Twelve people is the right AI workshop cohort size for a hands-on session with one facilitator, and forty is a different product wearing the same name. The reason is arithmetic rather than taste. What a team buys in a hands-on session is a trained person reading their real output and naming what is wrong with it. That attention is fixed, it gets divided by the number of people in the room, and a six-hour day split twelve ways gives each person thirty minutes. Split forty ways it gives nine.
Every vendor publishes a number and none publishes that division. The spread is extreme — one provider recommends 50 to 100 people for a 90-minute hands-on session, another caps the room near ten per instructor — and it tracks delivery model rather than evidence. Here is how to work it out, and the ten rules it produces.
The published numbers disagree by 10x
Four providers state a cohort size in public. All four numbers were on the vendors’ own pages on October 22, 2026.
| Provider | Published cohort size | Format | Stated reason |
|---|---|---|---|
| Lyzr | 50 to 100, recommended | 90 minutes, live, with real-time code reviews | “Maximum peer collaboration, live interactive engagement, and high-impact team output” |
| FutureProof | 20 to 100 per cohort | 4 to 6 week program | None published |
| Improving | 5 to 15 “ideal” for the two-day hands-on workshop; “any” size for executive sessions | 90 minutes to 12 weeks, by program | Published per program, not per customer |
| DataNorth | Around 10 per instructor | 2 to 4 hours | So “everyone can receive direct support and feedback during exercises” |
One provider’s recommended minimum is five times another’s maximum, and both sell hands-on corporate AI workshop delivery to the same kind of buyer.
The third row is the most informative, because it is one vendor disagreeing with itself in public and being right both times. Improving publishes 5 to 15 as ideal for its two-day hands-on build workshop and “any size” for its executive sessions. That is the real rule, printed by accident: hands-on intensity sets the ceiling, and nothing else does. A briefing scales to a few hundred. A session where everyone produces something does not.
Only one of the four states a number per instructor, which is the second tell. A room number is a logistics claim; a ratio is a staffing commitment, and staffing is the whole cost structure of hands-on teaching. A vendor who publishes 100 and no ratio is saying headcount does not change what they send.
None of which makes the large numbers dishonest. A 90-minute session for 80 people is worth buying as a demonstration. The error is pricing it against a working session.
The arithmetic nobody publishes
Facilitator attention is the scarce input, so measure it. A one-day session with six hours of working time and one facilitator holds about 360 minutes of attention for the whole room.
| People in the room | Attention per person | What that buys |
|---|---|---|
| 8 | 45 min | Several passes each, including the second attempt |
| 12 | 30 min | One close look each, plus the hard case |
| 16 | 22 min | One look each, nothing for the quietly stuck |
| 20 | 18 min | Triage — the loudest problems only |
| 40 | 9 min | A lecture with a queue at the front |
| 80 | 4 min | A webinar |
Those are upper bounds, assuming every minute goes to individual attention and none to teaching, setup, or transitions. In practice half the clock goes to the session itself, which halves every figure in the middle column.
A second calculation sets the real ceiling. The part of a session that changes behavior is the round where the facilitator reads several real outputs aloud and names the difference between the best and the worst. Done properly, one output takes four to five minutes: read it, say what the model was missing, show the fix. Twelve people is an hour, which fits in a day. Forty is more than three hours, which does not, so the round shrinks to three volunteers and a show of hands. Nothing on the invoice records that it left.
So twelve is not a preference. It is the largest number at which the central mechanic still fits inside the session. How to build a day around that mechanic is in the AI workshop design playbook.
What the randomized evidence found
Corporate AI training has no peer-reviewed literature on group size. Adjacent fields do, and the best study is more interesting than the vendor consensus.
Mahling and colleagues ran a prospective, randomized, double-blind trial at the University of Tübingen, published in BMC Medical Education (14:185, 2014). Seventy-four fourth-year medical students were taught basic life support in groups of three, five, or eight per tutor, with four trained tutors rotated across the group sizes so the teacher was not a confound. Sixty-nine datasets were analyzed.
The headline result is the part vendors would rather you skipped: group size did not affect the test score. Pass rates were 93%, 100% and 100%, and median checklist scores 48.0, 47.5 and 47.5 (p = .96). By the measure most training programs use, the big group did fine.
