AI workshop formats are sold by duration — 60 minutes, half a day, a full day, three days — and duration is the wrong variable to shop on. What separates one format from another is the number of minutes participants spend working on their own files, and whether those minutes arrive in one block or several. Everything else on a proposal is packaging.
That matters because one word covers products with wildly different ceilings. A 60-minute session can create interest. It cannot change how a task gets done on Monday. A vendor selling the first will describe it in the language of the second, and the proposal will never say anything false. This guide gives each format an honest ceiling, then hands you a one-minute test to run on any agenda before you sign it.
The six formats and their honest ceilings
Every catalog on the market arranges these six. The fourth column is the one no vendor page prints.
| Format | Typical length | Reliably produces | Cannot produce |
|---|---|---|---|
| Lunch-and-learn | 45–90 minutes | Awareness, a shared vocabulary, volunteers | A changed workflow |
| Focus session | 2–3 hours, one task | One task done a new way by one team | Breadth across roles |
| Half day | 3–4 hours | Two or three reworked tasks, a prompt set | A reviewed, documented process |
| Full day | 6–7 hours | The above, plus written procedures and a baseline | Retention past a few weeks on its own |
| Multi-day immersion | 2–3 consecutive days | Depth on a complex workflow, a trained internal facilitator | Habit, if nothing follows it |
| Spaced cohort | 4–6 sessions of 90 minutes over 6–10 weeks | Practice between sessions, visible habit change | Fast answers; it occupies a quarter |
Read the fourth column first. Most disappointing AI enablement spend is not a bad session; it is a competent session whose format could never have produced the result the sponsor wanted.
Lunch-and-learn. Someone demonstrates ChatGPT, Claude, Gemini, or Microsoft Copilot while 40 people eat. This is a marketing format, and marketing is a real need — if nobody in finance has opened an assistant, interest has to come before practice. Judge it on volunteers, not behavior. If three people stay afterward to ask how it applies to their close process, the hour did its job.
Focus session. Two or three hours on a single recurring task for a single team: the weekly pipeline summary, the first-pass read of a vendor contract. The narrow scope is the strength — one task, one team, one new method, done well enough that someone repeats it unsupervised on Tuesday.
Half day. The most common entry point, and for good reason. Three to four hours is enough to rework two or three tasks with a group that already uses the tools occasionally. Mila’s catalog is typical of where the market settled, offering half-day and multi-day sessions and a full day deliverable as “1 full day or 2 half-days.”
Full day. Six or seven hours buys the half day plus the two artifacts that make the spend defensible later: written procedures for the workflows you changed, and a baseline taken before the exercises. It also buys a problem. Attention collapses in the first hour after lunch, and an agenda that ignores that wastes its most expensive block. The detailed anatomy of a session that changes behavior applies at every length, but a full day is where ignoring it costs most.
Multi-day immersion. Two or three consecutive days, justified by one of two things: a workflow complicated enough that a day of it is a tour rather than a transfer, or a decision to train internal facilitators. Outside those cases, consecutive days are a scheduling convenience for the vendor.
Spaced cohort. Four to six short sessions across two months, with assigned work on real files in between. It is the format most likely to produce visible change and the hardest to get approved, because it asks for calendar commitment from people who have none. It is also the only one where what happens between sessions is designed rather than hoped for.
The two variables that actually separate them
Strip the names and two numbers remain.
Total hands-on minutes. Minutes in which a participant has their own document open and is being corrected. Not demonstration, not Q&A, not the agenda’s total. A 90-minute lunch-and-learn with a 20-minute exercise holds 20 hands-on minutes. A six-hour day with a two-hour guided build holds 120. That factor of six is the whole difference between the two products.
Distribution. Whether those minutes land in one sitting or across several weeks. This sounds like logistics and is not; it determines what kind of result you get, as the research below settles.
A third variable, group size, acts as a multiplier on the first, and the market mostly ignores it.
The agenda arithmetic test
Here is the test. On any proposal, add the minutes given to blocks where participants work on their own material. Divide by the agenda’s total. Below 40%, you are buying a presentation with exercises attached, whatever the cover page says.
Run it on a real example. One vendor’s workshop page states that a session spends “roughly seventy percent of its time on participants doing the work.” The same page publishes its curriculum: four modules totaling 270 minutes, of which the guided build is 120. That is 44%. Claim and agenda sit in the same scroll and disagree by 26 points.
