AI training theater is a session built to be reported rather than used. It produces a full room, a satisfaction score, a completion percentage, and a congratulatory email. Four weeks later nobody works differently, and the numbers on the slide still look fine, because none of them were measuring the work.
The term borrows from security theater, where a visible measure stands in for an effective one. The structure is the same. Attendance is an input that reliably goes up. Capability is an outcome that does not move unless something after the session moves it. Companies buy the first and report it as the second, usually without anyone meaning to mislead.
What AI training theater looks like
It rarely looks bad in the room, which is what makes it hard to catch.
A trainer arrives with a polished deck and clean examples: write a haiku, summarize a press release, plan a trip to Lisbon. People leave saying it was better than expected, the feedback form averages 4.6 out of 5, and the sponsor forwards that score to the CFO. Then the CFO asks what changed, and the answer is a number of seats.
Four tells, all visible from outside the room.
The exercises use the vendor’s material, not yours. A demo document cannot teach anyone to handle the mess in your intake queue, and the mess is the hard part.
Every reported metric is same-day. Completion, attendance, satisfaction: each knowable before people reach the parking lot, none of them surviving contact with a Tuesday.
Nothing in writing leaves the room. No procedure, no prompt set tied to a named task, no baseline to compare against later.
Nobody owns Monday. No named person, no review date, no second touch. The session was the deliverable.
The pattern predates AI. Training Industry’s February 2026 piece When Corporate Training Becomes Theater puts the diagnosis plainly: visibility beats effectiveness, and learning teams get judged on what is easy to count. What is new is the pressure. Companies have just bought AI licenses for everyone and are being asked to show the spend worked, which is exactly when a legible number beats a true one.
Why the theatrical numbers keep winning
Nobody sets out to buy theater. Five forces push toward it, each rational for the person deciding.
| Force | Why it favors attendance | What it costs |
|---|---|---|
| Legibility | A completion dashboard is one clean number; behavior change needs a baseline and 90 days | The easy number crowds out the real one |
| Defensibility | A roster is an audit trail that satisfies legal and risk whatever happened to capability | Documentation becomes the goal |
| Procurement shape | Vendors are paid on seats, so the success metric is fixed before anyone asks about outcomes | The invoice clears on the wrong event |
| Timing | Attendance is measurable the same afternoon; capability is invisible for weeks | Nobody is watching when the truth arrives |
| Incentives | The team that measures waste reports its own failure rate | The gap stays unmeasured on purpose |
The procurement line deserves a moment. If a vendor is paid for a day and graded on a feedback form, both sides have agreed to be judged on the easy part. Change what you measure and you change what you buy, which is the working form of treating licenses as a purchase and capability as a build.
The statistic we are not going to repeat
Nearly every article on this subject reaches for the same line: only 10% of training ever transfers to the job. It is a good line. It is also unsupported, and the field that popularized it has said so.
The figure traces to a 1982 conference paper, quoted secondhand in Baldwin and Ford’s 1988 review of transfer research. Ford, Yelon and Billington returned to it in 2011 in Performance Improvement Quarterly and found no empirical behavioral evidence behind the number. Their name for it, the ten percent delusion, is the honest summary. It survives because repeating a citation is cheaper than checking one.
We flag it because a page asking you to be skeptical of vendors has no business running on a number it did not verify. When a proposal quotes a dramatic adoption statistic, ask for the study. A vendor who cannot produce it has shown you how they handle evidence.
What the research does establish is narrower and more useful: transfer depends on the environment the learner returns to. Blume and colleagues’ 2010 meta-analysis in the Journal of Management, covering 89 studies, found support for predictors including motivation and a supportive work environment. What the organization does afterward outweighs what happened in the room.
What the evidence supports, and where it stops
Three findings worth building on, each with its limit stated.
