FEATURED ARTICLE
dembrane is a set of building blocks we built to help support dialogue processes: pick the ones you need, leave the rest. At its base it's a record-keeper: record the whole process without changing how you facilitate it, and you already have an audit trail and a memory. Everything above that is an optional feedback loop running on a tighter, faster cycle than the one below. The last loop is opt-in, and it runs the other way and pools what's learned across processes.
How to use this map
Walk your process design through the five layers and mark if a layer earns its place, then switch on only the features you need. Just the recording layer on its own is already worth it; the tighter loops are there when a moment calls for them; the commons is there if you ever choose to share.
The five layers
Layer
What it does
How participants relate to it
What it takes to switch on
1: RECORD
The whole process
Turns any number of phones into a secure system of record: every conversation transcribed, searchable, auditable.
Most people don't mind being recorded, as long as it's clearly optional, consented to and privacy-safe.
Facilitation stays the same, but there's some setup: create a project, set anonymisation and keywords. Then get informed consent, frame recording as optional and scan a QR code per recording.
2: ASK
A session
Merges and de-duplicates what was said across every table into a first draft of the recommendations, and surfaces what a session missed.
People are usually glad to see hundreds of hours pulled into something they can hold, as long as it is presented in their words, and reflects their conclusions, not the machine's verdict.
Add a synthesis step: generate the merged draft, then bring it back to participants to validate.
3: VERIFY
A conversation
Hands the summary back to the group, in the room, so they correct it and sign off on what gets saved.
Being asked 'did we get you right?' tends to land well. People will correct the record if you give them the moment. It can also feel like too much after a long session.
Design in a live review moment at each table, with time to read the summary back and correct it. Usually 5 minutes is plenty. Practice makes perfect.
4: EXPLORE
Within a live conversation
Offers a Socratic question, drawn from the friction in what's been said, to deepen the conversation as it happens.
A sharp question can feel like a gift that opens things up; a clumsy or ill-timed one feels like a gimmick. It only lands when the conversation is already alive.
Brief people on the button and decide where a live question might earn its place. It's the most experimental layer, and works well in settings without a facilitator.
0: SHARE
With the commons (coming soon)
Pools anonymised processes into a governed data commons so the whole field can learn what makes deliberation work.
In many contexts, people are honoured at the idea that their words will make an impact beyond the local context. In others, it's an unacceptable risk. Check in with the stakeholders.
Opt in, add a separate data-sharing agreement and inform participants.
The higher you climb through the layers, the tighter and faster the loop, from a look back over a whole process to a question inside a conversation that's still happening. SHARE (0) runs the other way: the widest, slowest loop, with the biggest meta impact.
The stories behind the layers
Prologue
Deliberation is the process of having frank, candid, often difficult conversations in order to arrive at outcomes that matter and actionable decisions that a maximum number of stakeholders can consent to.
Most people have experienced some kind of deliberation. It is often the default mode of collective decision making when there is no implicit hierarchy. Power delegates, peers deliberate.
Power delegates, peers deliberate.
At dembrane we believe that deliberation is the default mode of human coordination and decision making, but as societies have gotten larger and more complex, we have needed to come up with ways of scaling deliberation. We tried just letting the village elders deliberate, or electing representatives. These techniques have gotten us pretty far, but I'm sure we can all see that we are at the point where we should be having really big, candid conversations about how to get on, but instead we are stuck in old routines of power plays and back room gossip. Deliberation has stalled and we need new ways to bring back its mojo.
Luckily there are some excellent people in this world who have dedicated their careers to helping us all return to good old fashioned, bottom up deliberation. A few years ago, the founders of dembrane decided to try and help these people as much as they could, and dembrane was the result.
People have a diverse range of opinions about what constitutes good deliberation. We knew that if we built something that was overly opinionated, that we wouldn't make anyone happy. Instead, we were inspired by Lego: If we could provide building blocks where process designers, facilitators and deliberation hosts could pick and choose the blocks that they needed, we could create something that could help the widest range of processes.
Now that we have built a set of building blocks, we've noticed that they enable different layers of value. From basic record-keeping and data hygiene, to the most experimental features, process designers can pick and choose what they need to supercharge their processes, without being locked into a single way of working.
Layer 1: RECORD
Record keeping. Looking back on a whole process.
The first layer of value is as a system of record. A feedback loop to look back on a whole deliberative process.
Many deliberation processes we observed were using handwritten notes, post-it notes, whiteboards and flip-overs to record some of what the deliberators were saying. This is fine, but we heard a lot of people asking "Where are the post-it notes from the previous session?" or "Oh! I remember somebody had a great quote about that but I can't remember what it was..." or "Can you read the handwriting on this post-it note?"
In many deliberative processes, there simply wasn't a system of record. And when auditors and commissioners came along and asked how they could be sure that the outcomes and recommendations came from the participants, the process designers drew a blank. Fetching a stack of posters from the back of a car as post-it notes flutter away in the wind doesn't exactly inspire confidence.
What else could they do? They could try to transcribe verbatim what everyone was saying, but that's a dedicated job. They'd need to have note takers at every table and they are already stretched.
