A demonstration can make an AI workflow look straightforward. The person using it still needs to know which records to trust, what to check and what to do when a step fails.
Use a rehearsal when you have a proposed task and can involve someone who does that work. You will leave with a decision about the next test, a list of changes and a person responsible for each change. You can do the exercise on paper before connecting anything.
1. Choose a complete, small task
Start with a trigger and finish with a result another person can use. “AI for customer service” is too broad. “Prepare a reply to a delivery-date question for a service coordinator to approve” is testable.
Write down the current way of doing the task, including the tool or record that receives the result. Keep that process available during the rehearsal. Identify the employee doing the work, the reviewer and the person receiving it. One person may fill several roles; the responsibilities still need to be clear.
2. Agree what a usable result contains
List the facts the output must include and the claims it must avoid. Name the approved sources, their dates and their owners. Decide what happens when sources conflict or information is missing.
For a delivery reply, acceptance might require the correct order, a date supported by the current schedule and a clear statement that the coordinator must approve sending. A confident sentence with an invented date fails the task.
Check access before using real information. Use fictional or appropriately cleared records if the rehearsal environment is not approved for customer data.
3. Rehearse three cases
Run a normal case, an unclear case and a failed step. Let the employee follow the instructions without the builder translating every prompt. Observe where they hesitate or leave the workflow to finish the job elsewhere.
For each case, record preparation, review, corrections and recovery. Ask the employee to explain why they accepted or rejected the result. That explanation is useful training material: it shows the judgment the task needs.
Keep actions separate from drafts. If the proposal includes updating a system, specify the permission and approval needed, then check the result in the tool. A preview of an update does not establish that the update happened.
Worked fictional example
A service coordinator currently reads an order record and a delivery schedule, drafts a reply and gets approval from the service lead. The proposed AI step prepares the draft from those same two sources. It cannot send.
In the normal case, the draft cites the current schedule and leaves the delivery date for approval. The lead accepts it for sending through the existing process.
In the unclear case, the schedule has no date for that order. The draft must say the date needs checking and route the question to the scheduler. Guessing a date is a failed result, even if the prose reads well.
In the failed-step case, the order record is unavailable. The coordinator uses the existing manual process and logs the failed lookup for its owner. Nobody tells the customer that an unverified date is confirmed.
The team discovers that its draft instructions do not identify the scheduler. Its next action is to fix that handoff and repeat the unclear case. These are fictional observations demonstrating the method, not evidence of a working installation or time saved.
4. Decide what needs to change
Sort the findings by consequence. Missing permission or unreliable sources may prevent a live trial. Awkward wording may be easy to fix. A workflow requiring constant rescue from its builder needs more work before wider training.
Give each change an owner and a way to check it. “Improve the prompt” is vague. “Require the current schedule reference; repeat the missing-date case with the coordinator” is a checkable change.
Compare the proposal with keeping the current process or configuring a verified feature in an existing tool. Rehearsal can reveal that a simpler option serves the task better.
Where this exercise stops
A rehearsal with a few cases does not establish reliability, savings, permissions in production or employee adoption. Higher-consequence actions need appropriate specialist review and broader testing. Do not send messages or change customer records just to demonstrate a rehearsal.
Your next action: pick one task, name the person who will use the result and write the three cases before arranging training.
The blank Rehearsal card is available by email, without a Field Notes subscription. You can use the steps above without subscribing. Keep your completed card in your own files.
If the task depends on unstated knowledge, use the knowledge handover guide to capture the decisions behind it.
If the work needs help across several tools or teams, book the diagnostic call. You can also contact BR8N.