Dose matters. BCG’s AI at Work 2025 study, published 26 June 2025 across more than 10,600 workers in 11 countries, found only 36% of employees feel adequately trained in AI use while 72% use it regularly. Among frontline staff regular use drops to 51%, and just 25% say leaders give enough guidance. Workers getting five or more hours of training, in person and with coaching, were much more likely to become regular users. The caveat nobody else prints: that finding is correlational and cannot separate training hours from the kind of employer who funds five of them.
Access is not the constraint. Gallup’s June 2025 analysis of AI use at work found 40% of US employees use AI at least a few times a year, up from 21% two years earlier. Weekly use is 19%, daily use 8%. The number that should stop a sponsor is a different one: only 16% of AI users strongly agree the AI tools their employer provides are useful for their work. Only 22% say a clear AI plan has been communicated. People have the tools. They lack the application.
Leaders measure from the wrong seat. McKinsey’s January 2025 research found C-suite leaders estimate 4% of employees use AI for at least 30% of their daily work, while 13% of employees report that level. Executives are about three times wrong on the baseline, and a program measured against a wrong baseline proves nothing either way.
The four levels most companies stop at
The vocabulary has existed since the 1950s. The Kirkpatrick model sorts evaluation into reaction, learning, behavior and results. Most programs report the first two and never reach the third.
| Level | The question | Theatrical answer | Real answer |
|---|---|---|---|
| 1. Reaction | Did they like it? | 4.6 out of 5 | Useful only as a warning when low |
| 2. Learning | Did they learn? | 94% completion | A task done unaided after the session |
| 3. Behavior | Are they doing it? | Not measured | Share of trained staff using AI on a named task at day 60 |
| 4. Results | Did the work change? | Not measured | Cycle time or rework on that task, against a baseline |
The trap is that the levels are not in tension. A session can score 4.6 on reaction, hit 94% completion, and produce no level-three change at once. Those numbers are not contradictory, they are unrelated. A room full of happy people tells you the facilitator was likable.
Level three is cheap to measure once you accept it has to wait. Pick one task per team, time it in the room, then ask at day 30 and day 60 whether the written procedure is still in use and what was tried that failed. The second question surfaces more than any dashboard.
The proposal tells you before the session does
You do not need to run the session to find out. Almost every tell sits in the document you were sent, where an AI workshop for teams either commits to an output or avoids doing so.
| In the proposal | What it predicts |
|---|---|
| Full agenda arrives before any discovery call | The content is fixed and was written for somebody else |
| Deliverables are a deck and a recording | Nothing operational leaves the room |
| Exercises use sample documents | Practice happens on the easy version of your work |
| Per-head pricing, no follow-up line | Vendor revenue scales with the roster, not the result |
| Success is “participant satisfaction” | You have agreed to be graded on applause |
| No named owner or review date | Follow-through is nobody’s job once the invoice clears |
| Preparation hours unstated | Usually near zero, which is the generic deck problem |
| Cohorts above roughly 20 people | Individual correction is impossible, so it is a lecture |
Four questions to send back before any contract. Each is cheap to ask and expensive to fake.
- How many hours will you spend on our documents before arriving, and which do you need? A vendor who wants nothing in advance has nothing specific planned.
- What is in our hands on Friday afternoon, in writing? Procedures and a task-labeled prompt set, or a deck. The answer sorts the market.
- What gets measured at day 60, and who takes the baseline? If the only measurement is same-day, the engagement ends before the evidence arrives.
- What is the follow-up, and is it priced? Market rates for hands-on AI training run roughly $2,000 to $15,000 by length and preparation; a follow-up cadence costs a fraction of that and does most of the work.
A vendor who answers all four plainly may still be wrong for you. One who deflects all four has told you what you are buying.
The rule that quietly pays for theater
There is a structural reason attendance records became the artifact of choice. The EU AI Act’s Article 4 AI literacy duty has applied since 2 February 2025 and requires providers and deployers to take measures ensuring a sufficient level of AI literacy among staff. It also states that it does not require any guaranteed level of literacy in any individual.
That is a sensible way to write a law, and it creates an obvious incentive: the cheapest way to show you took measures is a roster with signatures. Compliance rewards documentation, capability rewards practice, and one session can serve both only if somebody writes the capability requirement into the brief, because the rule will not.
