In this note
The demo classifies a support ticket and drafts a convincing explanation. The team asks for another month of engineering. Nobody has watched a reviewer check a batch, agreed which mistakes are unacceptable, or measured whether fixing the drafts takes longer than doing the work manually.
I'd ask which unanswered question could change the decision to build. Then fund the smallest permitted piece of work that can answer it. In the hypothetical support example below, that means planning a short investigation while keeping the prototype out of the live queue.
Say why you cannot approve the next stage yet
The proposed service would prepare a ticket's category, urgency, and supporting facts for a reviewer. It would not contact customers, close tickets, or change access. The support manager has proposed the job, but no target team has approved its acceptance rule, evidence, or operating plan. A polished demo cannot supply those decisions.
| Gate | Status and missing answer in this example |
|---|---|
| Owned outcome | Unknown. Confirm which tickets qualify, what the reviewer must accept, the baseline, and who owns the result. |
| Independent verifier | Unknown. A senior reviewer is proposed, but the expected answers, rejection test, and available review time have not been checked. |
| Bounded authority and loss | Unknown. Owners must decide which data and destinations are permitted, what the service may do, which errors or losses are unacceptable, and when to stop and recover. |
| Positive value after full cost | Unknown. Review effort, useful accepted volume, support cost, and remaining harm or loss have not been established. |
Keep those four answers separate. A low model bill cannot make a missing permission acceptable. A reviewer agreeing that the idea sounds useful cannot establish that the service saves time. While any gate is failed or unknown, fund only authorized discovery; do not start or continue a live pilot.
Fund a question the team can answer
Here, I'd propose up to six staff-hours to investigate whether preparing these drafts could reduce review work without missing urgent tickets. Start by confirming the workflow owner and permitted source access. Only after access is approved should anyone inspect the agreed support records. If that permission is unavailable, stop the case work and return the access gap to its owner.
| Work allowance | Answer to bring back |
|---|---|
| Up to one hour: confirm owners and permitted evidence | Who can approve the work and its data use? Which records may be inspected? If access is unresolved, do not begin the conditional case work. |
| Up to three hours: permitted case walkthroughs | What does the reviewer check today? Which errors should stop a draft? Does the checker reject a deliberately wrong category or urgency recommendation whose correct answer was checked separately? Record the observations and their limits. |
| Up to two hours: compare options and record the decision | Compare the draft proposal with today's process or a smaller repair. List costs and unresolved risks, then recommend the next investigation, narrowing, deferral, or stop. |
- Maximum planned staff time
- 1 + 3 + 26 hours
- Labor allowance
- 6 hours × USD 150/hourUSD 900
Count everyone's time: a one-hour walkthrough with two people uses two staff-hours. USD 150 an hour is an invented planning rate, not a price recommendation. The USD 900 covers only this labor allowance. The proposal assumes an existing approved workspace and no new licenses or provider calls. Any other expense needs its own estimate and approval before it is incurred. Unused hours do not authorize the team to expand the task.
Choose the limit to suit the real question. Six hours may reveal a missing field or an unusable review step; it cannot establish a reliable error rate, sustained adoption, or production economics. If getting enough evidence needs more work, return with that specific request instead of calling the incomplete result a passed pilot.
Bring back what would change the choice
Keep each observation beside its source. A walkthrough shows how one reviewer handled those cases. A local test shows whether its checker rejected a particular bad output. An estimate shows what someone expects to spend. None is evidence that customers or the operating team have accepted a live service.
If reviewers spend most of their time finding a missing policy, investigate that source problem before improving the draft. If a simple form or rule removes enough work, compare that repair with the assistant. If review remains expensive or no qualified checker has time, narrow the job or stop. Preserve findings that weaken the original proposal.
Finish discovery with a new decision
At the time limit, or earlier if a prerequisite blocks the work, take the findings to the sponsor and workflow owner. Record what was learned, what remains unknown, the time and cost used, and what another bounded step would settle. The sponsor can approve funding; data, policy, risk, and release owners still control their respective decisions.
Reassess all four gates. Unknowns still permit authorized discovery only. Even if the four investment gates later pass, live use needs separate technical tests, operator acceptance, adoption and value checks, operating capacity, and release approval. Trying the service against live tickets without using its answers is often called a shadow trial; that also requires separate permission. The six-hour allowance authorizes neither kind of trial nor sending customer data to a model.
The proposed next decision is to authorize this limited discovery plan, subject to its access conditions, while deferring the pilot and rollout. No investigation has run or target gate passed. Name who will answer the first missing question and when the sponsor will review that answer.