Ninety minutes is enough to change what a team does on Monday, and not enough to do most of what gets put on a 90-minute agenda. The constraint is arithmetic: eight to sixteen people, one thing you can plausibly buy with the time, and every block you add taking minutes from the one block that produces the change.
Most published agendas spend the budget on the wrong half of the room. The best free one available — Atlassian’s team playbook for an AI training workshop — lists twelve steps with minute allocations, and the session steps add up to 85 minutes against a stated run time of one hour. Of those 85 minutes, 15 are labeled “time to play.” Fifty are three use-case walkthroughs delivered from the front. In the template at the top of the search results, participants watch for roughly three minutes per minute of hands on the tool.
The 90-minute run of show
This is the agenda we run when 90 minutes is the whole slot and the group is one function, eight to sixteen people with licenses already in hand.
| Minutes | Block | What it buys | Fails if |
|---|---|---|---|
| 0–6 | The frame | One sentence on the goal, one on what is off limits, one on what each person leaves with | It becomes a history of language models |
| 6–16 | One task, driven live | A worked example on a real document from this team, narrated including what went wrong | The example is a generic prompt on a made-up scenario |
| 16–46 | Everyone does the same task on their own material | The only 30 minutes that reliably changes next week | Anyone is still watching a screen at minute 20 |
| 46–60 | Check the output against the source | The habit that keeps the gains | It is folded into the practice block and quietly skipped |
| 60–74 | Compare and correct | Three people read out what they got, wrong answers included; the room sees the failure modes | It becomes a showcase of the best result |
| 74–84 | The never-paste list and the save | What does not go into an assistant, plus the two or three prompts worth keeping | The list is a policy slide, not the room’s own sentences |
| 84–90 | One commitment each | A named task each person will do this way before Friday | It ends on “any questions?” |
Fifty-eight of the 90 minutes — blocks three, four and five — have participants working on their own material. That is the number to defend.
Count the minutes somebody else is driving
Do the arithmetic on any AI workshop agenda in front of you. Add up the minutes in which the facilitator or a slide is the thing moving, then the minutes in which every participant is typing against their own work. One of those fractions is what you are paying for.
The convention puts the demonstration first and at length and the doing last and short, because a demonstration is easy to deliver well to any size of room and hands-on time is not. Thirty people in one room makes the three-to-one ratio almost inevitable — a room-booking decision disguised as a teaching decision.
You will also see a figure quoted that hands-on activity should exceed 50% of the time, circulated widely enough that AI assistants repeat it as a finding. It traces to one vendor’s design guide, alongside a claim that the threshold “drives behavior change in 80%+ of participants.” Neither has a study behind it. The instinct is right and the citation does not exist, so set your own fraction from what each block is for.
What the room can buy, and what it cannot
The training-transfer literature is unambiguous about where sessions lose their value, and it is not in the room.
Arthur, Bennett, Edens and Bell pooled 1,152 effect sizes from 165 sources on organizational training. Sample-weighted mean effects were large and consistent: d = 0.63 for learning criteria across 234 data points and 15,014 people, and d = 0.62 for behavioral criteria across 122 data points and 15,627 people. Only 5.8% of the effect sizes were negative. Training works, and that question is settled well enough to stop arguing about it.
The finding that matters for an agenda is their within-study comparison. Among studies measuring both learning and behavior on the same participants, the average decrease from the learning measure to the behavioral measure was 0.77 — a drop the authors call fairly large and attribute to transfer conditions, lack of opportunity to perform, and skill loss. People learn the thing. Then they do not do it, because nothing afterward requires them to.
The same paper removes a line most workshop pitches rely on. Lectures, the authors note, have an anecdotal reputation as boring and ineffective, and their mean effects were generally favorable: d = 0.45 on cognitive content against a learning criterion, 0.50 combined with discussion, and 0.71 against a behavioral criterion across ten studies and 8,131 people. Discussion alone came in at 0.20. So the case for a hands-on agenda is not that talking fails. It is that 90 minutes has to be spent on what survives the walk back to the desk, and watching somebody else succeed is not it.
The block almost no agenda has
Every agenda in the market teaches prompting. The one randomized trial that tested whether a short prompting primer helps found that it made people worse.
