Ask AI to play devil's advocate on a team decision

Garden · not yet reviewed Time: 25 minutes Good for: team lead, project manager, programme coordinator, committee chair, anyone about to commit a group to a plan

You'll end up with: A team decision that has been questioned before you commit to it, with a short table of the assumptions it rests on and one change or check you'll make because of them.

Tools: Any general AI chat assistant (free tier is fine). Leave out names, budgets or anything confidential your organisation wouldn't want in an AI tool.

Why this matters

Once a plan feels agreed, people stop raising doubts, partly to be polite and partly because the meeting is over. Most AI chat tools make this worse, because they tend to agree with whatever you bring them. Asking an AI to argue the other side, on purpose and one question at a time, gives a group a cheap way to hear objections before money or time is spent.

Smallest useful build

A short session with an AI devil’s advocate about one real decision, ending in a two-column table (the decision; the assumptions and alternatives it raised) and one sentence on what you’ll change or check. About 25 minutes.

Sample inputs

Use a decision your team has made or is about to make. If you need one, use this invented example:

Decision: Move our monthly all-staff meeting from in person to fully online, starting next quarter.
Reasons: two of our five offices are now remote-first; travel costs went up 30% this year; attendance at the last in-person meeting was 60%.
Goal: keep staff informed and connected without the travel.

Steps

1. Write your own objection first. Before opening the AI tool, write the strongest reason against the decision that you can think of, and one thing that would have to be true for the decision to work. Keep both; you’ll compare them later.

2. Start the devil’s advocate. Open a new chat and paste this prompt. It is adapted from the Wharton Generative AI Labs “Devil’s Advocate” prompt by Ethan Mollick and Lilach Mollick (CC BY 4.0):

   You're a friendly, direct teammate whose job today is to play devil's advocate. Teams often fall into a consensus trap, where nobody questions a decision once it feels agreed; your role is to question it respectfully. First ask me what the decision or plan is. Then ask one question at a time, and wait for my answer before the next, until you understand the goal. Then ask me, one at a time: what alternative views exist, what the drawbacks would be if we go ahead, what evidence supports the decision and what we're assuming. Keep questions short and aimed at hidden assumptions. When we've covered the main weak points, stop and give me a table with two columns: "Decision" and "Hidden assumptions or alternative views".

3. Answer honestly, and push back once. Answer each question as you would to a sharp colleague. If a question misses the point, say so and see whether it finds a better one. If it starts agreeing with you, ask: “What would someone who opposes this decision say next?”

4. Decide what changes. Compare the table with the objection you wrote in step 1. Write one sentence: “Because of ___, we will change ___ / check ___ before we commit.”

Check it

• The table lists at least one assumption you hadn’t written down in step 1.
• Each assumption in the table is specific to your situation, for example “remote staff will join on camera”, rather than general, such as “people may resist change”.
• Your final sentence names a concrete change or check, with an owner or a date.

Break it

Start a new chat and present the same decision as if you’re unsure about it (“I’m worried this is a mistake: …”). If the AI now argues the opposite of what it argued before, it’s mirroring your mood. Weigh only the points that appeared in both sessions.

Reflect

• Did the AI raise anything your team would have been uncomfortable saying aloud?
• Was your own objection from step 1 stronger or weaker than the AI’s best one?

Next iteration

Run it with the whole team: one person types, everyone answers. Then try two related prompts from the same library, “Team After Action Review” after a project ends and “Team Charter” when a new group starts, and adapt whichever helps into a team habit.

Connections

Skills it uses

• /expert-council (garden)
• /stuck-or-pivot (garden)

Concepts it practises

• Productive friction (garden)
• Sycophancy (garden)
• Clarifying questions (garden)
• Human in the loop (garden)

Competencies

• Frame the outcome
• Steer
• Check the result

Background reading

• Wharton Generative AI Labs Prompt Library (garden)
• The Cybernetic Teammate (garden)
• A top researcher says a new divide is emerging in AI use, and most people are on the losing side (garden)

Do this one first

• Let AI ask before it plans (garden)

How we know

The claim

Groups often fall into a consensus trap, where nobody questions a decision once it feels agreed. An AI set up as a devil's-advocate teammate, asking one question at a time about assumptions, drawbacks and alternatives, helps a team surface what it hasn't considered.

The evidence

Based on the "Devil's Advocate" prompt in the Wharton Generative AI Labs prompt library by Ethan Mollick and Lilach Mollick, licensed CC BY 4.0. The library describes its prompts as evidence-based; no study of this specific prompt was found. The adapted prompt below is a shortened rewrite for work teams. The step order (you write your own objection first) and the checks are the commons' own.

The limits

The AI only knows what you tell it, so it can question your reasoning but not your facts. Chat tools tend to agree with you, so a weak devil's advocate can make a plan feel tested when it wasn't. One session doesn't show the decision is right.

---

Plain Markdown for agents: /projects/prj-devils-advocate-for-a-team-decision.md