What AI can and cannot do in an innovation workshop
In the vendor demo the AI generates fifty ideas in ten seconds, and the sales rep pauses for applause. A facilitator in the room is thinking: that was never the problem. Teams do not lack ideas. They lack a clean record of what was said, a fair way to choose, and a sponsor who says yes. Here is what AI does well today, what it does badly, and the rules to read out before the session.
AI in the room is a scribe, not a participant
Facilitators fear a tool that takes over the craft and turns a conversation into a screen. Managers expect the tool to run the workshop so they can skip it. Both treat the AI as a participant. It is a scribe with a few extra tricks.
What AI does well today
Turning speech into text
Transcription is where AI most clearly takes load off the human scribe. It catches the half sentence that was interrupted and the remark from the back of the room. The transcript is raw material, not the record.
Sorting statements into ideas, questions, concerns and decisions
Given a transcript, a model can tag which statements are ideas, open questions, concerns or decisions. It gets most of them right, and the facilitator corrects the rest in minutes. Tools like IdeaHone do this directly from the transcript.
Spotting duplicates and near-duplicates
Four people proposing "a customer portal" in four phrasings becomes one idea with four authors. Tedious for humans, easy for a model, and the same idea no longer splits its votes.
Writing the first draft of the summary
A draft an hour after the workshop beats a polished one next week. The AI writes it from the record; the facilitator checks it against what was actually decided and sends it.
Scoring against written criteria and explaining the score
If the team has written criteria such as "reduces onboarding time for new customers", a model can score each idea against them and say why. The explanation is the useful part: the team argues with the reasoning, not with a number. The method is weighted scoring.
Playing a role on request
Ask it to be the sceptic, the customer or the finance director, and it attacks an idea from that angle. Used late, on the team's own ideas, this is a cheap stress test.
Translating
Mixed-language teams get the record in every language on the same day. Whoever kept quiet rather than speak English can read and comment in their own.
What it does badly or should not do
- Deciding. A score is an argument, not a verdict. Once a team accepts "the AI ranked it third" as a reason, the workshop has lost its purpose.
- Reading the room. It does not see the energy drop after lunch, the silence after the director speaks, or the quiet disagreement.
- Choosing the question. "How do we cut support costs" and "why do customers call us" lead to different rooms. The framing is the facilitator's craft.
- Generating ideas instead of the team. Fifty generic ideas crowd out the three real ones, and the team stops owning the result. If at all, generate late, to provoke, and label it.
- Replacing the sponsor's yes or no. The sponsor commits people and money. A sponsor who hides behind a tool has not decided.
Rules for using AI in the room
- The AI writes, people decide. Say it at the start and repeat it before the vote.
- Every AI output is labelled and editable. A tag or a different colour; nothing generated may pass for something a person said.
- Tell participants what the AI does before it does it. "It transcribes, sorts and drafts the summary. It does not score until you have written the criteria."
- Let it score only against criteria the team wrote. No criteria on the wall, no scoring; otherwise you get an opinion dressed up as a method.
- Use role play after the team's own ideas, not before. The sceptic is useful against something the team believes in; before that it is noise.
- Keep the human scribe for the first session. Run both until you trust the capture; AI meeting notes vs structured capture shows what to compare.
- Check the summary against the record before sending it. Every decision should point to the moment it was made; if you cannot find it, it was not decided.
The best test of an AI workshop tool is whether the facilitator talks less to the screen and more to the room.
The European AI Act (Regulation (EU) 2024/1689) brings transparency duties under Article 50 when people interact with AI, and content the AI produced should be marked as such. Check with your data protection officer before you record a workshop; transcription and GDPR covers the recording side.
Three ways AI costs you the room's trust
- The demo workshop. The first session shows off the tool. The team watches a screen for an hour and leaves with nothing it owns.
- Scores without criteria. Someone presses "rank these" before anyone has written what good means. The ranking looks objective and is not.
- Summary sent unread. One invented decision in the record costs more credibility than the tool saved you in time.
AI that suggests, people who decide
IdeaHone turns what is said into cards, scores ideas against your criteria with written reasons and drafts the summary. Every suggestion can be edited; no decision is made by a machine.
Can AI facilitate a workshop?
No. It can transcribe, sort, draft and score, which removes the clerical load. It cannot read the room, choose the framing or hold a sponsor to a decision, and that is the facilitation.
Should AI generate ideas in a brainstorm?
Not at the start. Fifty generic ideas crowd out the team's own, and nobody owns the result. If at all, use it late to provoke, and label everything it generated.
How do I use AI in a workshop without losing the team?
Announce what the tool does before it does it, label every AI output, let it score only against criteria the team wrote, and keep the human scribe until you trust the capture.
