What leaves the room
The dangerous output of an AI workshop is a tidy summary. It looks like a decision. Nobody owns it, and nobody can tell which parts were checked.
What changed
Workshops used to end with a messy wall and a tired team, and the mess was useful. You could see what was decided and what was still an argument. Now they end with a document. The model produces it, it’s formatted, it has headings, and it reads as though the team reached a conclusion on every point.
That document gets forwarded. The sponsor reads it as a decision. Three weeks later someone asks “who agreed to this?” and the answer is: nobody, the summary did.
Everything in the previous nine days leads here. Designing the workflow, briefing, sequencing, directing, stress-testing, prototyping, checking, testing, sharing: all of it is to get a team to a decision they actually made and can defend. The last skill is making sure that’s what leaves the room.
The skill
Close with a decision record, not a summary. It has five sections, and the point of the structure is to keep separate the things a summary blends together: what the model generated, what the team checked, what’s still assumed, what the team actually chose, and what would make them look again.
| Section | From the October sprint |
|---|---|
| Generated | 50 intervention ideas across 6 prompts. 3 prototypes: handler screen, follow-up letter, call script. 8 objections from the stress-test. |
| Checked | Regulator allows pre-bill contact (compliance, Tues 2pm). Billing data lag is 48h, not 24h (data team). 5 customer tests Thursday: 2 of 5 rang before reading the letter; all 5 understood the comparison line on the screen. |
| Still assumed | Call volumes manageable at scale (ops lead disagrees). Print turnaround of 48h (unconfirmed). Customers on the cheque-payment cohort behave like the five we tested (not represented). |
| We chose | Pilot the call-first route with the handler screen, 200 customers, next billing cycle. Letter route parked. Ops lead disagreed on volume and has committed; her objection is logged. Owner: Priya. Review: 12 December. |
| We’ll look again if | Call volume exceeds 15% of cohort in week one. Compliance changes its view. Fewer than 60% of called customers rate it helpful. |
Fig. 10 — A decision record from the October sprint. Names changed. The dissent is in the record, not smoothed out of it.
The team writes it, on the wall, in the last forty minutes. The facilitator reads each section out and asks the question that keeps it honest. Generated: what did the model make? Checked: what did a person or a customer confirm, and when? Still assumed: what are we going ahead on without checking? Chose: what, who owns it, when do we look at it? Reconsider: what would make us stop?
The model can draft it from the wall. The team must correct it. If you’ve done Day 7, you know what to look for.
What goes wrong
“Still assumed” gets left empty, because writing assumptions down feels like admitting weakness in front of the sponsor. It’s the opposite. A sponsor who sees “we’re assuming print turnaround is 48 hours, unconfirmed” trusts the rest of the record more, because it’s obviously been written by people who were paying attention. An empty assumptions box reads as either naive or dishonest.
“We’ll look again if” gets written vaguely. “If it doesn’t work.” That’s not a trigger, it’s a shrug. Numbers and dates. If you can’t name the number, you don’t yet know what success looks like, and that’s worth discovering on day two rather than in month three.
And the one that undoes everything: the record gets written, then the model’s summary gets sent instead because it’s prettier. Send the record. Attach the wall photo. The summary can go as an appendix if someone insists.
The ten, together
If you’ve followed all ten days, you’ve got a complete method for running a room with a model in it. Design the workflow so the model has jobs and reviewers. Brief it like a new starter. Sequence it on purpose. Direct it toward difference. Stress-test what it says and sort the objections. Prototype live with the question pinned up. Check its synthesis against the wall. Rehearse with it, then test with real people. Rotate the keyboard. Close with a record.
None of it is anti-AI. All of it is pro-decision. The model makes everything faster. These ten skills are how you make sure the thing that got faster was the right thing.
Run it Monday
- Block the last 40 minutes of the workshop for the record. Non-negotiable. Cut an exercise instead.
- Five sections on the wall. Let the model draft; the team corrects, line by line.
- Dissent goes in “we chose”, by name, with their commitment.
- “Look again if” needs at least one number and one date.
- Send the record, not the summary.
A decision is something a named person chose, on evidence they can point to, with a date they’ll look again.
If this sounds like your organisation. AI budgets are rising and work isn’t changing. Teams are faster at everything except deciding. A HyperSprint takes one real workflow, redesigns it with the team in two days, and ends with a record like the one above. Start a conversation.
All ten skills, every slide and timebox, the T.O.K.E.N. brief, the role cards, the decision record template and the prompts are in the facilitation deck. Free.
