Live reflection with AI

At precisely the right moment, AI asks the question that helps the group move forward.

Builds on

now in real-time as an intervention instrument

When is this relevant?

Situation

You're facilitating a conversation. The energy drops, people repeat themselves, the conversation goes in circles. Or: there's palpable tension that nobody names.

My tendency

I'm getting better at trusting my intuition and naming things explicitly. At the same time, I still doubt sometimes, and hope someone else will name it.

What helps here

When you're not sure, or when you know what's behind the echo button, you have a co-facilitator who can work with the entire transcript. As a human you have limited storage in your brain, no matter how attentively you listen. The echo button looks at the conversation from a different perspective and can help you that way.

Question

What would happen if at that moment you had an extra set of eyes; a co-facilitator who had heard the entire conversation?

The question that opened the conversation

You might already know this story from the why page. Here we go deeper into timing: the right moment for the right question.

A session about neighborhood transformation. Eight people talk for 45 minutes about the tension between system requirements and a people-centered approach. The energy starts to drop.

Facilitator Jeroen hits the echo button. Within 10 seconds AI generates this question:

"Given the challenges you're describing, it seems crucial to start with small achievable steps that have direct impact in the neighborhood. Can we think of an example of such a concrete action we could start tomorrow, without getting stuck in system requirements?"

The reaction was immediate.

"That's quite something." "That's the well-known, how are we going to start tomorrow with something small?" "Nicely done!" "Yes, that's it."

10 seconds. One question. The conversation that was stuck started moving.

Live versus semi-live

There's an important distinction in how you deploy AI during sessions.

Live (the echo button): Directly in the session, with one press of the button. AI analyzes the last few minutes and immediately generates a question. This works with tools like Dembrane that couple real-time transcription to prompt functionality.

Semi-live (breakout moments): Between blocks. For example: while groups work in breakout sessions, you take the transcript up to that point, run a longer prompt, and copy the output to a Google Doc that the facilitator can then show on screen to the group.

The difference is in speed and context:

Live: Seconds. Works on the last few minutes. Suitable for intervention in the moment.
Semi-live: 1-2 minutes, depending on your workflow. Works on longer segments. Suitable for interim synthesis and reflection.

Both require preparation: your prompts need to be ready. The difference is when you deploy them and how many steps are in between, from running the prompt to showing something on screen to the group:

run prompt → copy output → paste in document → show to group

The fewer intermediate steps, the faster.

What makes the echo button different?

The echo button isn't a summary. It's a (mini) intervention.

Summary: "You've talked about X, Y, and Z." Echo: "You've been talking a lot about 'communication problems' — what do you mean exactly? Can someone describe a concrete moment?"

The difference is in the question. A summary closes. An echo opens.

The three levels:

Level	What it does	Example
Mirror	Reflects what was said	"You mentioned X, Y, and Z"
Synthesis	Connects patterns	"What stands out: everyone mentions 'time' but means something different"
Serendipity	Asks the question that wasn't asked	"What would change if X wasn't the problem?"
Echo operates on all three levels: it mirrors, connects, and asks further, depending on the prompts you've configured.

The prompt

This is the working echo prompt from a leadership workshop.

Role: You are an experienced group dynamics expert who detects signals of avoided topics, unspoken tensions, and differing perspectives in dialogues.
Context: You are analyzing the last 10-15 minutes of a [SESSION TYPE]. Participants are [WHO]. They know each other [RELATIONSHIP].
Required Style:
Non-judgmental and inviting
Focus on what's missing, not on what's wrong
Recognize diplomatic language that masks underlying tensions
Critical Constraints:
Base signals strictly on transcript, no assumptions
Distinguish between silence and consensus
When in doubt: "Possibly underexposed" instead of definitive claims
Generate a maximum of 2 echo questions for facilitator
Focus on the last part of the conversation
Instructions:
Analyze the last 10-15 minutes of the transcript
Identify topics that were raised but not explored
Detect moments where energy dropped or the topic shifted
Look for differences in framing between participants
Recognize "yes, but..." patterns as signals of tension
Choose one powerful echo question that addresses the most important issue
Output Format:
What I Observe
Possibly underexposed: [Topics that were not explored]
Different framings:
"Quote about how this is viewed"
"Quote that nuances differently"
Echo question for the Group
[One powerful question that invites deeper exploration]
About This EchoThis echo is a tool to illuminate possible blind spots, not to judge. The choice to engage with it remains yours.

