An exercise earns its place in an AI workshop when it produces something a participant can use on Monday. Not a feeling of confidence, not a completed activity, but an artifact: a prompt that reproduces a document they write every week, a checklist for the next draft an assistant hands them, a scored list of tasks in their own job worth trying first. Every exercise below is specified by that output, and the three at the end produce nothing.
The default session is built the other way around. A vendor writes an agenda of block names and durations, the room watches most of them, and the closing survey measures whether people enjoyed it. Ask a provider what each participant will walk out holding and the answer is usually an adjective. That question separates a real curriculum from a slide deck, and AI enablement that survives the week is mostly a matter of choosing well below and protecting the minutes.
The four tests an exercise has to pass
Four questions, applied to every candidate. An exercise that fails two is entertainment.
- Does the participant do the typing? Watching a facilitator produce a perfect output teaches recognition, not production. In their review for Psychological Science in the Public Interest, Dunlosky and colleagues rated practice testing and distributed practice the highest-utility learning techniques, with rereading and highlighting at the bottom. A demonstration-heavy workshop is the rereading of corporate training.
- Does it run on the participant’s own material? A clean sample invoice behaves nothing like the real one, which is missing two fields and contradicts the email thread above it.
- Does it produce an artifact that leaves the room? A saved prompt, a written check, a scored shortlist. If nothing is written down, the capability existed for twenty minutes.
- Is failure visible inside the exercise? People need to watch an assistant get something wrong while someone who can explain why is in the room, or the first wrong answer they meet alone becomes the reason they stop.
The library, in one table
| Exercise | What each person produces | Time | Fails when |
|---|---|---|---|
| Document loop | A saved prompt that reproduces a weekly deliverable | 35–45 min | The document is a one-off, so the prompt is never reused |
| Verification drill | A written three-check routine for their outputs | 20–25 min | It becomes a debate about trusting AI in general |
| Workflow teardown | A scored shortlist of five to eight recurring tasks | 30–40 min | Scoring is skipped and the list becomes a wish list |
| Show-then-do ladder | Three completed tasks at rising difficulty | 45–60 min | The practice half is cut when the demo overruns |
| Handoff pair | One prompt per pair, plus a named next user | 20–30 min | Pairs self-select and the confident users group up |
| Rewrite-the-request | A before-and-after pair of one real request | 15–20 min | Treated as a vocabulary lesson, not practice |
| Boundary test | One sentence on what they will not put in a tool | 10–15 min | No written data rule exists, so the room invents one |
| Commitment round | One task named out loud, with a date | 10 min | No follow-up exists, so nothing is checked |
Eight exercises, 185 to 255 minutes at the extremes. A half day holds four or five, and fewer as the room grows — which is the argument in our case for a cohort of twelve rather than forty. How they fit a short run sheet is covered in our walk-through of a 90-minute session; the case for each one follows.
The handoff pair deserves a note, because most rooms hold someone who uses an assistant daily and someone who has opened one twice, and teaching to the middle loses both. Give the pair asymmetric jobs: the experienced person drives the first attempt and narrates it, then hands over and only answers questions. Assign pairs from the intake, or the confident users will group themselves.
The document loop
Each person opens a document they produce on a recurring basis — the weekly pipeline summary, the client update, the incident write-up — and prompts an assistant to produce the next one. The output is wrong. They name what is wrong in a sentence and prompt again. Three to five rounds, then they save the prompt that worked somewhere they will find it next Thursday.
This is the whole product in one exercise. The loop — attempt, read critically, correct, repeat — is the transferable skill, and the saved prompt converts a session into a habit. It is also the first exercise to get cut, because it is slow and the room goes quiet while it runs. Quiet is the point.
Two details decide whether it lands. The task has to recur, because a prompt for a document written once a year will never be opened again. And the facilitator circulates rather than standing at the front: the teachable moment is the first failed attempt, invisible from the podium. This is where the material gathered during the pre-work a buyer owes the facilitator gets spent.
The verification drill
The facilitator hands out a plausible AI-generated draft with errors planted in it: an invented figure, a real-sounding citation that does not exist, a confident claim that contradicts the attached source. Participants have fifteen minutes to find them, then each writes the three checks they will run on their own outputs from now on.
