A mixed-ability AI workshop fails when the facilitator picks a single pace, and the person it fails first is your best user. Cognitive-load research has a name for this: guidance a beginner needs is dead weight to someone who already holds the skill. A middle pace does not split the difference between the two ends of the room. It taxes one end to subsidize the other.
The fix is not a faster or slower session. It is to differentiate the task and keep the explanation shared, because one room can absorb one instruction and twelve pieces of work at once. Here is how to size the spread you have, where the standard playbook stops short, and a running order that holds both ends of it.
The room, measured
Most AI enablement planning treats the spread as a vague worry. It is published data. Gallup’s workplace AI indicator, as of May 2026, puts US employee use at 15% daily, 30% a few times a week or more, and 52% a few times a year or more. Gallup prints no non-user figure, but the arithmetic leaves roughly 48% who use AI rarely or never. Apply that to twelve people from across a company:
| Starting frequency | Share of US employees | People in a room of 12 |
|---|---|---|
| Daily | 15% | About 2 |
| A few times a week | 15% | About 2 |
| Monthly or a few times a year | 22% | About 3 |
| Rarely or never | About 48% | About 5 |
Two people in that room reach for an assistant every working day. Five have barely opened one. That is a difference of hundreds of attempts against a handful, not a half-step of pacing.
A second spread crosses the first. Pew Research Center’s American Trends Panel, fielded June 22–28 2026 with 3,488 US adults, found 52% more concerned than excited about rising AI use, against 9% more excited. That is a majority position, and some of those people are in the room because a manager put them there. One Gallup figure explains the mood: 25% of employees say their organization has communicated a clear AI plan, while 47% say it has adopted the tools.
Skill and willingness are two different spreads
Published guidance on mixed-ability AI training conflates the two. It reads a quiet person as a beginner and an objection as a knowledge gap. Cross the axes and four people appear, each needing something different from the same two hours.
| Low skill | High skill | |
|---|---|---|
| Willing | The recruit. Needs one win on real work before lunch, and permission to be slow. | The quiet power user. Needs a harder problem than the one on the slide. |
| Reluctant | The conscript. Needs an answer to “why me” before any tool appears. | The skilled skeptic. Holds a specific, often correct objection. Most expensive to get wrong. |
The bottom row is what breaks sessions. Neither reluctant attendee can be managed by pace, because slow content was never their problem, and neither can be managed by enthusiasm.
Your intake should produce this grid, not the single confidence score most vendors collect. Two questions do it: how often do you use an assistant for work, and what is the one thing you want it to stop getting wrong. The second answer separates the conscript from the skeptic in a line, and it belongs in the pre-work that changes the session. A facilitator who never asks is planning to teach the middle.
Why a middle pace costs more than it looks
Kalyuga, Ayres, Chandler and Sweller published “The expertise reversal effect” in Educational Psychologist (38(1), 23–31, 2003), reviewing studies that tested instructional methods against learners with different prior knowledge. The finding: techniques that work well for inexperienced learners lose their effect, and can produce worse results, with experienced learners.
The mechanism is plain. A novice has no stored pattern for the task, so step-by-step guidance builds one. Someone who holds that pattern must also reconcile the guidance against what they know, and the reconciling consumes the working memory the task needs. Support is not neutral for them. It is interference.
So a middle pace is a transfer rather than a compromise: the beginner gets partial scaffolding, the advanced user gets measurable interference, and the form reports a reasonable average. The thing to vary is the task. One explanation plus two difficulties of work costs nothing and puts each person at their own edge, which is why a session should be built around what people are asked to do rather than what they are shown.
What the standard playbook does, and where it stops
The most widely used corporate structure is Atlassian’s AI Training Workshop play, worth reading before you write your own. It is honest about its audience, aimed at beginners in less technical roles across functions like marketing, HR, sales and finance. It budgets two to three hours of prep for a one-hour run and asks attendees “to complete a basic survey to gauge their familiarity and confidence with the AI tool(s).” It then walks use cases in rising difficulty, “one basic, one intermediate, and/or one more advanced,” with a show-then-do pattern and a closing survey.
For a beginner audience that is a good hour. Three things in it stop short of a genuinely mixed room.
The survey does not branch. Familiarity is measured at the door and the agenda does not change. A diagnostic that cannot alter the session is a satisfaction baseline.