Everything else moved.
| Measure (medians, except where noted) | Groups of 3 | Groups of 5 | Groups of 8 |
|---|---|---|---|
| Question-and-answer exchanges per student | 3.0 | 2.0 | 0.5 |
| Hands-on time per student | 4:07 | 4:32 | 2:16 |
| Unrelated conversations (mean) | 2.9 | 16.0 | 17.0 |
| Self-rated confidence after training | 8.2 | 8.0 | 7.0 |
Students in groups of eight asked a sixth as many questions as students in groups of three (p < .001), got about half the hands-on time (p = .02 against groups of three, p = .006 against groups of five), sat through roughly six times as much off-topic conversation, and rated their own competence lower (p = .007).
Two things follow for anyone buying hands-on AI training. A post-session quiz will not detect what a large room costs you, because the quiz is the one measure that did not move — so a buyer who sizes cohorts by the score keeps choosing the larger room. And the measures that did move are the ones that carry a new habit back to the desk.
The authors note their task was a well-defined procedural skill, and that smaller groups may matter more for complex skills and less advanced learners. Getting useful work out of a general-purpose assistant on your own messy documents is harder than chest compressions, so read the study as a floor on the effect rather than a ceiling.
The one published staffing standard outside the vendor set agrees: The Carpentries, whose curriculum is used for hands-on technical teaching at hundreds of institutions, recommends “1 helper for every 8-10 students, not counting the two Instructors.”
The 10 Rules of cohort size
1. Size the cohort to the attention you are buying, not the room you booked
The conference room seats thirty, so thirty people get invited. Do the division first and let the answer pick the room.
2. Twelve is the number where every output can be read out loud
Four to five minutes per output, twelve people, one hour. That round is the product. Pick the largest cohort that still leaves room for it — twelve to sixteen in a one-day session.
3. Count attendance, not invitations
Every published cohort number is an invite-list number. At a 70% show rate a 40-person invite is a 28-person room, and a 14-person invite is a 10-person room. Both errors cost money: the over-booked room gets a lecture, the under-filled one pays capped-fee rates for empty chairs, as the per-head versus flat-fee arithmetic shows.
4. Above twelve, add a facilitator — not chairs
Published standards run from one helper per five learners to one per ten. Near twelve, a facilitator stops reading outputs and starts triaging. Ask any vendor one question: how many people are you sending, and what is each doing during the hands-on block.
5. Budget for stuck minutes, not heads
AI sessions break differently from software training. Every output differs, so there is no correct screen to point at, and the failures are individual: a refused request, a login carrying the wrong account, a document the model misreads. Each is a two-to-eight-minute private conversation. Estimate three per person — at forty people, two days of someone’s time compressed into six hours.
6. Expect a big room to pass the test and learn less
Scores hold while hands-on minutes halve and questions collapse. Measure a session by what people do in the following three weeks, not by what they recall at 4 p.m. Which measures to collect is in our case for weekly office hours after the workshop.
7. Halve the number when the session is remote
Remote delivery removes the facilitator’s cheapest diagnostic: glancing at twelve screens while walking. Reading one shared screen at a time costs several times as long, and muted tiles hide the person who gave up twenty minutes ago.
8. Split by starting fluency before you split by headcount
If forty people need training, the first cut is not alphabetical and not by department. Separate the people who use an assistant daily from those who have opened one twice. A mixed room forces the facilitator to teach to the middle, which bores one half and loses the other.
9. Two rooms of twelve beat one room of twenty-four
Two sessions cost more. They also produce roughly four times the individual attention and leave you with twelve people who already did it, sitting next to the twelve about to. Run the second cohort two to three weeks later and recruit two of the first as helpers — the cheapest facilitator you will ever staff, and where internal AI enablement capability starts. Sequencing it across a larger population is in our guide to rolling AI workshops across 200 staff.
10. If you cannot follow up with the cohort, it is too big
Follow-through means somebody reads what each person produced in week one and answers what they got stuck on in week two, and that labor scales with headcount exactly like the session does. A cohort sized for the room but not for the three weeks after it decays to baseline — the difference between a one-off workshop and ongoing coaching.
Why the 40-person room keeps getting proposed
Nobody designs a 40-person hands-on session. It arrives by subtraction, and two forces produce it.
The first is budget per head. ATD’s 2026 State of the Industry report, covering 2025, put direct learning spend at $846 per employee, down from $1,254 the year before, while formal learning hours per employee rose to 16.7 from 13.7. More hours on a third less money resolves in one direction: more people per session. A $6,000 day across 40 people is $150 a head; across 12 it is $500. The first number clears a budget review and the second needs an argument.