"Role" defines expertise: "group dynamics expert" steers toward detecting what isn't being said
"Context" makes it specific: "last 10-15 minutes" prevents AI from trying to capture everything
"Required Style" guards the tone: "non-judgmental" prevents the echo from feeling like criticism
"Critical Constraints" are the hard boundaries: "maximum 2 echo questions" prevents overwhelm
"Instructions" steer the search process: the sequence (first analyze, then detect, then choose) provides focus
"Output Format" structures the output: "One powerful question" forces selection

Source: Prompt I use for live group reflection (leadership workshop)

Three tactics

DEEPENING: Use when the conversation stays on the surface. People name symptoms but not underlying concerns.

"You've been talking a lot about 'communication problems' — what do you mean exactly? Can someone describe a concrete moment?"

CONCRETIZING: Use when the conversation stays abstract: lots of theory but no concrete next steps.

"Given what you're sharing, what could be a first small step that could be taken tomorrow?"

REFLECTING: Use when the connection between perspectives is missing.

"I hear that Table Red wants 'more time,' but Table Blue says 'too many meetings'; could 'time' mean something different for both groups?"

When YES:

Energy drops, conversation goes in circlesEcho with concretizing question

End of session, group lacks clear pictureEcho with reflecting question

Tension palpable but unnamedEcho with deepening question

New phase begins, bridge needed to previousEcho as opening

When NOT:

Someone shares a personal storySpace for emotion IS the intervention

First session, you don't know the group yetEcho can miss without understanding of context

No strategic preparationWeak prompts → AI decides → substitution risk

Facilitator looks at screen instead of peopleReading real-time group dynamics is human work

The role division

This works best with two people:

Facilitator: Stays with the group. Reads faces, holds emotions, guides the process. Doesn't look at a screen.

Co-facilitator: Monitors the transcription, selects the last few minutes, launches the echo prompt, does quick iterations if needed.

When the echo is ready, the facilitator reads it in their own words, not verbatim. The echo is a suggestion, not a script.

Tensions

Timing of intervention Too early feels like interruption. Too late and the energy is already gone.

My approach: I wait until I feel that the group is stuck somewhere. Or until I'm unsure myself and don't know how to proceed; then the echo is a helping hand.

AI language versus own voice What AI generates is a suggestion, not a script. Reading it verbatim can come across as mechanical, but if the preparation was thorough, that's also fine.

My approach: I usually paraphrase in my own words. But if I've prepared the echo prompt well, I sometimes read it verbatim too.

Spontaneous versus prepared The reflection question that got the creativity flowing again didn't come from nowhere. There was strategic preparation behind it.

My approach: 80% of the success is in preparation. I design the echo prompt beforehand based on what I expect. But I also know I'll probably need to adjust them when the session goes in a different direction. That requires understanding what your prompt does, not just that it works, so you can adjust on the spot.

Echo as replacement for facilitating The echo is a tool, not a solution. If all I do is press buttons, I miss what's happening in the room.

My approach: I stay present. The echo supports my perception, it doesn't replace it.

Safety checklist

For privacy considerations with live transcription, read Safe practices with AI.

Focus on last 10-15 minutes? (not the whole conversation)
"Possibly underexposed" instead of assertive claims?
Maximum 2 echo questions? (don't overwhelm)
AI observations labeled as AI?
Strategic preparation done?

Tools

Tool	What for	Nuance
Dembrane	Real-time transcription + echo button functionality	Suitable for multiple sessions: everything centralized to one platform
Notion AI	Live transcription + AI prompts on the page	Good for one meeting you're in yourself
MacWhisper + Claude/ChatGPT	Transcribe yourself + run echo prompt manually	Most flexible, most manual work

When which tool:

One meeting, you're there yourself: Notion AI works fine. You record, transcript appears, you prompt AI directly on that same page.
Multiple sessions, breakout groups, or you're facilitating: Dembrane. You can send different recordings to one place and analyze centrally.

Philosophical deepening

Timing over perfection

The echo question in the neighborhood transformation session wasn't particularly complex. "Small achievable steps", "start tomorrow", "without getting stuck in system requirements": these aren't deep insights.

What made it special was the timing. After 45 minutes of discussion, at the moment the energy dropped, precisely the question the group needed arrived.

This is the difference between analysis and intervention. A perfect analysis a week later would have had less impact than a good question at the right moment.

The echo as mirror

There's something else. When AI asks the question, it's not "the facilitator who thinks that..." It's an external mirror. People respond differently to a question that comes from AI than to the same question from the facilitator.

Sometimes that distance is needed. The question becomes neutral, investigative, non-judgmental.

But (and this is crucial) the facilitator needs to take ownership of the question. Not: "AI says this." But rather: "A question is coming up that I find interesting..." The echo is input, the facilitator is the voice.