AI facilitation: don’t automate the facilitator

# The pitch is an AI agent with a marker

Strip the demo down and the AI facilitator does four things: it sets an agenda, states the norms, prompts whoever is next, and writes the recap. Three of those are documents. The fourth is a timer.

The part that looks like facilitation, deciding the room is ready to move on, is a judgment made from material that never reaches the transcript: who has gone quiet in the last ten minutes, which agreement sounded rehearsed, whether the two people who actually have to execute this said the same thing or just used the same word. Authority to say “let’s move” is granted by the room, and rooms grant it to people.

You can automate a timer. Nobody needed that automated.

# Facilitators adopted AI everywhere except the room

The State of Facilitation survey ran across 1,050 practitioners in late 2024, and its picture of AI use is lopsided in a specific direction. Everything before the workshop and everything after it, heavily. The workshop itself, barely.

One respondent summed the posture up neatly: “I use it for creating first drafts, not final agendas.” Notice where summarizing sits in that list, too. 58.6% use AI to summarize information, which sounds like the exact capability a facilitator needs mid-session, except that in practice it happens on the train home.

Preparing sessions

85.8% of AI-using facilitators

Wrapping up and post-workshop activities

42.2%

Marketing

29.5%

Facilitating sessions

19.6%

Sentiment about AI in facilitation

62% positive, 30% neutral, 8% negative

SessionLab, State of Facilitation 2025 (third edition, 1,050 respondents) ↗

# The objection is about sensemaking, and it is a good one

Ask a working facilitator whether AI should synthesize the room and you will not get a complaint about AI agents. You will get a claim about where meaning comes from, and it goes roughly like this: the struggle to reach a synthesis is the thing that makes it hold. Sam Kaner’s groan zone earns its name because the group has to grind its own conflicting framings into something shared, and a tidy summary handed down from outside gives you alignment-shaped output without any of the alignment. We went through that whole diamond in how AI changes convergence, so take it as read here.

There are three sharper versions of the objection underneath it. A facilitator reads hesitation and tone, not just words. A human synthesis can be argued with in the room, while “the AI said” quietly launders a judgment nobody has to defend. And blending averages, so the one person who saw the real problem gets smoothed into the middle of the paragraph.

All of that is correct, and any piece that waves it off is not worth reading. So the argument has to be made on the objection’s own ground.

While AI can fast-track a group’s decision-making, it cannot drive participants’ buy-in if they haven’t been involved in the solution-finding and sense-making process. Just because AI can do it, doesn’t mean it should.

Dr Myriam Hadnes, expert insight in State of Facilitation 2025

# Your affinity map was always one person’s reading

Here is the part of the ritual nobody defends out loud. Sixty sticky notes come off the wall at the break. One tired person, working alone against a twelve-minute coffee queue, sorts them into five clusters and writes a label on each. Those five labels go back up on the wall, and for the rest of the session the room treats them as what it said.

That clustering is sensemaking, performed by a single human under time pressure, with no version history and no second opinion. Nielsen Norman Group’s guidance on affinity diagramming (Brown, 2023) catalogs how it goes wrong: keyword matching, where notes get grouped because they share a word rather than a meaning, and dominant voices setting the grouping strategy early, so that better ways of cutting the same data never get the space to emerge.

Facilitators only ever get one clustering for a boring reason. Re-sorting a physical wall a second way costs twenty minutes you do not have, so you pick a reading and live with it. Stasser and Titus showed in 1985 that this is exactly the moment groups are most vulnerable: discussion systematically favors information members already shared and information that supports what they walked in believing. The single reading you can afford to make is the one most likely to be the obvious one.

Which means the status quo already hands sensemaking to one person, working invisibly and getting a single attempt at it, while the room receives the result as though it were the data.

Nielsen Norman Group: Avoiding 3 Common Pitfalls of Affinity Diagramming ↗

# The second reading: same answers, new criteria

The move we want a name for is this: leave the question alone, leave the answers alone, and change the instruction that governs how they are read. Then read them again. Call it the second reading, because in a normal session there has never been one.

In One Voicer the instruction is a field called response criteria, one or two sentences long. Editing it arms a re-blend, and the blend re-forms over every response already collected, usually inside about fifteen seconds. Nothing is filtered and nothing is deleted; criteria steers how the voices are woven together, and every voice is still there and still counted.

That last property is what makes this different from asking a follow-up question. A new question asks the room for another round of typing, and the material it produces then needs a synthesis of its own. A second reading asks the room for nothing. The fifty answers on the board stay exactly as they were, and you can change the lens over them as many times as the session has room for.

# Six readings of the same fifty answers

The two worth reaching for first are the two a wall of stickies is worst at. Dissent is the obvious one: a default blend smooths a room toward its average, and typing one sentence that asks for the fault line instead will often produce the most useful fifteen seconds in the middle of a session, because it turns a vague sense that people disagree into a named disagreement the group can work.

The second is the antidote to the Stasser and Titus problem. Ask the blend for what only one or two people mentioned, and you get the unshared knowledge that ordinary group discussion is structurally bad at surfacing, handed to you before the room has had a chance to bury it under the obvious.

Ranking by frequency is last because dot-voting could already give you that one.

# Show the room your lens

This is where the laundering objection gets its answer, and it turns on a detail that is easy to miss: response criteria is shown to respondents as they answer. The instruction that shapes the synthesis sits on screen next to the box they are typing into.

Compare that to the break. Nobody in the room can see the sorting rule in the facilitator’s head, ask what a different rule would have produced, or notice that “Button Issues” got its label from a shared keyword. Put the rule in writing, show it to everyone, and change it in front of them, and the interpretive layer becomes the most auditable part of the session rather than the least.

One honest wrinkle. If you change the criteria mid-session, anyone who answered earlier saw the previous wording, so say out loud that you are re-reading their answers under a new instruction. One sentence covers it.

# Automate the reading, not the deciding

Three things stay with the facilitator, and no criteria field touches any of them: deciding when the room is ready to move, deciding what the room’s disagreement actually means, and holding the argument that follows. A second reading does not resolve a conflict. It puts the conflict on the projector in words the room recognizes as its own, which is more work for the facilitator, not less.

That is the reason to be unimpressed by the AI agent with the marker. It automates the chair and leaves the reading alone, when the reading was the only part that a human genuinely could not do fast enough. Response criteria come with the Creator plan and up, alongside the group and present features a facilitator runs a session on.

The profession’s own summary of where this is going, printed on the cover of the 2025 report, is “human-led, tech-enhanced.” Fair enough. Just be precise about which half the machine is doing.

Asking your question: response criteria, length, and context →