Dell’Acqua and colleagues ran a randomized trial with 758 consultants at Boston Consulting Group across three conditions: no assistant, an assistant, or an assistant plus a brief prompt-engineering overview. On the task inside the assistant’s competence, the AI groups completed 12.2% more subtasks, worked about 25% faster, and were graded roughly 32% higher on quality. That is the prize, and it is real.
On the second task — a brand-strategy case requiring spreadsheet figures to be reconciled against a contradicting clue buried in interview notes — the results inverted. Correct answers: 84.5% with no assistant, 70.6% with the assistant, and 60% with the assistant plus the prompting overview. The assistant handled the surface numbers, missed the insight, and the consultants trusted it. The training add-on inoculated nobody; it appears to have raised confidence in output that was wrong.
Read that as an instruction about minutes 46 to 60. The skill a 90-minute session can install is not prompt construction, which people pick up from use. It is the reflex of checking a plausible answer against the source before forwarding it. That block needs a task where the assistant is wrong on the team’s own data, and it needs its own protected quarter hour, because folded into the practice block it is always what gets dropped.
Noy and Zhang’s experiment points the same way. Across 453 college-educated professionals on occupation-specific writing tasks, time taken fell 40% and graded quality rose 18% — and the authors were explicit that outputs were not fact-checked, and that real gains would be smaller once checking time is counted. Verification is not an add-on to the productivity case; it decides whether the measured gain is real.
Three things to cut, in order
When the content does not fit the clock — and it never fits — cut in this order.
First, the tool tour. A pass through four assistants and their interface differences costs 12 to 15 minutes and transfers nothing, because the interfaces will have changed by next quarter. Teach one assistant deeply and name the others once. Whether to teach one assistant or several is worth settling before the agenda is drafted.
Second, the capability overview. Model families, context windows, how training works. Ten minutes, zero behavior change, and the block that makes a session feel substantial to whoever booked it while teaching the room nothing it can use. If leadership needs that orientation, it is a different session.
Third, the closing demonstration. The impressive multi-step workflow at the end is the most memorable ten minutes and the least useful, because nobody in the room can reproduce it on Tuesday. Keep it only if the slot runs long, and label it a preview.
What you do not cut is the comparison block at minutes 60 to 74. It looks optional on paper. It is where the room learns that everybody’s first attempt was mediocre, which is the most useful thing a group of adults can learn together about a new tool, and it cannot be delivered from the front.
Why two separated sessions beat one morning
If you have 180 minutes of AI enablement budget, two 90-minute sessions two weeks apart will beat one three-hour morning.
Cepeda, Pashler, Vul, Wixted and Rohrer synthesized 839 assessments across 317 experiments in 184 articles on distributed practice. Their central result is that the interval between study episodes and the interval before the final test operate jointly: the gap producing maximal retention increases as the period over which you need the skill retained increases. If you want the capability present in a month, the second sitting should not be on the same morning.
Be careful how much weight that carries. The corpus is verbal recall, not workplace software skill, and no published trial has tested a 90-minute AI fluency session against a half-day one. This is design inference, not a measured result, and anyone quoting you a figure on it is making it up. The practical version: the second sitting is where people bring the task that broke, and a one-to-three-week gap gives them time to have one. Nothing gets demonstrated; the facilitator triages. Whether that follow-up should be a session, a standing slot or nothing depends on what the first one taught, which is the argument in our comparison of a single session against an ongoing coaching arrangement.
What has to be true before and after
Two conditions sit outside the 90 minutes and decide most of its value.
Before: every participant arrives with one real document and one recurring task. Not a scenario, not a sample. If that pre-work does not land, the practice block degrades into generic prompting on invented problems and you have bought a demonstration with extra steps. Send one line with a deadline the day before, and expect to chase a third of the room.
After: somebody asks to see the changed work. The 0.77 drop between learning and behavior is the measurable shape of nobody asking. A manager who says “show me how you did that draft” in a regular one-to-one costs nothing and is the highest-return item in the plan. A standing half hour for whoever is stuck is the next cheapest, and the attendance pattern it runs into is covered in our look at weekly office hours after the session.
