I did not arrive with a plan for how AI should be used in that process. What happened was much more organic. As the day unfolded, I could see places where my experience with AI might help us process and synthesise a large amount of material quickly. I volunteered
that capacity, and one of the facilitators, with whom I already had a good working relationship and who also understands technology well, was comfortable experimenting with me.
So this was not an “AI-enabled workshop” designed in advance. It was a participatory workshop in which we progressively found ways to use AI as a support tool. This post describes what we tried, what I think worked, where the risks became more visible, and what I would design more deliberately if I were involved in a similar process again.
Oh and... When I refer to AI in this post, I mean a small set of AI-enabled tools we used for transcription, synthesis, analysis and re-presentation of participant-generated material.
What were we trying to do?
The Commission had a fairly ambitious task. We wanted participants to move from a broad discussion of challenges in teacher professional development towards a shared diagnosis of the underlying problems, possible responses, priorities and, eventually, a Commission declaration.
The workshop was designed as a sequence. We started with expert inputs and discussion to establish some common ground. Participants then worked in groups to identify challenges and think about their immediate and deeper causes. We brought those inputs together into a common root-cause structure, used a World Café-style process to interrogate and develop the issues further, asked people to prioritise them through dot voting, and then moved towards possible actions and commitments.
That sounds reasonably tidy when written afterwards. In practice, the room generated a lot of information very quickly. That was the first reason AI became useful.
The first use of AI happened in the groups
During the root-cause exercise, groups discussed the challenges they thought mattered and mapped the causes underneath them. Some worked on newsprint. Some used their own AI tools to turn their discussion into a neater diagram. Others simply photographed what they had produced.
I liked this use of AI because the people who had generated the ideas could immediately check the output. If the diagram did not represent what they had meant, they were right there to correct it.
It also allowed groups to surface quite a lot of material. In a conventional workshop, one spokesperson may spend several minutes explaining one or two points while everybody else listens. Here, groups could retain a much richer set of ideas without needing to report every point verbally to the whole room.
By the end of the first round, we had several different group products. But I was also conscious that a photograph of newsprint or a neat diagram does not tell you everything that happened in the conversation. So we asked group representatives to give us a short narrated explanation of their work as well.
We used the narratives together with the photographs and diagrams. That helped us retain some of the language, emphasis and context behind what had been written down. For me, the combination of the group product plus a short human explanation of it was much more useful than either on its own.
I gradually became part of the synthesis process
As the group products accumulated, I started helping the facilitation team bring the different inputs together. This was not a role that had been planned for me in advance. It emerged because the combination of the workshop need and the tools I was able to use made it possible to synthesise material much faster than we otherwise could have.
I used AI to help integrate the group products into a common fishbone structure. We developed several versions, gradually incorporating the original group work, narrated explanations and further discussion in the room.
This worked partly because there was already trust between me and the facilitation team. I was not working separately from the process and handing over an AI-generated analysis from the outside. We were talking continuously about what was emerging, what looked right, what needed checking and what should go back to participants.
By the end, the fishbone contained six broad areas: system coordination and alignment; quality, relevance and recognition of teacher preparation and TPD; teacher workforce and pathways; support around teachers and practitioners; data, research, evidence and learning; and resources and enabling conditions.
This stage was where AI was particularly powerful. It helped us move quickly between messy material and something participants could actually look at together. We could identify overlaps, move ideas, merge some things, separate others and return a revised version to the group while the discussion was still fresh.
But it also made me think quite hard about power.
AI did not remove the power of the person doing the synthesis
Once I became the person through whom much of the material was flowing, I acquired a degree of interpretive power that had not been explicitly designed into the process.
I decided what material to put into the tool, what questions to ask, which outputs looked sensible, what needed to be revised and when something was ready to go back to the group.
In other words, I was still holding the pen.
And because AI allowed me to process so much information so quickly, I was probably holding a more powerful pen than I would have in a normal workshop.
That made checking especially important.
The importance of being able to check
One of the things I think we got right was that the AI-produced synthesis did not simply become the final answer.
We kept returning the fishbone to the participants. People could challenge the wording, add things that were missing, move ideas around and question the categories. There were several iterations.
I also deliberately kept the later detailed versions together instead of simply replacing the old one with the new one. That meant the people doing the final synthesis could go back and compare versions if they wanted to check whether something important had disappeared or changed.
The Commission declaration eventually described the root causes as being identified through analysis, validated through stakeholder engagement, supported by AI and prioritised through voting.
I like that description because it reflects the balance reasonably well. AI helped us organise the material, but people still had to decide whether the representation was credible.
This also exposed a risk I had not thought about quite so clearly before: people are less likely to interrogate something once it looks neat.
A wall full of newsprint looks unfinished. Everyone understands that it is provisional. A nicely structured framework with clean headings already looks like analysis. It carries a kind of authority simply because it has been organised well.
AI is extremely good at producing that effect.
So next time I would be even more deliberate about asking people to challenge the synthesis. Not only, “Does this look right?”, but questions like: What has disappeared? What has been merged that should not have been? Whose perspective is missing? What looks more settled here than it actually felt in the room?
The neater the AI output, the more deliberately we probably need to invite people to disagree with it.
The dot voting reminded me not to automate everything
Once the fishbone was sufficiently stable, people prioritised the issues by putting dots against the things they thought mattered most.
We could have done that through an online survey. It would have been quicker to capture and analyse.
I am glad we did not.
People stood around the diagrams together. They talked about what the categories meant. They puzzled over where they wanted to place their dots. They could see what other people were responding to. The voting produced data, but the process of voting also created another round of collective sense-making.
We then captured the voting data and checked it manually. The strongest signals included fragmentation and alignment, data and evidence, programme quality and the support structures around teachers.
