At a fully powered intake desk, every coordinator prepares the packet your best coordinator would prepare, on every new inquiry. The party names for the conflicts clerk are complete. Missing details have already become questions for the prospective client. The inquiry with a hearing date sits at the top of the pile. That is 10X what your intake desk can do, with the team you already have.
Proof makes that possible. We plan from the outcome, map the work that crosses teams and land each capability on the roles that do it. Then we prove it against your best person’s standard on real work, and only then roll it out.
Start with the best-person check.
The best-person check is the standard behind every test we run. Build the workflow from how your best person does the work: the sources they trust, the fields they never skip, the moment they stop and pick up the phone. Define what good is before the build, in terms a reviewer can see on the page. A result can differ in wording from a past example and still meet the standard, and a fluent one can fail because it leaves out something essential. Hold back hard cases the build never sees, so the test measures the instructions rather than the examples behind them.
Then let your expert judge the output on real work. Where a grading rule and your expert disagree, the disagreement goes into the record and you examine it together. Your expert has the final say.
Ask five questions of any workflow before your team relies on it:
- Who agreed the standard, and what makes a result good?
- Which cases were held back from the build?
- Who grades the output, and who has the final say?
- Will we see omissions and uncertain results alongside the passes?
- What did the workflow change about the whole job?
At a law firm’s intake desk, the expert is the senior intake coordinator. The firm administrator owns the result, and the conflicts clerk and the responsible attorney rely on the packet. From new inquiry to open matter maps that work between teams.
The test record, case by case.
The workflow under test is the inquiry handoff, which turns a new inquiry into a packet for conflict review and acceptance. People make every decision: the conflicts clerk runs the conflict review, the responsible attorney decides whether the firm takes the matter, and the intake coordinator sends every message.
Each row is one held-back case, with its expected behavior written before the first run. The fourth column says what the tester captures at each run.
Illustrative example with invented data. An invented 40-person litigation and employment firm. Expected behavior only; the actual output is recorded at each run.
| Case and input | Expected behavior | Exception to flag | Record at each run | Reviewer’s decision rule |
|---|---|---|---|---|
| Ordinary. Inquiry 1: lease dispute, both parties named. | Complete packet, every party listed, routed to the conflicts clerk. | None expected. | The packet and its routing line. | Fails if any field lacks a source or anything says a conflict is clear. |
| Missing information. Inquiry 2: no opposing party named. | Marks the party list incomplete, names the gap and drafts one question for the coordinator to send. | Opposing party missing. | The flag and the drafted question. | Fails if the packet reads as complete or a party is guessed. |
| Contradictory inputs. Inquiry 3: the form says tenant; the call notes say landlord. | Shows both with their sources and asks the coordinator to confirm. | Conflicting role. | Both versions as shown. | Fails if one version silently wins. |
| Ambiguous match. Inquiry 4: Company B LLC, close to existing client Company B Holdings. | Lists the possible match for the conflicts clerk. Makes no finding. | Possible related entity. | The match note, exactly worded. | Fails if the packet says the two are, or are not, the same. |
| Stop and ask. Inquiry 5: papers served and a hearing date mentioned. | Stops and flags it for the responsible attorney through the coordinator. Works out no deadline. | Time-sensitive matter. | The stop notice and anything else produced. | Fails if any date is calculated or the inquiry joins the ordinary queue. |
| Out of scope. Inquiry 6: the prospective client asks whether it can withhold rent. | Carries the question to the responsible attorney. Drafts no answer. | Request for legal advice. | How the question appears. | Fails if the packet contains any answer. |
The senior intake coordinator marks each run, and the firm administrator signs off. Each case runs more than once, because AI can answer the same request differently from one run to the next. A miss sends the work back to the instructions, the inputs or a rule nobody had written down. Then the whole set runs again.
Pilot acceptance-test preparation builds a record like this for any workflow your team is piloting, with a rubric and a pass rule. Our technical companion on testing covers run records, held-back answers and retesting after every change.
Ready means this task, in this environment.
Ready is exact: this version of the instructions, doing this task, in this firm’s approved AI workspace, met the expected behavior on every held-back case, and the expert agreed with every mark. That is the evidence a firm administrator needs to put it in front of the whole intake desk.
Decisions stay where they were, and every packet still gets the coordinator’s review. Change the matter type, the intake form or the AI platform and the affected cases run again. The limits go into the record in plain words, next to the passes.
Illustrative example with invented data. The limits section of the same record, as it would read after a passing test.
Limits written into the record
- Tested on web-form inquiries and typed call notes. Emailed inquiries with attachments are untested.
- Tested on lease and commercial disputes. Employment matters need their own cases before use.
- The party list is only as complete as the inquiry. The conflicts clerk runs the search.
- Retest after any change to the instructions, the intake checklist or the workspace.
Then the whole team gets it.
The owner introduces a ready workflow to the people who will use it every day and watches first use. The rollout checklist covers each step. Misses from daily use become new test cases.
The test runs where your team already works, in ChatGPT, Claude, Copilot or Gemini, and your records stay in your account. Your people set the standard and keep control of how the work gets done. Evals, validation and verification sets out the full method.
Bring one workflow your team wants to rely on and the cases your best person worries about.