Verification deserves its own exercise rather than a bullet on a governance slide, and it is the only one that makes a room safer rather than faster. A team that has spent twenty minutes catching a fabricated citation does not need a memo explaining why that matters.
The failure mode is predictable: someone will want to debate whether these tools are trustworthy in general, and that conversation has no end. Keep it concrete — this draft, these errors, three checks.
The workflow teardown
Participants list the tasks they repeat weekly, then score each on two axes: how often it happens, and what it costs if the output is wrong. Frequent and low-consequence is where to start. The ambitious, high-consequence corner is where the enthusiasm goes and where the value is not.
Several vendor curricula include a version of this as process mapping, scoring each action for feasibility, effort, and the need for a human check. The scoring is the part that gets dropped, and without it the exercise produces a wish list. With it, the output is a shortlist a manager can approve, so the session has produced a decision. For a multi-function rollout the teardown also says who should go next, the question our framework for running several tracks in one company exists to answer.
Show-then-do, and why the third rung matters
Demonstrate a task, then have the room do the same task, at three rising levels of difficulty. Atlassian’s Teamwork Lab publishes this structure as a play with per-step minutes and reported from its own session that 33% of attendees felt more confident using agents and weekly use of its assistant rose 24%. A measured before-and-after is rare enough here to be worth borrowing.
Two things in that play are more instructive than the headline numbers. It lists two to three hours of facilitator preparation against a one-hour session, and its timed steps add up to roughly 85 minutes rather than 60. Both are the economics of exercise design showing through: preparation outweighs delivery two or three to one, and agendas overrun. That ratio is why a per-seat catalog course cannot run exercises on your documents.
The rung that matters is the third. If the advanced step is cut for time, the strong users learn nothing and conclude the session was remedial. Trim the demonstration halves instead.
How much of the session is practice
More than half, and nobody who sells sessions disputes it. Published guidance converges above 50% of agenda time, with one practitioner guide putting its applied phase at 40 to 50% of the clock and reserving separate time for reflection and commitments.
The reason to take the high end is the training-dose evidence rather than facilitator preference. BCG’s 2025 survey of more than 10,600 workers across 11 countries found that only 36% of employees felt adequately trained, and that those receiving five or more hours of training — particularly in person and coached — were significantly more likely to become regular users. Five hours is a small budget; spend half on presentation and the real dose is two and a half, and the coached part of that finding only happens during exercises.
So count the minutes in which participants are typing. Under half the session and you are buying a presentation with activities attached, whatever the agenda calls it. The broader case for designing backward from behavior sits in our workshop design reference.
Real documents or sample data
Real documents, redacted to one standard by one person rather than by thirty people improvising. This is the largest single difference between a session that changes work and one that demonstrates software.
Confidentiality is a fair objection with practical answers. Redact against a written rule — no client names, no personal data, no unreleased financials — and show one worked example before anyone touches a tool. Use the enterprise tier of the assistant the company has sanctioned, where retention settings are administrable. Where redaction is unsafe, substitute a structurally identical document from a closed matter; structure is what the exercise teaches. The sanitized sample is what fails, because a clean first result removes the correction loop.
Three exercises that are theater
These appear on real agendas, including good ones, and none survive the four tests.
The scavenger hunt. A list of tasks to complete with AI tools, first team done wins a prize. Fun, and it measures speed at finding features. Nobody’s job is finding features. If you want competition, award it for a saved prompt another team can run on their own document.
Stump the AI. Participants try to make the assistant fail. Useful once, as a two-minute framing device before the verification drill. As a standalone segment it trains the room to collect failures rather than catch them in work that matters, and hands the skeptics ammunition without teaching anyone a check.
The build sprint. Ninety minutes assembling a custom assistant or a small automation. Strong closing demo, and an artifact nobody owns: two months later it has drifted out of date and whoever built it did not volunteer to maintain it. Custom builds are a project with a budget and an owner, not an exercise. Keep the session on fluency and let the teardown say which build to scope.