Neither is a training purchase, which is why no vendor agenda includes them, and it is the same confusion that lets license spend get counted as capability building — the argument in the case against mistaking licenses for capability. The wider design question, across longer formats, belongs to our guide to designing a session that changes Monday morning.
What a session this size costs
Hands-on AI enablement sessions sit in a market range of roughly $2,000 to $15,000, and a single 90-minute session for one function lands at the bottom of it. The default unit for a corporate AI workshop is a day: BrainStation’s AI Fundamentals workshop is a full-day in-person format running 9:00 to 4:30 across nine named modules, with no price published and a contact form instead.
A 90-minute session is not a cheaper version of that day. It is a different purchase — one function, one task family, one habit — and the right first one when you do not yet know which team will benefit most. Buy the 90 minutes, check in three weeks whether anything in anybody’s output looks different, and let that decide whether you buy the day. If a short session quotes near the top of the range, ask what the preparation is: two to three hours of build per hour delivered is realistic on custom material.
Frequently asked questions
What should a 90-minute AI workshop agenda contain?
Six blocks: a short frame, one task driven live on the team’s real material, 30 minutes of everyone doing that task on their own work, a protected block on checking output against the source, a comparison round where wrong answers get read out, and a commitment each. Roughly 58 of the 90 minutes should have participants working rather than watching.
Can 90 minutes actually change behavior?
For one task family in one function, yes — if the practice runs on real work and somebody asks to see the result the following week. It will not produce fluency across a company, and no agenda claiming otherwise is being straight with you.
How much of an AI workshop should be hands-on?
More than half. The widely quoted 50% figure is a vendor invention rather than a research finding, so set the fraction from what each block is for. In 90 minutes, three blocks totaling 58 minutes is the number to check in any proposal.
Is 90 minutes or a half day better for a first session?
Ninety minutes first, for one function. It is cheap enough to run before you know which team benefits most, and the distributed-practice evidence suggests a second separated sitting beats extending the first one.
How many people should be in a 90-minute session?
Eight to sixteen. Above that the comparison block stops working and one facilitator cannot support the practice block, which is how sessions drift back into watching rather than doing.
Should the workshop teach prompt engineering?
Only as a by-product. The one randomized trial that added a short prompt-engineering overview found accuracy on the harder task fell to 60%, below the 70.6% of the group with the assistant alone and the 84.5% with none. Spend those minutes on verification.
What does the verification block look like in practice?
A task where the assistant is confidently wrong on the team’s own data — a figure that contradicts a source document, or a summary that drops the clause that mattered. People find the error, say how, and write down the check they will apply from now on.
What pre-work do participants need?
One real document and one recurring task each, requested the day before in a single line. Without it the practice block runs on invented examples and stops being training on the team’s work.
How soon should we see changed behavior?
Within about three weeks on the tasks the session covered, or not at all. If nothing in anybody’s output looks different by then, more sessions will not fix it — the problem is the pre-work, the task choice, or whether any manager is asking.
What does a 90-minute AI workshop cost?
Hands-on sessions sit in a market range of roughly $2,000 to $15,000, with a single short session for one function at the low end. Expect two to three hours of build for every hour delivered when the material runs on your own documents.
Key takeaways
- The constraint is arithmetic. Of 90 minutes, defend the 58 in which participants work on their own material.
- Audit the ratio in any proposal. The most-cited free agenda in this category gives 15 of 85 timed minutes to hands-on work and 50 to walkthroughs from the front.
- Training works; transfer is where it leaks. Across 1,152 effect sizes from 165 sources, the within-study drop from learning to behavioral measures averaged 0.77.
- Teach verification, not prompt tricks. A short prompting overview dropped accuracy on the harder task to 60%, against 70.6% with the assistant alone and 84.5% with none.
- Cut the tool tour, the capability overview and the closing demonstration before you cut the comparison round.
- Two separated sittings beat one three-hour morning, and the pre-work request plus one manager asking to see the work are worth more than any block inside the room.
Before you book a room, work out which function and which task the 90 minutes should buy. A free AI-readiness call will tell you where a first session is likely to show a result, what the pre-work needs to be, and whether to start with a short session or a full day.
Arthur Wandzel
