# Client discovery lab: earn the first delivery decision LockedIn Labs · Original training material · Edition 2 Practice the client conversation before building the assistant. In 30 minutes, turn a sponsor's request into a bounded experiment that an operator and receiving engineer can challenge. This is a free role-play with supplied synthetic facts; it needs no account, model, patient data or production access. Use two people: one interviewer and one partner playing the four stakeholder roles below. The interviewer reads only the opening brief before starting; the partner reads all four role cards. With five people, assign one role each. Working alone, write your questions first, then uncover the cards and mark which facts your questions would have elicited. Solo work cannot establish interview skill with another person; retain that limitation in your reflection. ## Opening brief — interviewer reads this A fictional front desk answers parking-access, form-location and desk-hours requests. The sponsor wants an AI assistant and asks: “Can you show automatic replies next week? We spend too long answering the same things.” Your task is to establish what happens today, identify a small useful change and read back a proposed experiment. You are not authorized to promise integrations, automatic sending, a delivery date or measured savings. Ask for a concrete recent example and evidence that could change your proposed solution. ## Run the conversation | Time | Task | Keep | | --- | --- | --- | | 0–3 min | Read the opening brief; write your first three questions | Questions before seeing answers | | 3–15 min | Interview the four roles; follow the strongest discrepancy | Notes with role and evidence ID | | 15–20 min | Read back a first-slice agreement | Outcome, comparison, boundaries and owners | | 20–25 min | Introduce the change card; revise one decision | Before/after wording and reason | | 25–30 min | Receiver critique and one unanswered question | A specific next evidence request | In the Academy discovery lesson, the first 15 minutes can supply the interview portion of your existing exercise. The full standalone lab adds the change and receiver critique. Do not count the same work twice in your study record. ## Role cards — partner reads these Answer the question asked, in your own words, without inventing missing facts. Reveal related evidence when the interviewer follows up. This is not a hidden keyword game: any question with the same intent earns the relevant detail. If a fact is absent, say “not established in this packet” and identify who could know. The interviewer can ask to see the whole packet at any point; record that choice. ### 1. Operations sponsor Your opening position is speed: you remember a reply taking two minutes. - **S1 — unit of work:** when asked what the two minutes includes, explain that it is composing the reply after finding the answer. You have not timed the complete task or sampled the whole queue. - **S2 — decision:** you can accept an experiment for one administrative queue. You need the operator to judge usefulness and the receiving engineer to judge operation. You cannot approve another team's data access or source versions. - **S3 — business question:** you want to learn whether operators can finish the same task with less effort. A simple procedure card is an acceptable candidate if it helps. No saving, budget or rollout commitment is established. Ask the interviewer at read-back: “What will this teach us that a fast demo won't?” ### 2. Front-desk operator Start with: “Finding the right instruction is the difficult part.” - **O1 — reconstructed request:** for synthetic request R-014, you spent one minute reading, six comparing procedure copies, two composing and one checking. It waited 40 minutes before you opened it. No timestamp export is supplied; these are authored exercise timings, not a real time study. - **O2 — accepted output:** you need a short form-location answer supported by the current source. You reject added submission requirements or a guarantee that a form will be completed today. You retain the communication decision. - **O3 — comparison sample:** supplied active handling times for six manual requests are 8, 9, 10, 10, 11 and 12 minutes. Their categories and individual source versions are not supplied, so do not call them matched comparison pairs. - **O4 — alternative:** a reliable current-source card might remove repeated lookup. Whether generated wording adds value beyond it is still unknown. Ask at read-back: “Where will you count time spent correcting or rejecting a draft?” ### 3. Procedure owner Start with: “Two copies can look convincing and still disagree.” - **P1 — authoritative fixture:** ADMIN-FORM-04, version 3, tenant-a/location-a, says general administrative forms are available at reception. It establishes neither submission