Your Team Has Never Agreed on What Makes an AI Answer Good Enough. This Workshop Forces the Conversation.

Most teams using AI right now are operating on an assumption nobody actually tested: that everyone knows what "good enough" means. They don't. Ask five people on the same team when they'd reject an AI output and you'll get five different answers, some based on tone, some on speed, some on a gut feeling they can't quite explain. Nobody wrote it down because nobody thought to.

This gap rarely shows up in a demo. It shows up three weeks later, when one person on the team ships an AI drafted client email nobody else would have sent, or when a report goes out with a number nobody double checked because "it looked right." By then it's not a training problem, it's a trust problem, and trust problems are expensive to repair.

The fix isn't a better prompt library. It's a conversation the team has never had out loud: what specifically makes an AI answer usable here, in this role, on this kind of task. The workshop below is built to force that conversation in 90 minutes, using a real task, not a hypothetical one.

Objective

Get a working team to agree, explicitly and in writing, on the criteria that decide whether an AI output is good enough to use, then pressure test those criteria against something they're actually working on this week.

Format

Participants: 4 to 8 people from the same working team. Mixing departments dilutes the exercise, this works best with people who share the same tasks and the same stakes.

Duration: 90 minutes.

Materials: sticky notes or a shared digital board, a laptop with access to whatever AI tool the team already uses, a timer, and one real task per pair that they are currently working on or about to start.

Run sheet

1. Frame, 5 minutes. Explain that today isn't about writing better prompts. It's about naming, out loud, what "good enough" actually means for this team's work, because almost no team has done this.
2. Silent brainstorm, 10 minutes. Everyone writes on sticky notes what makes them reject an AI output on sight, before reading further. One idea per note, no discussion yet.
3. Cluster, 10 minutes. Group similar notes on the board and name the clusters: wrong tone, missing context, too generic, factually shaky, whatever emerges.
4. Rank, 10 minutes. Dot vote on the three or four clusters that matter most for this specific team's work.
5. Draft the threshold, 15 minutes. In pairs, turn the top clusters into a one page checklist: "before I use this, I check for X, Y, Z."
6. Live test, 30 minutes. Each pair picks one real, current task, runs it through AI live, and scores the output against their own checklist, out loud and honestly.
7. Revise, 10 minutes. Something the checklist missed will almost always surface here. Edit it on the spot.
8. Commit, 10 minutes. Each person states one line they'll add to their own workflow starting Monday.

Debrief questions

Where did your checklist fail to catch a weak output. Was there a moment you were tempted to accept a mediocre answer because you were rushed, and what triggered that. What does a bad answer actually cost in your role, and does your checking match that cost. Whose checklist item surprised you the most.

Common failure modes

Skipping the live test. Groups that stay theoretical produce a checklist that sounds good and gets ignored the following week. Protect the 30 minutes, it's the point of the exercise.

Letting senior voices dominate round one. Silent writing before discussion is what prevents this. Don't skip it to save time.

Choosing a task that's too safe. If the task is one nobody's worried about, the exercise proves nothing. Push the group toward a task they're genuinely unsure of.

Turning the checklist into a policy document. Keep it to one page, one team, plain language. The moment it reads like compliance, people stop using it.

Why this matters more than another prompting session

This exercise sits closer to the Knowledge and Alignment parts of AI adoption than to the tools themselves. Teams don't fail at AI because they lack prompts, they fail because nobody agreed on the bar. A shared, explicit threshold is cheap to build and expensive to skip.

Try this week

Before your next team meeting, ask everyone to write down, silently and without discussing it first, the one thing that makes them personally reject an AI output. Read them out loud together. That's the whole exercise compressed into ten minutes, and it usually surfaces at least one disagreement the team didn't know it had.