Rising difficulty is sequential, so somebody waits. Run the basic use case first and the two daily users spend the opening block watching something they can already do, in the segment where attention is highest.
The advanced attendees are recruited as staff. The play’s first step is finding champions, described as AI Super Users, Facilitators or AI Liaisons. Enlisting your strongest users is right. Asking them to help run a session with nothing in it for them is how an advocate becomes a cynic.
Published approaches elsewhere all handle the spread by splitting it: self-selected tracks after a shared opening, a second facilitator per level, or separate beginner and advanced products. None says what to do with one budget line, one day on the calendar, and no option to split, which is the position most companies of 50 to 500 people are in.
The one-session running order
This shape holds a spread in one room with one facilitator. It assumes twelve to sixteen people, their own documents, and the intake above. Timings are for two hours and stretch for a half day.
| Block | Minutes | What happens | Why it works at both ends |
|---|---|---|---|
| Frame | 0–10 | Why this team, this quarter, and what the day will not cover | Answers the conscript before any tool appears |
| Failure demo | 10–25 | Facilitator runs a real request badly, then repairs it live | Beginners see the method, advanced users the diagnosis |
| Floor and ceiling | 25–45 | One task, two finish lines: defined output plus named extension | Nobody waits, nobody is abandoned |
| Own work, paired | 45–85 | Each person works a real task in assigned pairs | The spread becomes teaching capacity |
| Read-out | 85–105 | Three outputs read aloud, one per level, with faults named | Correcting strong work in public is where a power user learns |
| Commit | 105–120 | Each person writes the task they will do this way next week | Differentiated by definition |
Three details carry most of the weight.
Run the failure demo before the success demo. A clean prompt that works is a slide. A request that returns something confidently wrong, followed by the facilitator saying why and fixing it, is legible to a beginner and new to an advanced user, because diagnosing a bad output is the skill nobody picks up from daily use. It also puts the session’s first honest word in front of the skeptic.
Give every task a floor and a ceiling. The floor is the output everyone produces. The ceiling is a harder version for whoever finishes in half the time: chain two steps, work a messier document, get the same result without naming the format. Five minutes of prep removes the most common complaint from experienced attendees.
Pair with asymmetric roles. Assign pairs from the intake, or confident users will cluster. The experienced person drives the first attempt and narrates it, then hands over the keyboard and may only answer questions. Peer teaching has an evidence base with adults: Balta, Michinov, Balyimez and Ayaz’s meta-analysis covered 29 studies and 4,738 participants in International Journal of Educational Research (86, 2017) and found a positive effect against traditional teaching. The narration constraint is what makes it work, because it forces a tacit skill into words and surfaces the strong user’s own gaps.
Avoid generic practice material. A toy prompt collapses the room onto one axis of speed, and the fast finisher wins a meaningless race. Own backlogs are what let twelve people run at twelve difficulties with nobody managing the pace, which is the quiet reason to insist on training on last week’s real work.
The power-user bargain
Your two daily users set the tone in the first fifteen minutes. Engaged, they make the room read the technology as serious. Visibly bored, they give every reluctant attendee permission to disengage.
Make an explicit trade with them on the intake call. They bring their hardest current problem rather than a demonstration, and if the facilitator cannot solve it in the room, that is a finding worth the day. They take a teaching role in the paired block. They are told in advance that the opening is not aimed at them, which costs nothing and buys their attention for the read-out.
What they must not be is unpaid staff. Whether the session gave them something back decides whether an advocate or a cynic walks out, and the same mechanism decides whether capability survives the vendor leaving, as the workshop design playbook argues at length.
The skilled skeptic
The high-skill reluctant attendee is the one person who can end a session early, because the objection is usually specific and often right. They have watched a model invent a citation, or they own the system that a colleague’s AI-written summary described wrongly.
Handle the objection, not the person. Ask for the exact case and reproduce it live. Either the failure reproduces and the facilitator shows the verification step that catches it, which is the most useful ten minutes of the day, or it does not reproduce and the room hears the honest reason why. What fails is answering a correct objection with enthusiasm, which tells the room the facilitator cannot distinguish a bug from a mood.
The feedback form will hide the failure
A mixed room produces a bimodal happy sheet and an average erases it. Power users mark it low for being easy, reluctant attendees mark it high for being less threatening than expected, the mean lands near four out of five, and the session looks fine.
