One strong answer can save an afternoon. A reliable team process takes a little more work: the same kind of input, a clear result, a person who checks it, and a way to recover when it goes wrong. This is the step from “I tried AI” to “we know when and how to use it.” It does not require building an agent or buying a new platform.
Keep the process small. The fictional workshop team wants help drafting attendee replies about the schedule. It does not want the assistant promising refunds, deciding accessibility arrangements, or sending messages without approval. That boundary makes the first workflow useful and testable.
A small team workflow
Question
A real, eligible item
Draft
AI uses an approved source
Review
A person checks and may reject
Send + learn
Record edits and fix the source
Write the handoff in ordinary language
A workflow is a set of steps people can repeat. For this one: a staff member selects a routine attendee question; the assistant drafts using an approved fact sheet; a person checks the answer against the sheet; the person sends or rewrites it; the team records corrections. Those steps are more important than the name of the AI tool.
Separate “draft” from “send.” In a low-risk personal task you might accept a first answer yourself. In a customer-facing process, a reviewer needs permission, enough time, and a clear way to reject the draft. An approval checkbox with no source material is not a meaningful check.
Try this
Copy a reusable drafting prompt
Use only the fictional facts below; a person must approve the reply.
Draft a reply to the attendee question below using only the approved fact sheet. Keep it under 100 words. If the sheet does not answer the question, say we will confirm rather than guessing. Do not promise a refund, accessible accommodation, parking, or a registration change. After the draft, list each factual sentence and the supporting line from the sheet. APPROVED FACT SHEET: The fictional workshop lasts two hours and has space for 40 attendees. A hands-on activity lasts at least 30 minutes. The date, venue, price, parking, accessibility arrangements, and refund policy are not confirmed. ATTENDEE QUESTION: Will there be parking, and can I get a refund if I cannot attend?
Why it works: A good result here says what is known and routes both unanswered questions to a person. It does not invent a customer policy.
Test the awkward cases before the easy cases become habit
Make a short test set: five routine questions, three missing-fact questions, and two cases that should be escalated. A missing-fact question is not a nuisance; it tells you whether the assistant knows when to stop. Record how many drafts a reviewer accepted, how many needed edits, and whether any unsupported promise appeared.
Set a stop rule before wider use. For example: pause the trial if a draft promises a refund not in the fact sheet or if the reviewer cannot check the answer within the agreed time. The specific threshold belongs to the team, but it must exist before volume makes mistakes harder to spot.
Keep the process repairable
When a reviewer corrects a draft, ask why. Was the fact sheet stale, the question outside scope, the instruction unclear, or the model wrong despite clear evidence? Fix the right part. Rewriting the prompt cannot repair a missing policy; a meeting cannot fix a broken source link.
Choose a named owner who can update the fact sheet and a small review cadence. If the process depends on one enthusiastic person working late, it is not yet a stable team workflow. The goal is not to remove every human touch. It is to make the human touch valuable and predictable.
Practice assignment
Build a one-page workflow card
Design a draft-and-review process for a repeated question your team handles, using fictional or approved material.
- 1Name the input, approved source, AI task, reviewer, and final action.
- 2Write three questions the AI may answer and two it must escalate.
- 3Test ten sample questions and record accepted, edited, and rejected drafts.
- 4Write one pause rule and one owner for source updates.
What to save
A workflow card, ten-case test log, and a go, change, or stop recommendation.
You are ready when
Another teammate could follow the card, spot an unsupported promise, and know who can stop the trial.
Further reading
Sources and guidance behind this lesson, not required reading to finish it.
- OpenAI Academy: Prompting
Updated September 2026; clear task, context, output, and iteration.
- OpenAI Academy: Responsible and safe use of AI
Check workplace rules and use appropriate human review.
- MIT Sloan: Hidden work behind AI experimentation
Why sustained expert testing and review need time and support.