requirements nor individual status or completion timing. - **P2 — conflicting copy:** a saved version-2 copy points elsewhere. In this exercise it is superseded; popularity or a recent file-copy date cannot make it current. Version 3 is the supplied authoritative fixture for the rehearsal. - **P3 — open dependency:** the owner can resolve source disputes. The mechanism for distributing updates into a future live service has not been established. A correct local fixture does not resolve that integration dependency. Ask at read-back: “Who notices a withdrawn procedure before the next draft?” ### 4. Receiving engineer Start with: “Show me what we would have to own.” - **E1 — environment:** local synthetic fixtures are available. Live identity, source-store integration and deployment access are not established by this packet. Do not substitute request-body fields for trusted authorization. - **E2 — operation:** the first slice can read authorized administrative guidance and propose a draft. Missing, withdrawn or unauthorized sources need a clear outcome and the existing manual path. No outbound sender is part of this slice. - **E3 — handoff:** you need the scope, result contract, failure behavior, source ownership and unresolved dependencies. You cannot accept a release solely from a screenshot or a sponsor's approval of a demonstration. Ask at read-back: “Which dependency prevents a live trial, and what can we learn locally?” ## Change card — introduce after the first read-back The sponsor now says: “The complaints I hear are about the 40-minute wait before anyone opens the request. Faster writing may not solve that. Can the assistant prioritize the queue instead?” Do not invent a new queue study. Record this as a changed priority and an unverified explanation, not proof that every request waits 40 minutes. Ask what queue timestamps, categories and assignment decisions would clarify the problem. Name who can supply approved synthetic or appropriately authorized evidence. Explain whether you would pause, narrow or retain the drafting comparison. Automatic prioritization requires its own scope and impact decision; this card provides no authority for clinical triage or other consequential decisions. ## Your first-slice agreement Fill this in before looking at the protected worked interpretation: ```text Participant roles / simulation date: Recent request and evidence IDs: Active effort / queue waiting / how each was obtained: Observed discrepancy and its explanation: Useful output / user who decides it is useful: Simplest candidate / alternative / evidence that could favor the alternative: Comparison unit / treatment of checking, correction, rejection and fallback: Permitted categories / source and scope / forbidden side effects: Decision owners: Unresolved dependency / owner / next evidence request: Observable stop condition and manual next step: Change-card decision / reason / uncertainty retained: ``` Keep a short evidence ledger: **statement | role and ID | supplied fact, hypothesis or decision | next action**. A stakeholder quotation is evidence of what they said; it is not automatically evidence that their causal explanation is true. Mark invented extensions separately from the supplied packet. ## Optional extension: compare the whole task Spend another twenty-five minutes preparing a recommendation from the synthetic comparison below. This is a separate, deliberately matched teaching fixture; it does not supply the missing categories or source versions for O3's six requests. Each row represents the same administrative task completed under three authored conditions: the existing manual procedure, a current-source card and an assisted draft. No people were timed and no model or production service ran. Before reading the rows, keep this supplied sponsor constraint: a candidate must finish every task in this packet within twelve active minutes, including rejected drafts and manual fallback. This is an exercise decision rule, not a real service level. The operator retains sending authority in every condition. None of these times measure queue waiting, and this comparison cannot answer the change card's question about prioritization. For the assisted condition, draft preparation, review, repair and fallback are sequential, non-overlapping minutes. Review is required even when a draft is rejected. A rejected draft has zero repair minutes and then a full manual fallback, which is additional to the draft and review already spent. The manual and source-card columns already include their checking and completion work. Accepted means accepted within this fictional packet, not independently judged safe or qualified for live use. | Case | Manual total | Source-card total | Draft | Review | Repair | Fallback | Draft decision | | --- | --- | --- | --- | --- | --- | --- | --- | | C1 | 8 | 6 | 1 | 2 | 0 | 0 | Accepted | | C2 | 9 | 6 | 1 | 2 | 1 | 0 | Accepted after repair | | C3 | 10 | 7 | 1 | 3 | 0 | 10 | Rejected; manual fallback | | C4 | 10 | 7 | 1 | 2 | 0 | 0 | Accepted | | C5 | 11 | 8 | 1 | 3 | 2 | 0 | Accepted after repair | | C6 | 12 | 8 | 1 | 2 | 0 | 12 | Rejected; manual fallback | | C7 | 8 | 6 | 1 | 2 | 0 | 0 | Accepted | | C8 | 12 | 8 | 1 | 2 | 0 | 0 | Accepted | ### Make a first recommendation before comparing Compute an assisted total for every row. Keep all eight tasks in each condition's denominator, then calculate the total and mean active time. Separately report the draft acceptance count, tasks exceeding twelve minutes and the largest task time. Compare assisted work with the source card as well as with manual work. A faster candidate than the old procedure may still be weaker than a simpler alternative on the decision that matters. Write a six-sentence recommendation: the supplied question, comparison unit, full-task result, constraint result, candidate to investigate next and one piece of evidence that could overturn your recommendation. Keep the result before reading the worked comparison. Do not multiply eight synthetic tasks into an annual saving or infer that shorter task time reduces queue waiting. ### Worked comparison — keep rejected work in the denominator The assisted totals are **3, 4, 14, 3, 6, 15, 3 and 3 minutes**. Across all eight tasks, manual work totals **80 minutes** (mean 10), the source card **56** (mean 7) and assisted work **51** (mean 6.375). The assisted condition uses five fewer minutes than the card in aggregate, but its fourteen- and fifteen-minute tasks violate the supplied per-task constraint. Both other conditions meet it in this packet. Those observations support investigating the source card under the declared rule; they do not establish that a source card is always the best design. Six of eight drafts are accepted, including two after repair. Counting only those six tasks gives 22 assisted minutes and a mean of about 3.67. Presenting that as the full workflow average would omit the 29 minutes spent on the two rejected attempts and their fallback. The original denominator and all four time components must survive the recommendation. A one-minute draft is not a one-minute completed task. The fixture gives each case equal weight. It supplies no real queue mix, repeated runs, learning effects, error severity, uncertainty estimate or operating costs. A mean here describes these rows; it is not a measured productivity gain or an investment return. If the sponsor changes the twelve-minute constraint, retain the original decision and explain the new tradeoff rather than rewriting the comparison to hide the slower tasks. ### Turn the recommendation into an evidence request Ask the operator which completed-task evidence would distinguish source lookup from drafting effort. Ask the procedure owner how to establish the current source for each comparison. Ask the receiving engineer how review, rejection, fallback and source maintenance would be observed without collecting private request content. The sponsor must name the decision and acceptable failure boundary before an authorized trial; this worksheet does not authorize one. For a future approved comparison, propose a record with task category and source version, condition and order, start/finish definitions, component times, result disposition, correction/fallback, assistance and missing observations. Explain how different operators, practice and order could affect the result. Do not call a missing or abandoned task a zero-minute success. Your receiving partner should be able to locate C3 and C6 and reproduce your full-task recommendation from the record. Keep the first recommendation, worked comparison and revised evidence request with the discovery agreement. ## Receiver observation sheet Record an observed example or “not observed” for each item. This is peer feedback, not a score, credential or independent qualification decision. - The interviewer asks about a completed request before choosing a tool. - The two-minute and ten-minute accounts are reconciled without dismissing either role. - Active effort is distinguished from queue delay, and the small sample's limits survive. - The agreement could reasonably select a source card instead of model drafting. - Authority, source ownership, failure and receiving responsibilities are explicit. - The changed priority causes a reasoned revision rather than an unsupported promise. For the final minute, the receiver names one decision they still cannot accept. The learner writes one follow-up question, its intended respondent and the specific evidence it would obtain. Save the original agreement and the revision as your discovery portfolio artifact. Worked interpretation and guided review remain in the existing Academy discovery lesson.