They could buy a bunch of microphones and record every conversation, but then they would have to keep a stack of microphones in the car, and someone would have to worry about keeping them charged, emptying the SD cards, plus where does the data live? How can you store it safely? What if there's more tables than microphones?
You could ask people to record the conversations on their phones, but then you've got to get everyone to transfer the files to you. Have you ever tried to upload an hour of audio from a stack of phones to some kind of cloud storage? It's not exactly easy. What if they keep the audio on their device? That's a data leak.
What about consent forms? Transcription?
We'd seen every variation under the sun of people trying to establish some kind of system of record for what was said. None of them were convenient. The default was the bare minimum. Not ideal.
We built dembrane so that people can turn any number of phones into conversation record keepers just by scanning a QR code. The conversation doesn't get stored on device, it gets sent immediately to our secure cloud environment where it is transcribed with state of the art models and saved on encrypted servers where process designers, hosts, and auditors can access it and trace what was said whenever they need to. Now process designers can decide to keep the audio or anonymise the conversations and only keep redacted transcripts. Consent forms and data retention are all managed centrally and policies are enforceable.
Nothing about the deliberation process itself had to change. Just one QR code scan per conversation, the phone just sits there on the side. At this point, we have helped hundreds of processes go from handwritten notes to a fully compliant system of record.
That's layer 1.
Layer 2: ASK
What were we thinking? What have we missed? Looking back on a session.
If you are recording all the conversations in a deliberation, a natural question becomes: what can you do with all that information?
Merging and de-duplicating findings across conversations
In the early days of dembrane back in 2023, we had a pilot client who was conducting a deliberation with about 90 stakeholders in partnership with the stakeholder engagement and participation team from a major bank. They had facilitators and note takers at every table, and after the event, they would compare notes to draft a report which they could hand back to the commissioner. At that time, we were still recording conversations with a suitcase full of microphones, and Jorim would run around the room switching out SD cards and transcribing all the conversations on his laptop.
Using a data processing pipeline powered by GPT-4, we had finished a first draft report before the final plenary.
The room read the recommendations and the supporting quotes, and we heard little gasps of "Hey I said that!" and "Did your group discuss that as well?!" The client was happy, and a few weeks later we checked in with the bank to see what progress they had made on their report. The team had decided not to write it, as the note-takers and facilitators all agreed the dembrane outcomes were sufficient.
"Hey! I said that!"
It's not impossible to draft collective outcomes without dembrane, but every process we had encountered faced a similar bottleneck. At a single table, you can get participants to formulate recommendations deliberatively, but how do you reconcile all the different recommendations at all the different tables?
Often, results are compiled after the events, which kills momentum and participant ownership. Other times, facilitators meet in between sessions to re-structure the recommendations to arrive at a final list. Other times still, a participant per table is asked to share their group's outcomes with the whole group (to varying degrees of success).
With dembrane, it becomes possible to ask for a de-duplicated list of outcomes based on the transcripts from every table. Not replacing the process of drafting, but merging the drafts from multiple tables simultaneously into a big picture. This machine generated big picture needs validation, of course, but the participants are still in the room: once they double check and sign off, the process is once again empirically rooted in their experience.
Since that first report in 2023, dembrane has been used to generate preliminary outcomes in hundreds of processes. We have learned a lot!
The point is not to feed a language model a bunch of transcripts and ask any kind of question and present the answer as some kind of oracle, that's how participants lose trust in the process. The point is to do the best possible deliberative process that you can, and ask dembrane for the final step of merging and de-duplicating the findings.
Looking for blind-spots
Another valuable pattern is using dembrane to ask: what have we missed? After a deliberation on mobility at the municipality of 's-Hertogenbosch in the Netherlands, we asked dembrane what the municipality had missed. dembrane responded that the discussions often mentioned the needs of delivery drivers and emergency services, but that those people were not actually present: they were being used as a rhetorical device rather than being represented. The process designers: "Oh, wow, that's so true, we need to understand things from their perspective also."
As long as the output from dembrane is treated with the scepticism it deserves, as a first draft subject to maximum scrutiny, participants start to feel more ownership over the outcome and trust in the process, not less. The added benefit of speed means processes can hold onto an underrated resource: momentum.
Layer 3: VERIFY
Missed a spot: looking back on a conversation.
As we built out dembrane and it got easier to record conversations and draft collective recommendations, we realised how important feedback loops were to the entire process of deliberation.
In many sessions, we noticed conversations could get very long, addressing many issues simultaneously and weaving in and out of big picture thinking, and nitty gritty details. A skilled facilitator could give a comprehensive overview of the discussion after the fact, or be fully present and emotionally attuned to the dynamics, but doing both was incredibly straining.
Just dropping such a transcript into a system of record, or using it to generate recommendations, sometimes flattens the very thing that made the conversation worth having. The moment where someone connects two things nobody else had connected gets smoothed into a tidy bullet point, and the connection, the whole reason the conversation mattered, is lost.
