Jonah Bergström is a senior HR business partner at a manufacturing company with about 600 employees. His VP of Operations asked him to run a series of AI workshops after the company bought a suite of AI tools that no one was using three months after deployment. Jonah had attended AI workshops. He'd even liked some of them. But he had never designed and run one for adults who were, at best, skeptical and, at worst, actively afraid that the tools were there to replace them. His first session had 22 people in the room and felt, he said, "like presenting a tax audit to people who didn't know they were being audited." By the fifth session, he had a design that actually worked. This lesson captures what he learned.

Running a good AI workshop is not primarily a matter of knowing more about AI than the people in the room. It is a matter of designing a session that respects how working adults actually acquire a skill: on their own tasks, under light supervision, with enough repetition to build confidence before they leave. Almost every failure mode below is a violation of that principle dressed up as something else.

Why AI Workshops Fail

Before designing one that works, it is worth understanding the failure modes, because they are consistent and they are all avoidable. Each one feels reasonable at design time. That is precisely why they survive into so many sessions.

Too much concept, too little doing. Adults in a professional context learn skills through practice, not lecture. A 60-minute presentation on how large language models work, followed by 10 minutes of "now you try it," inverts the effective ratio. Participants walk out understanding AI in the abstract and incapable of using it in their actual work. The session feels successful from the front of the room, because the material was covered and nobody looked confused, and the covering is exactly the problem.

Generic examples that don't map to participants' work. Showing a workshop audience how AI can write a marketing email when their jobs are in operations, finance, or customer service creates a gap they cannot bridge on their own. The facilitator sees a clear demonstration; the participants see a demonstration of somebody else's job. If the examples don't match participants' real tasks, the skill transfer rate approaches zero, and no amount of enthusiasm in the delivery closes that gap.

Fear not addressed explicitly. In any room of working professionals learning about AI tools, a meaningful fraction are worried about job security. If you don't name and address that concern early, it does not go away; it sits underneath the session as background anxiety that blocks learning for the rest of the time you have. People who are calculating whether this tool is going to replace them are not simultaneously learning to prompt it.

No follow-through mechanism. A workshop that ends without a clear "here's what you do Monday morning," and without a system to follow up, has a shelf life of about 72 hours. Most participants leave with genuine good intentions, and those intentions evaporate under the pressure of their normal workload. The absence of follow-through is what turns a well-received session into no measurable change in behavior.

The Design Principles That Work

Start with the Audience's Actual Pain

Before designing the workshop, Jonah started doing 15-minute interviews with a cross-section of participants, six conversations in total, asking a single question: "What tasks in your job take the most time but feel like they shouldn't?" That phrasing does a lot of work. It surfaces the specific friction points where AI tools are most likely to create real value, and it does so without asking people to speculate about AI, which they are not well positioned to do.

Those answers become the examples for the workshop. In Jonah's manufacturing context, the top answers were writing shift-handover reports, summarizing maintenance logs to brief the operations manager, and answering repetitive employee questions from an HR policy document. His fifth session was built entirely around those three tasks, using real documents from the business rather than sanitized samples. Completion rates jumped from 60% of participants trying the tool during the session to 95%. The content of the session was not more sophisticated than before. It was simply about their work.

Flip the Ratio: 70% Doing, 30% Explaining

A 90-minute workshop should contain at most 25 to 30 minutes of explanation, demonstration, and Q&A. The remaining 60-plus minutes should be participants using the tool on tasks they actually care about. This ratio is uncomfortable for facilitators who prepared well, because it means leaving prepared material unsaid, and it is the single highest-leverage design decision in the session.

Structure the time as alternating short bursts rather than one block of teaching followed by one block of practice: 5 minutes of instruction, 10 minutes of independent practice, 5 minutes of debrief, then repeat. This cadence keeps attention high and, more importantly, ensures that every concept is immediately applied before the next one is introduced. Participants never accumulate a backlog of unpracticed ideas, and the debrief gives you a live read on who is stuck while there is still time to help them.

Address Fear in the First Ten Minutes

Jonah's breakthrough came when he started every session with a direct acknowledgment. Something close to: "Some of you may be wondering if these tools are here to replace your jobs. That's a fair question and I want to address it directly before we start." He then spent five minutes on what the data actually shows about AI and job displacement in their specific context, not minimizing the concern, but providing honest information about which tasks AI handles well and which require the human judgment their roles depend on.

After this addition, he saw a measurable shift in room energy. People who had come in with crossed arms began participating within the first 20 minutes. One participant told him afterward: "I felt like you weren't trying to sell me something." That sentence is the whole mechanism. The credibility of everything else you say in a workshop depends on whether you were willing to say the uncomfortable thing first.

Use Graduated Challenge Levels

Participants in any workshop have widely varying starting skill levels, and a single practice exercise will be trivial for some people and impossible for others. Design three tiers of practice exercise instead, and let participants move between them at their own pace rather than assigning tiers in advance.