Split every score by the starting frequency you collected at intake. Two averages tell you which half of the room you served, and they usually disagree by more than a point. Then measure behavior three weeks out at each end: whether the recruits are doing the committed task the new way, and whether anything the power users do changed at all. If the advanced half reports no change, it was a beginner session with an audience problem, whatever the average said. That is the discipline behind our case that buying licenses is not the same as building capability.
When to stop trying and split the room
One room absorbs a wide spread, not every spread. Three thresholds:
- More than sixteen people. Past that the read-out does not fit, and the read-out is where a mixed room pays off. The arithmetic is in our rules on cohort size.
- Daily users above a third of the room. With five or six advanced attendees, floor-and-ceiling becomes two sessions running in parallel with one facilitator. Give them their own ninety minutes on a harder problem.
- A reluctant majority with no answer to “why us.” If leadership has not said what the tools are for, no session design fixes it, and Gallup’s 25% figure says the gap is common. Spend the first meeting on the plan, not the prompts.
Hands-on AI enablement for a team runs roughly $2,000 to $15,000 in the current market, and the split-versus-single decision moves that total more than almost any other choice, because the cost is facilitator days. Check any provider’s current group limits and pricing on their own pages, since corporate training catalogs change quarterly. For a second opinion on your own spread, a free AI-readiness call is where we read your intake numbers and say whether it is one room or two.
Frequently asked questions
How do you run an AI workshop with mixed ability levels?
Keep one shared explanation and differentiate the work. Give every task a floor, the output everyone produces, and a ceiling, a harder version for whoever finishes early. Have people work their own real documents, so twelve people run at twelve difficulties with nobody managing the pace.
Should beginners and advanced users be in the same AI workshop?
Up to about sixteen people, yes, if tasks are differentiated and the advanced attendees have a role. Split when daily users exceed a third of the room, or when the group is too large for the read-out.
How do I keep experienced users engaged in a beginner-level session?
Give them their own hardest current problem instead of the demonstration task, a teaching role in the paired block, and advance notice that the opening is not aimed at them. A bored power user sets the tone for every reluctant attendee.
What does research say about teaching novices and experts together?
Cognitive load theory addresses it directly. The expertise reversal effect (Kalyuga, Ayres, Chandler and Sweller, Educational Psychologist, 2003) found that guidance which helps inexperienced learners loses its effect, and can produce worse outcomes, for learners with more prior knowledge.
How mixed is a typical corporate AI workshop room?
Wider than most planning assumes. Gallup’s May 2026 figures put 15% of US employees using AI daily and 52% a few times a year or more, leaving roughly 48% who use it rarely or never. A twelve-person cross-section holds about two daily users and five with almost no hands-on time.
Is a skeptic the same as a beginner?
No, and treating them alike is the common error. A skilled skeptic usually carries a specific objection drawn from a real failure, and answering it with enthusiasm costs the facilitator the room.
How do you pair participants of different levels?
Assign pairs from the intake, since confident users cluster when left to self-select. The experienced person drives the first attempt and narrates it, then hands over the keyboard and only answers questions.
Will the feedback form tell us whether a mixed session worked?
Not as an average. Advanced users rate it low for being easy and reluctant attendees rate it high for being unthreatening, so the mean looks acceptable. Split the scores by starting frequency and check behavior three weeks later at each end.
Key takeaways
- At Gallup’s May 2026 rates, a twelve-person room holds about two daily AI users and about five people who use it rarely or never.
- Skill and willingness are separate axes. The reluctant pair, the conscript and the skilled skeptic, is what breaks sessions.
- A middle pace is not a compromise. Guidance that helps a novice is processing overhead for an expert, so the stronger half pays for the weaker half.
- Differentiate the task, not the explanation: one shared instruction, plus a floor and a ceiling on every exercise.
- Run the failure demonstration first. Diagnosing a bad output is the skill advanced users lack and beginners most need to see.
- Trade with your power users for their attention; do not recruit them as unpaid staff.
- Average feedback hides a bimodal result. Split every score by starting frequency, measure behavior change at each end, and split the room itself above sixteen people or when daily users exceed a third of it.
Collect two numbers before you book anything: how often each attendee uses an assistant, and what they want it to stop getting wrong. Those two columns decide whether you are running one session or two. A free AI-readiness call is where we read them with you and come back with a running order.
Arthur Wandzel