The second is the calendar. Pulling 40 people out of billable work three times is harder to schedule than once, so one large session looks considerate.
Both pressures are legitimate, and neither changes the division. Decide what the session is for before sizing it. Awareness across a department is one large session, priced and measured as one. Forty people working differently in November is three or four cohorts, and the real question is whether the capability is worth the difference. Treating a 40-person room as a cheap version of the second thing is the one option that reliably wastes the money — the broader argument in our case that buying licenses is not the same as building capability.
Hands-on training for a team runs roughly $2,000 to $15,000 in the current market, and cohort size moves that total less than buyers expect, because the cost is facilitator days rather than seats. Verify any provider’s current pricing and group limits on their own pages before comparing — corporate training catalogs change quarterly. A free AI-readiness call is where we run your attendance, fluency spread, and follow-up capacity and come back with a cohort count.
Frequently asked questions
What is the ideal cohort size for an AI workshop?
Twelve to sixteen for a hands-on session with one facilitator, and eight if it is remote. The limit comes from the round where the facilitator reads real outputs aloud and names what each got wrong, at four to five minutes per person. Twelve fills an hour; forty needs more than three, so the round gets cut.
How many participants can one facilitator handle in hands-on AI training?
About eight to twelve for close work. The Carpentries recommends one helper per 8 to 10 learners on top of two instructors, and one AI training provider publishes roughly ten participants per instructor. Above twelve, a single facilitator shifts from reading outputs to triaging the loudest problems.
Is a 40-person AI workshop worth it?
As awareness, yes, if it is priced and measured as awareness. As behavior change, no. One facilitator across 40 people has about nine minutes each in a six-hour day before teaching time is subtracted, and the feedback round does not fit. Forty people working differently is three or four cohorts.
Why do vendors publish such different cohort sizes?
Because the number describes their delivery model rather than the evidence. Published recommendations run from about 10 per instructor to 50 to 100 per cohort — a 10x spread inside one category. One provider publishes both: 5 to 15 for its hands-on build workshop, “any size” for its executive briefings.
Does research support smaller training groups?
The best randomized study comes from medical education, and the answer is more precise than yes. Mahling and colleagues (BMC Medical Education, 2014) taught 69 students in groups of three, five and eight per tutor. Pass rates were 93% to 100% across all three, but groups of eight asked a sixth as many questions, got about half the hands-on time, and sat through six times more off-topic conversation.
Will a post-session quiz tell us whether the cohort was too big?
No, and this is the most expensive thing to get wrong. In the randomized data the assessment score was the single measure that did not move with group size, so a quiz will report that your 40-person session worked. Measure what people produce in the following two to three weeks instead.
Should managers and their direct reports be in the same cohort?
Avoid it where the session involves showing work in progress. The output round depends on people volunteering their worst attempt, which is a different decision when the person who sets your review is watching. Where headcount forces the overlap, split by reporting line first.
What is the right cohort size for a remote AI workshop?
Eight, with cameras on during the feedback round. Remote delivery removes the facilitator’s ability to scan screens while circulating, so each review costs several times as long, and a muted grid hides the person who got stuck and stopped trying.
Key takeaways
- A session is a quantity of facilitator attention. Six hours across 12 people is 30 minutes each; across 40 it is 9.
- Twelve to sixteen is the largest cohort in which every output can still be read aloud and corrected inside the day. Eight if remote.
- Published vendor sizes run from about 10 per instructor to 50 to 100 per cohort. One vendor prints both, segmented by hands-on intensity, which is the real rule.
- Randomized evidence found test scores unchanged across groups of three, five and eight while questions per learner fell from 3.0 to 0.5 and hands-on time halved. A quiz cannot detect what a large room costs.
- Count attendance, not invitations: at a 70% show rate, a 40-person invite is a 28-person room.
- The 40-person session is a budget artifact — learning spend per employee fell to $846 while hours rose. Right for awareness, wrong for capability.
- Two cohorts of twelve beat one of twenty-four: the first becomes the second’s helpers, and the enablement capability starts staying in-house.
Do the division before you book the room, ask how many facilitators are coming, and size the cohort to the follow-up you can sustain. A free AI-readiness call turns your headcount, fluency spread, and calendar into a cohort plan with a number attached.
Arthur Wandzel