That was a useful reminder that a workshop activity can have two functions. It may collect information, but it may also create interaction. If we automate the first function, we can accidentally remove the second.
Not every inefficiency should be removed.
The point where I think we relied too much on AI
After the dot voting, we had very little time to analyse what the pattern meant.
AI made it possible to do quite a sophisticated analysis very quickly. We could look at which individual issues received the most votes, which broader categories attracted attention, and which themes seemed to cut across several parts of the fishbone.
I ran a few additional queries to check some of the claims. We manually checked the voting numbers. I also sense-checked some of the interpretations with people in the room.
But this stage was harder to verify properly because of the time pressure.
Looking back, I think we handed too much analytical authority to AI at the point where our capacity to check the output was weakest.
The analysis was plausible. Much of it made sense. But plausible and correct are not the same thing.
Fortunately, this was not the final product of the day.
The final synthesis was much more human
There was a team responsible for turning the material into the Commission declaration. They had been in the room. They had listened to the discussions. They knew the policy and institutional context. They could see where the AI-assisted analysis fitted what had happened and where it needed to be adjusted.
That also meant that my own interpretation did not simply carry through to the final product. Other people could temper it with what they had heard and understood.
The team then had to translate the emerging analysis into a somewhat different format: proposed actions, lead stakeholders, supporting stakeholders, timelines, first steps and indicators.
That required much more than summarising what people had said. It required judgement.
And it highlighted another limitation of the workshop design.
Workshops are better at diagnosing problems than designing solutions
I think the process worked quite well for diagnosis. People could identify problems, challenge one another and gradually build a richer picture of the system.
Developing workable solutions was harder.
That is partly because people in these rooms work inside real institutions. They cannot necessarily volunteer their organisation to do something without checking with their leadership. They may not want to tell another organisation what it should do. They may know that an idea crosses somebody else's mandate or depends on a budget or policy process that is not in the room.
There is also a more basic point: workable solutions often need people with quite specific knowledge. It is one thing to agree that coordination is fragmented. It is another to know exactly what coordination mechanism could work, who has the authority to convene it, how it should relate to existing structures, and what would make it sustainable.
We did try to create another route for ideas by opening a short anonymous survey. I still think that was useful, because it allowed people to suggest things without publicly volunteering themselves or somebody else.
But we introduced it too late. By that time, people were tired.
Next time I would open that channel much earlier and let people add ideas throughout the day.
AI saved time, but it did not save mental effort
The process looked fast from the outside. That was partly because a lot of the complexity was being absorbed behind the scenes.
I was listening to the room, deciding what needed to be captured, receiving group products, looking at photographs, processing narrated recordings, transcribing them, prompting AI, checking outputs, revising structures, checking voting data and sharing products back into the process.
By the end I was mentally exhausted.
AI had reduced the time needed to produce things. It had not removed the need for judgement. In some ways, it increased the amount of judgement required because we were able to process much more information and produce another iteration very quickly.
The bottleneck moved from producing the analysis to checking and thinking about the analysis.
There were other people in the room who were also very comfortable with these tools and could probably have helped. But because we were making the workflow up as we went along, I did not have a good way of delegating while keeping everything connected.
For example, someone could take responsibility for the recordings. But later I realised that what I really needed was not simply a transcript. I needed the narrative linked to the correct group product and available in a form that could easily feed into the synthesis.
A better workflow would have helped enormously.
What I would do differently next time
I would use AI again. I think it added real value. But I would design the process around it much more deliberately.
I would probably build the workflow around a simple chain:
Raw group material → AI-assisted synthesis → human checking → participant validation → AI-assisted analysis → independent human review → final human synthesis.
I would assign different people to capture, synthesis and checking rather than allowing too much of that to sit with one person. I would also keep the original material and intermediate versions so that people can work backwards if something in the final synthesis does not look right.
Most importantly, I would think about checkability at each stage.
The question would not only be, “Can AI help us do this faster?” It would also be: Can we check the result? Who is able to check it? How much does this output influence what happens next?
The more consequential the output, and the harder it is to verify, the less comfortable I am allowing AI to carry the interpretation without another human layer.
We also needed basic guardrails
There were some things we did deliberately from the beginning.
We used paid tools with privacy and sharing settings clarified. We did not put confidential information or personally identifiable information onto the AI platform. The use of recordings and AI-assisted processing was disclosed, and participants actively agreed to this through the colloquium platform.
Those things are easy to treat as administration around the “interesting” AI work, but I think they are part of the method.
If we use AI in participatory processes, we need to think about consent, privacy, who controls the material, who gets to interpret it, and who can check what has been produced.
So, was it worth it?
I think it was.
AI helped us surface and retain more ideas than we would otherwise have managed in the available time. It helped us move between messy discussion and visible structure, and it allowed us to return synthesis to participants quickly enough for them to respond to it.
By the end of the process, the Commission had moved from a very broad collection of concerns to a fairly coherent diagnosis: the problem was not simply that South Africa needed more teacher professional development activity, but that there were weaknesses in alignment, evidence use, support, implementation and accountability across the system.
But AI did not make that conclusion legitimate.
The legitimacy came from people generating the material, checking how it had been represented, arguing about it, voting, revising it and finally deciding what they were prepared to carry into the declaration.
That is probably the main lesson I will take into the next workshop: AI can help a room process much more, much faster. But the design still has to make sure that people remain able to see, question and own what is being produced.
AI disclosure: I used AI as a thinking and writing partner in developing this reflection. It helped me organise the sequence of events, test the argument, and refine the wording. The account, interpretations and judgements are my own, based on my participation in the Commission process and the workshop materials. Charlie's joke for you: Knock knock. Who’s there? AI. AI who? AI thought I’d summarised that perfectly. Human: “Not quite. Try again.”