The vendor is probably not lying — the headline was written by someone who did not add up the modules. That is the point. The ratio is the hardest thing to fake in a document you can read before signing, and almost nobody does the addition. Ask for the agenda in minutes. A vendor who will not provide one is telling you something.
Two refinements worth applying:
- Count demos as instruction, not practice. Watching a facilitator do something impressive feels like practice and transfers like a video.
- Count the setup honestly. If the first 45 minutes go to logins and access, those minutes are overhead, and they come out of the hands-on column rather than the lecture column. Settle access before the day starts.
Headcount is part of the format
Hands-on means a facilitator watches someone work and corrects them, which has an arithmetic limit. One published catalog caps in-class sessions at “up to 50 participants” with “options available for bigger groups.” Another vendor’s employer page offers AI training “delivered for 1-1,000+ employees.” Neither describes the same product. Fifty people with one facilitator is a lecture that permits laptops; twelve is coaching.
The useful rule is roughly where the more specific vendor pages land: one facilitator per 12 to 15 participants. Doubling the room without doubling the facilitators does not halve the quality — it changes the format, because nobody can look over 40 shoulders in 120 minutes.
That gives you a second screening question with a right answer: how many facilitators, for how many participants, in which blocks? If the answer is one facilitator for 30 people during the build, you have priced a workshop and purchased a seminar — the same substitution that makes a session built from a generic slide deck so expensive relative to what it delivers.
One long block or several short ones
This is the one design question where evidence exists, and most vendor pages answer it with “it depends.”
The best field study available is a quasi-experimental follow-up with 64 bank employees, 32 per group, comparing spaced against massed training practice. Its finding, in the authors’ words: “Spaced rather than massed training practice resulted in greater transfer quality.” And immediately after: “Spaced training did not surpass massed training in terms of transfer quantity” (Kauffeld and Lehmann-Willenbrock, Sales training: effects of spaced practice on training transfer, 2010).
Read both halves. Spacing did not make people apply the training more often. It made them apply it better. For a buying decision, that resolves cleanly:
- If your problem is breadth — 60 people have licenses and 15 have logged in — a single block is the right instrument. A day gets more people started per unit of calendar pain.
- If your problem is quality — people use the assistant and the output is mediocre, unchecked, or embarrassing in front of a client — spacing is. Quality comes from repetition with correction, and a single day offers one shot at it.
Most companies of 50 to 500 have the second problem and buy for the first, because the first is easier to schedule. The compromise that survives contact with a calendar: one full day to establish the method, then two 90-minute clinics at weeks three and six for the people doing the work. That is the cheapest structure that still contains what spacing contributes.
It only works if someone owns the gap between sessions, which is where most of this spend quietly dies — the follow-up architecture most sessions never get is a separate design job from the agenda.
Where the build sprint fits, and where it does not
Hackathons and build sprints sit in the same catalogs as fluency sessions and are not interchangeable. OpenAI’s own event playbook publishes four of them with minute-level agendas — a one-hour sprint, a half day, a full day, and a three-day version — and every one of the four ends in demos (OpenAI Academy, Hackathon Playbook).
That is an excellent design for its purpose, which is a prototype and a room full of energy. It is not a design for 20 people doing their recurring work differently next week, because nothing in it asks a participant to repeat an ordinary task under correction.
Buy a sprint to learn whether something is possible. Buy a fluency session when you want the work to change. McKinsey’s research on AI upskilling names the failure plainly: training disconnected from employees’ actual jobs does not land, and the programs that work build learning into daily work rather than hypothetical scenarios.
Same budget, three different shapes
Market rates for corporate sessions run roughly $2,000 to $15,000 depending on length, customization, and facilitator count. Hold the budget at the middle of that range and the shape of the spend is a real choice:
| Shape | What you get | Best when |
|---|---|---|
| One full day, 24 people, two facilitators | Breadth. Everyone starts. Procedures written for two workflows. | Adoption is near zero and you need a floor |
| One full day plus two 90-minute clinics, 12 people | Depth with correction. Quality improves and is observed twice. | Usage exists and output quality is the complaint |
| Five 90-minute sessions, 12 people, over eight weeks | Habit. Work between sessions, measured at the end. | You have executive patience and one team that matters |
The first shape is the most popular and the least likely to move a metric. The third is the most effective and the hardest to get on calendars. The second is what actually gets approved, and it is the right default in this size band.