This structure prevents the two common failure modes that otherwise occur simultaneously in the same room: fast learners disengaging because the pace is too slow, and slower learners falling behind and disengaging because they cannot keep up. A tiered design also gives you something useful to say to both groups, which a single-track exercise does not.

Logistics That Actually Matter

Most of the logistics discussion around workshops is about rooms and catering. The three below are the ones that determine whether the session teaches anything.

Account access before the session. If participants need to set up an account, approve terms of service, or install software, do it before the workshop rather than during it. A 90-minute session that spends 25 minutes on access troubleshooting achieves very little, and the cost is worse than the arithmetic suggests, because the people whose access fails are usually the least confident people in the room and they spend the rest of the session behind. Send instructions in advance and have a colleague available for technical support in the 15 minutes before the session starts.

Group size: 12 to 20 is the sweet spot. Fewer than 12 and the energy drops, because there are not enough people generating different results to compare, and comparison across participants is a large part of what makes the practice segments work. More than 20 and you cannot give enough individual attention to participants who are stuck or struggling, which means the people who most need help are the ones who do not get it. For larger organizations, run multiple smaller sessions rather than one large one, even though that costs more facilitator time.

Physical versus virtual. In-person workshops consistently produce better engagement for AI skill building. Participants can see each other's screens, ask questions by turning to a neighbor rather than by interrupting the whole room, and the social learning from watching someone else figure something out happens immediately and without anyone having to organize it. If you must run virtually, use breakout rooms of 3 or 4 people during practice exercises to replicate some of that dynamic, and accept that you will need to be more deliberate about surfacing who is stuck.

The Monday Morning Commitment

The last ten minutes of any AI workshop should be structured around a specific commitment rather than around summary or thanks. Ask each participant to write down one task they will attempt using AI in the next five working days. Make it concrete: not "use AI more" but "use AI to draft the Thursday operations briefing." The specificity is what makes the commitment survive contact with a normal week, because a vague intention has no trigger and a named task does.

Two weeks after the workshop, send a one-question follow-up: "Did you try the task you committed to? If yes, what happened? If no, what got in the way?" This does two jobs at once. It signals that the commitment was real and someone was going to ask, and it gives you data to improve the next workshop, particularly from the people who did not follow through and can tell you exactly what blocked them. Jonah found that participants who made a written commitment applied the skill within five days at twice the rate of those who did not.

Anti-Patterns

These are the design choices that feel professional and reliably produce sessions nobody applies afterward.

• Front-loading the theory. Explaining how large language models work before anyone has touched the tool spends the room's attention on the least actionable material. Concepts land far better when they explain something a participant has just experienced.
• Demonstrating on your own examples. A demonstration built around a task the facilitator understands well is a demonstration of the facilitator's job. Build every example from the interview answers instead.
• Skipping the fear conversation because the room seems fine. Job security anxiety is rarely voiced and almost always present. A room that seems fine is often a room that has decided not to say anything.
• Running one difficulty level for everyone. A single exercise pitched at the middle loses the fast learners to boredom and the slower learners to embarrassment, and both groups disengage quietly.
• Handling account setup live. Troubleshooting access during the session burns the practice time that the session exists for, and it burns it on the participants who can least afford to fall behind.
• Ending with a summary instead of a commitment. A recap produces agreement in the room and nothing on Monday. A written, specific, dated commitment is the only ending with a measurable effect.

Practice Prompts

Use a workshop you are actually scheduled to run, or one you would run if asked. These exercises produce the raw material for the session design rather than describing it.

• Run the six interviews. Book 15 minutes each with six people who will attend, and ask only the core question: "What tasks in your job take the most time but feel like they shouldn't?" Write the answers down verbatim.
• Pick three tasks. From those answers, choose the three tasks that appear most often and that AI tools plausibly help with. Every example in your session should now come from that list, using real documents from the business.
• Build the time grid. Lay out a 90-minute session as alternating blocks of 5 minutes instruction, 10 minutes practice, 5 minutes debrief. Add up your explanation time. If it exceeds 30 minutes, cut material until it does not.
• Write the fear paragraph. Draft the opening acknowledgment in your own words, plus the five minutes of honest context about which tasks AI handles well and which need human judgment in your specific setting. Read it aloud. If it sounds like reassurance rather than information, rewrite it.
• Design one exercise at three tiers. Take a single task and write a foundational, an applied, and a challenge version of it. The foundational version must be impossible to fail.
• Draft the commitment and the follow-up. Write the closing instruction and the one-question message you will send two weeks later, before the session runs rather than after.