What published programs actually have people do
Worth checking rather than assuming, because every catalog describes itself as hands-on. Verified against the live pages in October 2026:
| Program | What a learner actually does | Terms |
|---|---|---|
| Google AI Essentials on Coursera | Five courses, under 10 hours; the applied project writes prompts and evaluates outputs against supplied workplace scenarios | $49 per month after a 7-day trial; no prerequisites |
| Work smarter with AI on Microsoft Learn | One module, reading plus knowledge checks; no lab environment | Free; assumes familiarity with the productivity suite |
| SessionLab’s generative-AI activity list | The filter returns a single method: generate a fictional monster from three characteristics | Free to browse |
| Private delivery | Exercises built from the team’s own documents, collected in advance | Market range roughly $2,000–15,000 per session |
None of this is dishonest. They are different products, and per-seat economics make preparation on your material impossible — the same arithmetic the Atlassian play exposes from the other side. The corpus of published, buyer-facing workshop exercises is close to empty: a fact about the market, not a complaint about any vendor in it.
How you know the exercises worked
Not the satisfaction score. A room that enjoyed itself is a weak signal, and a 98% recommendation rate tells you the facilitator was likable. Three checks that mean something:
- Count the artifacts. How many saved prompts, written checks and scored shortlists exist in a shared place two days later. One person can collect that in ten minutes.
- Re-run one exercise at 30 days. Give the same task from the document loop and see whether people reach a usable output in fewer rounds. That is a practice test, the method with the strongest evidence behind it.
- Ask one manager, not the attendees. Did anything in their team’s weekly output change. Attendees report intention; managers report behavior.
A session that passes those three has produced capability. One with positive surveys and no artifacts has produced an afternoon — the distinction behind treating enablement as capability rather than license deployment.
Frequently asked questions
What exercises should a corporate AI workshop include?
At minimum a document loop on each person’s own recurring deliverable, a verification drill against a draft with planted errors, and a commitment round. Add a teardown if the session has to produce a decision about what to try next.
How much of a session should be hands-on practice?
More than half, measured in minutes where participants are typing. Published guidance converges above 50% and the training-dose evidence argues for the high end. An agenda under half is a presentation with activities attached.
Should exercises use real company documents or sample data?
Real documents, redacted against one written rule by one person. Clean samples produce clean first attempts, which removes the correction loop that is the lesson. Where redaction is unsafe, substitute a structurally identical document from a closed matter.
Are AI games and scavenger hunts worth the time?
Rarely. A scavenger hunt measures how fast people find features, which is nobody’s job. If you want competition, award it for a saved prompt another team can run on their own document.
How many exercises fit in a half-day session?
Four or five. The document loop needs 35 to 45 minutes and should not be compressed; a drill, a teardown and a commitment round fit around it.
Does a beginner session need a failure or verification exercise?
Yes, more than a room of daily users does. The first wrong answer someone meets alone is usually the reason they stop. Twenty minutes catching planted errors with an expert present turns that into an expected step.
How do you run exercises in a room with very different skill levels?
Pair deliberately with asymmetric roles: the experienced person drives and narrates the first attempt, then hands over the keyboard and only answers questions. Assign pairs from the intake, or the confident users will group themselves.
How do you tell afterwards whether the exercises worked?
Count the artifacts in a shared place after two days, re-run one exercise at 30 days to see whether people reach a usable output in fewer rounds, and ask a manager whether the team’s output changed.
Can a vendor build these exercises from a standard catalog?
No, and the arithmetic says so. One published play prices facilitator preparation at two to three hours for a one-hour session. A per-seat course cannot absorb that per client, which is why its exercises run on supplied scenarios.
Key takeaways
- An exercise earns its minutes only if the participant does the typing, works on their own material, leaves with an artifact, and sees failure happen while an expert is present.
- The document loop is the one exercise no session should omit: three to five correction rounds against a recurring deliverable, ending in a saved prompt.
- Verification deserves its own exercise. Planted errors and a written three-check routine beat a governance slide.
- Score the teardown on frequency and cost of error, or it yields a wish list instead of a shortlist a manager can approve.
- Count the typing minutes in any proposal. Under half means a presentation with activities attached.
- Scavenger hunts, stump-the-AI segments and build sprints leave nothing behind. Cut them before cutting practice.
Pick the document three people on your team produce every week and ask whether a session could end with a saved prompt that reproduces it. If the answer is yes, you have the first exercise and most of the agenda. A free AI-readiness call will tell you which function has the best candidate document and what the rest of the session should ask.
Arthur Wandzel
