The outcome.
Every pilot ends in a clear decision, backed by evidence the whole team can read. Before the workflow touches real work, the pilot owner holds a complete acceptance test, written to your best person’s standard. A pilot that passes becomes a capability the whole team relies on from its first week.
Who starts the work, and when.
The pilot owner starts it: the operations lead, department head or firm administrator accountable for the workflow’s result. They start once the purpose is agreed, before anyone has seen an output. The expert, usually the person who does the work best today, supplies the hard cases.
The AI assistant prepares the test pack. The owner and the expert run every test, mark every result and decide.
What they supply.
- The workflow’s purpose.What it prepares, for whom, and where its job ends.
- Its inputs and output.The records it reads and the output a person reviews.
- A few past examples the team trusts.Finished work the expert would sign, with the inputs behind it.
- The known awkward cases.The files, accounts or requests that trip people up, one sentence each.
What comes out.
An acceptance test pack: cases with their expected behavior, a rubric, a record sheet and a pass rule.
Illustrative example with invented data. An invented 60-person software company’s client-onboarding handoff. The owner’s brief, then the pack prepared from it.
Pilot brief: client-onboarding handoff
- Purpose: prepare the handoff record a client success manager reads before kickoff.
- Inputs: the signed order form, the account record and the account executive’s call notes.
- Output: what was sold, the start date and whether it is confirmed, contacts, sales commitments, open questions, next action.
- Awkward cases: an unconfirmed start date, a missing billing contact, a promise the order form does not show.
| Case | Invented record | Expected behavior |
|---|---|---|
| Ordinary | Order 3317. All three sources agree. | Complete record, every item sourced. Next action: the client success manager confirms kickoff. |
| Missing information | Client C. No billing contact anywhere. | Marks the record incomplete and drafts one question for the account executive. |
| Contradictory inputs | Order 3320. The form gives one start date; the notes say the client asked for an earlier one. | Shows both dates with their sources, marks the start unconfirmed and asks the account executive. |
| Ambiguous match | Client D. Two contacts titled project lead. | Shows both and asks which leads onboarding. |
| Stop and ask | Order 3325. The notes promise a discount the form does not show. | Stops and refers it to the sales lead. States no discount as agreed. |
Rubric, record sheet and pass rule
- Completeness
- Every required field present, or marked missing with a question.
- Factual support
- Every statement traces to one of the three inputs.
- Correct recipient
- Record to the client success manager, questions to the account executive, promises outside the order form to the sales lead.
- Next action
- One next action, with its owner.
- Record sheet
- Case, instruction version, AI workspace, output as produced, a mark per criterion, the reviewer’s reason.
- Pass rule
- Two runs per case. Every run meets its expected behavior, and the expert agrees with every mark. A miss on the stop-and-ask case fails the pilot. Any other miss is fixed and the set rerun.
Expected behavior only. The owner and the expert record the actual outputs when they run the test.
A good pack lets a reviewer who has never seen the workflow mark any run.
Where it fits.
- Operations and process improvement, where a changed process proves itself in a pilot.
- IT and systems, to confirm what the workflow reads and where its output goes.
- Law firms, for intake and matter opening, where stopping to ask matters most.
- Software technology, for the handoff from sales to client success.
- Growth & scaling, so new hires inherit tested workflows.
What has to change for your company.
Your owner names the expert and sets the local rules: which record wins when two disagree, who receives each kind of question, how many runs each case gets and the pass rule. The test runs in the AI workspace your team already uses, on ChatGPT, Claude, Copilot or Gemini, with records your business permits there.
What has been checked.
Fully Powered AI wrote this capability template. It was editorially reviewed on 9 November 2026. It has not yet been tested in a client environment. It has not yet been tested as a packaged version in ChatGPT, Claude, Copilot or Gemini.
Testing it in your environment means preparing the pack for one real pilot. Your owner and expert then confirm that the cases, rubric and pass rule cover what they would check by hand. What we test before we call an AI workflow ready shows the record this produces.
Next step.
Bring the workflow you are about to pilot and two or three cases your expert worries about.
Related capabilities.
- SOP exception capture: find the awkward cases your test should cover.
- Account handoff: prepares the kind of record the example tests.
- Inquiry handoff: a law-firm intake workflow worth testing this way.
- From working workflow to daily use: the rollout checklist: takes a passed pilot to the whole team.