5 min |
| Verify facts, correct, and resolve inconsistencies | 20 min | 30 min |
| Final review and approval | 10 min | 10 min |
| Total active human effort | 60 min | 59 min |
The draft appears sooner, but the reviewer spends longer tracing unsupported statements and reconciling sources. In these illustrative cases, 13 minutes less writing is accompanied by 12 minutes more preparation and checking. Elapsed turnaround and waiting for approval have not been measured.
These observations challenge the assumption that writing is the main constraint. They do not establish that AI cannot help, that the one-minute difference is repeatable, or that unusual reports behave like the routine examples.
3. Compare the actual alternatives
A useful decision includes the current process and simpler interventions. The options below address different possible constraints. Their relative value remains a hypothesis.
| Option | Potential benefit | Main uncertainty |
|---|---|---|
| Keep the workflow; clarify templates and source responsibilities | Reduce ambiguity and repeated reconciliation with little new software | How much checking is caused by inconsistent inputs? |
| Build the general drafting assistant | Reduce initial narrative effort | Will verification and maintenance consume the gain? |
| Try source-linked extraction and a reviewer checklist | Make important facts easier to inspect before drafting | Can evidence links reduce total checking effort without increasing omissions? |
The recommendation favors testing verification because the illustrative observations locate substantial effort there. If source inconsistency is the dominant issue, fixing ownership and input structure may be more useful than adding AI. The experiment should keep that alternative visible.
4. The smallest discriminating experiment
Use twelve historical reports with accessible source records: eight routine cases and four selected for known exceptions. Keep those groups separate in the analysis; this selection is designed to reveal problems, not estimate their production prevalence. Define the required facts and approval criteria before testing the intervention.
Compare the existing method with a source-linked evidence view on the same underlying cases. Balance reviewer assignment and order where practical; a reviewer who has already solved a case may otherwise make the second method appear easier. Record that learning effect as a limitation if the team cannot control it.
- Record preparation, verification, correction, approval, and any later rework. Count unsuccessful attempts and the work spent recovering from them.
- Have a domain reviewer check required facts against the records. Treat an accepted report, rather than a generated draft, as the unit of completion.
- Keep the output in an internal draft environment. Report approval remains with the authorized person; no external delivery is part of this experiment.
- Compare the evidence view with the simpler template and source-ownership improvements. Explain which part of an observed gain each could account for.
5. Decide in advance what would change the plan
For this hypothetical planning decision, suppose the service lead considers a repeatable reduction of at least ten human minutes per routine report worth a larger trial. That is a local investigation threshold chosen for this example, not a general rule or a statistical acceptance test.
The quality condition is equally explicit: required facts remain covered, known exceptions are visible, and the reviewer can trace material statements to the relevant record. Any material unsupported statement prompts investigation before expansion. A small sample with no observed failures cannot establish that production failure risk is negligible.
- Continue to a limited live trial if the effort reduction is consistent enough to investigate further, the quality conditions hold in reviewed cases, and ownership is workable.
- Modify the approach if effort merely moves between stages, source preparation dominates, or the benefit depends on a narrow subset of reports.
- Stop this implementation direction if review becomes less reliable, source access cannot be provided appropriately, or ongoing costs exceed the plausible benefit.
6. Bound the economic claim
If a future representative trial supported ten minutes less human effort across all 80 monthly reports, the gross recovered capacity would be about 13.3 hours per month: 80 × 10 ÷ 60. With three human hours of monthly maintenance, the illustrative net would be about 10.3 hours before accounting for implementation effort and monetary operating costs.
This is a conditional calculation. The current observations do not support those inputs, and an exception-heavy workload might have a different result. Tooling and inference charges, access and integration work, reviewer training, and the cost of errors belong in the investment decision.
Recovered hours have value only if the team can use them. They may improve turnaround, reduce overtime, or free capacity for another task; they do not automatically become a cash saving. Establish the expected use of that capacity with the decision owner.
7. Recommendation, ownership, and handoff
Recommend the verification experiment and defer the full drafting build. The hypothetical service lead owns the continue, modify, or stop decision. A domain reviewer defines and checks acceptance; an engineer prepares the limited evidence view and records its operating assumptions.
The handoff consists of this memo, the case list and inclusion rationale, the effort log, the acceptance checklist, and the results with unresolved questions. Set the review for when the agreed cases have been examined, with a time limit agreed before work begins. If access or reviewer capacity is unavailable, report that constraint and revisit the plan.
The client can run the next experiment internally with these artefacts. Additional engineering support becomes useful when there is an agreed implementation task or an evidence gap that the team cannot reasonably address itself.
8. What this memo cannot conclude
Every observation in this example is synthetic. Even in a real engagement, a short investigation of selected cases would provide evidence for the next bounded decision, not a validated annual return or a production-readiness certification.
Before expanding, examine a broader workload, important exceptions, source changes, the actual maintenance burden, and how the intervention affects the people responsible for the complete report. The method is useful when it makes those commitments more explicit and directs the next investment toward evidence that could change the decision.