Reflection

Think about the last training session you attended, on any subject, that actually changed how you worked. What proportion of it was you doing the thing rather than watching someone else do it? Now think about a session you enjoyed but never applied, and ask what was missing at the end of it. Most people find that the difference was not the quality of the content but whether they left with a specific, dated intention and something they had already done once successfully. Then consider your own audience honestly: what is the unspoken concern in the room that you have been planning to work around rather than name? Whatever it is, addressing it directly in the first ten minutes will cost you less time than working around it for ninety.

Glossary

• Practice-to-explanation ratio: The proportion of workshop time participants spend using the tool versus listening. A working target is 70% doing and 30% explaining.
• Burst structure: Alternating short blocks of instruction, independent practice, and debrief, so that each concept is applied before the next is introduced.
• Graduated challenge levels: Foundational, applied, and challenge tiers of practice exercise, designed so that participants of different starting skill levels all have appropriate work.
• Foundational tier: A straightforward task with a model prompt supplied, which participants modify and observe. Success is guaranteed by design; the purpose is familiarity.
• Applied tier: A real task from the participant's own work, using a prompt structure they write themselves with light guidance.
• Challenge tier: An open-ended problem where participants design the prompt approach from scratch, for those who move quickly through the earlier tiers.
• Monday morning commitment: A written, specific task each participant undertakes to attempt with AI within the next five working days, captured in the last ten minutes of the session.
• Follow-through mechanism: The system that carries intent out of the room, typically a written commitment plus a short follow-up message two weeks later.

Related Lessons

Workshop facilitation is one delivery mode inside a wider capability effort. Creating Training Materials covers the artifacts that support a session and the reference material participants take away from it. Building Learning Infrastructure addresses what has to exist around individual sessions so that skills persist rather than decaying after the follow-up message. Measuring Training ROI and Skill Development takes the commitment and follow-up data from this lesson and turns it into evidence a sponsor will accept, and Sustaining AI Adoption Beyond Initial Enthusiasm deals with the longer arc after the workshop series ends.

Closing

The distance between Jonah's first session and his fifth was not expertise. He did not know appreciably more about AI by the end of the series. What changed was that he stopped designing a presentation and started designing a practice session: built on six conversations about real work, weighted heavily toward participants doing rather than watching, opened with the concern everyone was already carrying, tiered so that nobody was bored or lost, and closed with a written commitment somebody was going to follow up on. Those are not sophisticated moves. They are the ones most workshops skip, and they are the reason a room of skeptical people walked out able to do something they could not do when they walked in.

Key Takeaways

• Interview participants before designing the workshop. Six 15-minute conversations asking "what tasks take too much time?" give you the specific examples that make the session immediately relevant to the people in the room.
• Flip the ratio to 70% practice. Adults learn skills through doing, not listening; a 90-minute workshop should contain no more than 30 minutes of explanation, demonstration, and Q&A.
• Use a burst structure, alternating 5 minutes of instruction, 10 minutes of practice, and 5 minutes of debrief, so every concept is applied before the next arrives.
• Address fear explicitly in the first ten minutes. Job security anxiety is in the room whether you name it or not; naming it directly and providing honest context converts resistant participants into engaged ones.
• Design graduated challenge levels. Foundational, applied, and challenge tiers prevent both the fast learner disengagement problem and the slower learner dropout problem simultaneously.
• Solve logistics before the session, not during. Account access, software installation, and technical setup should be completed in advance; troubleshooting during the session destroys momentum and goodwill.
• End with a written commitment to a specific next action. Participants who commit to one concrete task in the next five days apply their learning at twice the rate of those who leave with only general intent.

Frequently Asked Questions

How long should an AI workshop be? The design in this lesson is built around 90 minutes, which is long enough to include several practice bursts and short enough to hold attention. What matters more than the total is the split: no more than 25 to 30 minutes of explanation, demonstration, and Q&A, with the rest spent on participants using the tool on their own tasks.

What if I cannot interview participants in advance? Six 15-minute conversations is a modest investment and worth protecting, but if it is genuinely impossible, ask the same question at the start of the session and build the practice segments around the answers you get. The cost is that you cannot prepare real documents in advance, which is a meaningful loss.

How do I address job security fears honestly if I don't have data? Then say so, and talk about what you do know: which tasks in their specific roles the tools handle well, and which require the human judgment those roles depend on. The credibility comes from not minimizing the concern, not from having statistics. Participants can tell the difference between honest uncertainty and a sales pitch.

Does this design work virtually? It works less well. In-person sessions consistently produce better engagement for AI skill building, because participants can see each other's screens and learn from watching someone nearby solve a problem. Virtually, use breakout rooms of 3 or 4 people during practice exercises and expect to work harder at noticing who is stuck.

What do I do if adoption still does not move after the workshops? Look at the follow-up responses from people who did not attempt their committed task, since they will tell you what actually blocked them. The most common answers are workload, access problems that were never fully resolved, and an example set that did not match their real work closely enough to transfer.