Two cost notes the formats hide. Participant time is the larger expense: 24 people for six hours is 144 hours of payroll, which dwarfs most session fees and belongs in the business case. And the cheapest format is the one whose ceiling matches the result you were asked to produce, not the shortest one. A lunch-and-learn bought to fix output quality is a total loss at any price.
Choosing in one page
Pick from the problem, not the catalog.
- Nobody has started. Lunch-and-learn or a half day. Measure volunteers and first logins.
- One team’s weekly task is the bottleneck. Focus session, 2–3 hours, that team only.
- Several teams use assistants and the output is uneven. Full day plus two clinics. The clinics fix quality.
- One workflow is complex and high-value. Multi-day immersion, with a named internal facilitator in the room from day one.
- You have been asked to show a return by next quarter. Any format, but take the baseline before the exercises begin. Without it, the February conversation has no arithmetic in it.
- Leadership needs to decide what to fund. A briefing, not a workshop. Do not mix the audiences; the executives will set the pace and nobody will practice.
Underneath all six sits the question this category avoids: licenses are not capability, and no format produces capability by being long. That is the subject of our argument on why buying licenses is not the same as building capability, and the mechanics of designing any of these sessions are set out in our guide to building a session that changes Monday morning. If you are still settling what the thing is before choosing how long it runs, start with what a session should actually produce.
Frequently asked questions
What are the main AI workshop formats?
Six: lunch-and-learn (45–90 minutes), focus session (2–3 hours on one task), half day (3–4 hours), full day (6–7 hours), multi-day immersion (2–3 consecutive days), and a spaced cohort of four to six short sessions over six to ten weeks.
How long should an AI workshop be?
Long enough to hold at least 90 minutes of participants working on their own files, which makes a half day the practical minimum for behavior change. Past seven hours in one sitting, extra length returns little.
Is a lunch-and-learn worth running?
Yes, for awareness. It creates vocabulary and surfaces the people who want to go further. Judge it on volunteers and first logins, not on changed work, and do not buy one to fix a quality problem.
What is the best AI workshop for teams that already use the tools?
A full day followed by two 90-minute clinics at roughly weeks three and six. Existing users need correction on work they have already attempted, which is what the clinics provide.
How many people should attend?
Twelve to 15 per facilitator for any hands-on block. Rooms of 40 or more are seminars that permit laptops — a valid format with a different ceiling. Ask how many facilitators cover the build block.
Should sessions run on consecutive days or be spread out?
Spread out, if quality is the problem. Field research on spaced versus massed practice found spacing produced greater transfer quality but not greater quantity. Consecutive days suit complex single workflows and facilitator training.
What does a corporate AI workshop cost?
Market rates run roughly $2,000 to $15,000 per engagement depending on length, group size, facilitator count, and how much customization to your own documents is involved. Most vendors quote case by case.
How is a hackathon different from a workshop?
A hackathon produces a prototype and ends in demos. A fluency session produces changed procedures and ends in written steps someone repeats next week. Published hackathon agendas give their final block to demonstrations, which tells you what they optimize for.
Is a virtual session as good as in person?
For focus sessions and clinics, close enough, provided breakout rooms let facilitators watch people work. Virtual full days lose the room after lunch and are better split in two.
What should I ask for before signing?
The agenda in minutes, the facilitator ratio during hands-on blocks, the deliverables that leave the room, and confirmation that participants work on your own documents.
Key takeaways
- Formats differ on two numbers: total hands-on minutes, and whether they are spaced. Duration alone predicts nothing.
- Divide hands-on minutes by total minutes on any agenda. Below 40%, you are buying a presentation. One published vendor agenda lands at 44% under a 70% headline.
- Group size is part of the format. One facilitator per 12 to 15 participants, or the build block becomes a demonstration.
- Spacing improves how well people apply training, not how often. Buy a single block for breadth; buy spacing for quality.
- A full day plus two short clinics is the best-value structure for most companies of 50 to 500 people.
- Build sprints and fluency sessions are different purchases. One ends in a prototype, the other in procedures.
- Participant payroll usually exceeds the session fee. Put it in the business case, and take a baseline before the first exercise.
Still deciding which shape fits the problem you have been handed? Book a free AI readiness call and we will work through it against your team’s own workflows.
Arthur Wandzel