Who does what, when the model is in the room
Most workshop agendas still don’t say where the AI goes. So it goes everywhere, and the facilitator ends up reacting to it all day.
What changed
Two years ago a workshop agenda was a list of exercises with times next to them. Lightning demos, 9:30. Sketching, 11:00. Everyone knew who did each one, because the answer was always “the people in the room”.
That’s no longer true. In every workshop I’ve run this year, somebody has opened a model during a break and come back with something. Twenty ideas. A rewritten problem statement. A clickable prototype. Sometimes it helps. Often it hijacks the next hour, because the room now has to react to an output nobody planned for.
The agenda hasn’t caught up. It still lists exercises. It needs to list jobs.
The skill
Take tomorrow’s agenda and, for every task on it, answer three questions. Who does this: a person, the model, or both? If the model is involved, who reviews what it produces? And is there a decision here that a human has to make before we move on?
Write the answers on the agenda. That’s it. It looks like admin. It changes the day.
Here’s a real one, from a two-day HyperSprint with an energy retailer in October. The challenge was reducing the calls that come in after a customer gets an unexpectedly high bill.
| Task | Person | AI | Review | Human decides |
|---|---|---|---|---|
| Map the current call journey | ● | |||
| Pull patterns from 200 call transcripts | ● | Ops lead checks 20 at random | ||
| Write the sprint question | ● | ● Decide | ||
| Generate 30 intervention ideas | ● | ● | People sketch first, then the model widens | |
| Pick three to prototype | ● | ● Decide | ||
| Build the three prototypes | ● | Team directs every revision | ||
| Draft the customer test script | ● | Researcher edits | ||
| Run the five customer tests | ● | |||
| Decide: build, change, combine, retest, stop | ● | ● Decide |
Fig. 1 — A HyperSprint agenda with jobs assigned. Three moments where a human has to decide before anything moves.
Three things to notice. The model never has a task without a named reviewer. The three decision points are where the day actually turns, and nothing is allowed to be generated past them until a person has chosen. And the people do far more than you’d expect: mapping, sketching, testing, deciding. The model does the volume work. That’s the split.
What goes wrong
The first time we tried this we assigned too much to the model. Transcript analysis, idea generation, prototyping, the test script and the summary. It was efficient and the team felt like spectators. By the afternoon of day one they’d stopped arguing with anything. We’d designed them out of their own workshop.
The fix was to move the “generate ideas” task back to people first, with the model second, and to make the review column a named person every time. The energy came back the moment someone had to put their name next to “checks 20 transcripts at random”.
The other failure is subtler. If you don’t mark the decision points, the model will quietly make them for you. A prototype that exists is a decision that’s been made. More on that on Day 6.
Run it Monday
- Print the agenda. Add four columns: Person, AI, Review, Human decides.
- Fill it in. Anything the model does gets a named reviewer.
- Mark no more than three or four decision points. If there are more, the workshop is too big.
- Share it with the room before you start, and say why the model sits where it does.
- Afterwards, note which cells you’d move. That’s your next agenda.
The agenda used to say what we’d do. Now it has to say who does it, and who checks.
The agenda template above, with every slide and timebox we run a HyperSprint from, is in the facilitation deck. It’s free.