So we built a feedback loop that closes while the conversation is still warm. We call it Verify. After people talk at a table, dembrane drafts a summary and hands it straight back to the group, and they push on it. Not in a report weeks later that nobody reads, but right there in the room, while they still remember exactly what they meant.
In a peacekeeping assembly in Northern Ireland, a participant read the summary their table had been given and said:
"Addiction links to trauma. Trauma links to the Troubles. The Troubles links to sectarianism, racism, homophobia, violence against women. I didn't see that in the AI summary."
dembrane had caught the topics, but it had missed the thread running through them. To the participant, that thread was the entire point. Because Verify put the summary back in front of the group while the conversation was still happening, the critique was captured, and it let the hosts pivot in real time: away from the traditional theme tables on day two, the "Mobility" table and the "Food" table, toward the higher question the participants were actually reaching for, like how the work of prioritising resources should itself be redesigned in an interconnected system.
The reason everyone could nod at the day-two results was that the system had been corrected. Verify keeps the participants as the source of truth and ensures the participants retain authority over their own words. dembrane asks "did I get this right?"; the people at the table decide whether to push back or not. The summary that gets saved is one they corrected and stand behind.
Layer 4: EXPLORE
A sharp question. Looking forward, inside the conversation.
Every layer so far looks backward. RECORD holds the whole process; ASK and VERIFY reach back over a session or a single conversation. EXPLORE it is the only one that looks forward: a single question, offered into a conversation while it is happening.
While people are talking, dembrane listens for the friction in what they have been circling and, when prompted, offers one Socratic question to take the conversation a level deeper.
At a corporate conference in 2025 we tried it with a room of board directors and young talents. They paired up, walked, recorded their conversations, and could press explore whenever they liked. Left alone, a lot of these conversations did what conversations between accomplished people often do: they stayed up high, fluent in the language of stakeholders and systems and good governance, a little abstract, a little resigned.
Then a question would land, and you could watch the altitude drop in real time.
One pair had been talking, smoothly, about "breaking out of the bubble and anchoring societal breadth into the governance DNA." Explore asked them what only an impersonal agent without a sense of the power dynamics could ask: what it would actually take to break their own bubble before they redesigned anyone else's. One of them stopped performing: she said she feels caught in her bubble all the time, that her own kids laugh her out of the room when she says she is for diversity, that she lives in a comfortable part of Amsterdam and her whole circle of friends looks exactly like her.
It did the same thing again and again. A lofty question about putting the broader interest ahead of short-term gain got poked by Explore, and the conversation turned concrete: would they be willing to not go on holiday three times a year, to not remodel the kitchen but invest in a vegetable garden instead? Another question asked, directly, whether anyone present had ever been genuinely vulnerable inside a formal leadership role: someone took the dare ("I've got one, shall I start?") and told a real story.
dembrane can't replace the perspectives of the people in the room (by design). Give one conversation to dembrane a hundred times and you'll get roughly the same question back; give it to a hundred different people and you'll get a hundred different responses. The machine is good at finding the next layer of a conversation that is already rich. It can't seat a missing person at the table, and it won't think a thought the group doesn't already half-hold. It's up to process designers to shape the day so that dembrane sharpens what the people have to say instead of vice versa.
Layer 0: SHARE
A commons for deliberation. Looking across every process, over years. (An invitation)
There is one more loop, and it runs the opposite way from all the others: not tighter and faster, but wider and slower. It sits underneath everything, call it layer zero, and we haven't built it yet.
Every process that records itself with dembrane builds a small, careful archive of how one real deliberation actually went. Imagine those archives, opt-in and anonymised, pooled into a commons so that the field of deliberation can learn about itself.
To compare how deliberation works in one country and another.
To see how a method travels.
To let researchers build and test real measures of what makes a process good.
Verify hints at why this could work. Every time a participant corrects a summary, we are left with two things side by side: what the machine first understood, and what a human said it got wrong. That is close to exactly the material you would need to make deliberation tools that genuinely improve over time with real deliberation as the source of truth.
This is the longest game, and the impacts your process the least. It does nothing in the room, and it only ever happens if a process chooses to opt in. We are building toward it, and looking for partners who would be glad to share what they learn. One question we will leave for you:
if your process could help every process that comes after it, what would you be willing to share?
How will you use it?
Those are the building blocks so far!
RECORD, so a process has a memory and can show where its conclusions came from.
ASK, to draft and redraft and surface what a session missed.
VERIFY, to let participants sharpen the summary while it is still warm.
EXPLORE, to deepen a conversation as it happens.
And, one day, SHARE: to let your process help the field evolve.
None of them are a way out of the hard part. The deliberation stays the people's to do; every layer is defined as much by what it leaves alone as by what it adds. See it as opting out of convenience by design. "Design for friction" maybe.
But we are curious: in the process you are building, where would a system of record change what's possible? Where would a second draft, a live correction, or a sharper question earn its place? Even just recording the conversations can be hugely valuable, and if you're ready and the context allows it, maybe you can ask, or explore, or verify. Take what helps and leave the rest. Then tell us where it helped and what needs improving.