If your AI training exists to satisfy a policy, say so and buy the cheap version. The error is spending capability money on a compliance artifact.
What a real session leaves behind
The difference is not energy or production value. It is what exists on Friday at 5pm that did not exist on Monday.
- Written procedures for the workflows you changed, in your own language, specific enough for someone who missed the session to follow.
- A prompt set labeled by task and owner, six to twelve items, not a library nobody opens.
- A baseline for the tasks the session touched, taken in the room. Without it, nothing later can be compared.
- A one-page safe-use sheet covering what goes into an assistant and what does not.
- A named owner and a date. One person, one review, 30 days out.
None of it is exotic, and all of it is testable in advance. How the pieces fit a running order is covered in the playbook for sessions that change Monday morning, and keeping a mixed room engaged long enough to produce them is the subject of the skeptic in row three.
You already ran one that did nothing
Most companies reading this have already paid for a session that went nowhere. That is recoverable, and cheaper than the first attempt, because people have now seen the tools and hold an opinion about them.
Pick one task, not a program: one recurring piece of work, one team, something someone complains about monthly. Breadth is what killed the first attempt. Take the baseline now: how long that task takes, how often it gets reworked. Run a half day on that task alone, with live documents, hands on keyboards, written procedure out at the end. Book the review before anyone leaves: day 30, one owner, two questions.
The honest limit: one session changes two or three workflows for the people in the room. It does not change an organization, and a vendor promising otherwise is selling the next round of theater.
Frequently asked questions
What is AI training theater?
A corporate AI workshop designed to be reported rather than used. It yields a full room, a high satisfaction score, and a completion percentage, none of which measure whether anyone works differently afterward. The name follows security theater: a visible measure standing in for an effective one.
How can I tell before signing whether a session will be theater?
Read the deliverables and the preparation. A fixed agenda sent before any discovery call, exercises built on sample documents, per-head pricing with no follow-up, and a deck as the only output all point one way. Then ask how many hours the facilitator will spend on your documents beforehand. Near zero is the strongest predictor.
Isn’t attendance a reasonable first measure?
As a check that the logistics worked, yes. As evidence of capability, no. Attendance and satisfaction sit at the first two Kirkpatrick levels; behavior and results sit at levels three and four. A program can post excellent level-one numbers and produce no level-three change, because the two are unrelated.
How soon after a session should behavior change show up?
Check at day 30 and again at day 60. Earlier measures novelty; later and the memory of what people tried has gone cold. The useful figure is the share of participants still using a written procedure on a named task.
How many hours of AI enablement does a team actually need?
BCG’s 2025 workforce study found employees receiving five or more hours of training, in person and with coaching, were much more likely to become regular AI users. Treat five hours as a floor worth clearing rather than a guarantee: the finding is correlational and cannot separate training hours from the employer who funds them.
What should a workshop leave behind?
Written procedures for the workflows you changed, a prompt set labeled by task and owner, a baseline taken in the room, a one-page safe-use sheet, and one named person with a 30-day review booked. A deck and a certificate only prove a day happened.
Who is responsible for follow-through, the vendor or us?
Both, and the split belongs in writing. The vendor owns the procedures, the baseline method, and any priced follow-up. You own the internal owner, the review date, and whether managers ask about the new behavior. No vendor can supply that last one, and it decides whether any of it holds.
Does a compliance requirement count as AI enablement?
No, though one session can serve both if the brief says so. The EU AI Act’s Article 4 literacy duty, applicable since February 2025, asks for measures taken and does not require a guaranteed level of literacy in any individual. That standard is satisfiable with a roster. Budget the two separately or the cheap one absorbs the other.
Key takeaways
- AI training theater is a session built to be reported, not used. Attendance always rises; capability moves only when something after the session moves it.
- The theatrical metrics win because they are legible, defensible, procurement-friendly, available the same afternoon, and graded by the team whose program they describe.
- Skip the famous “only 10% of training transfers” line. It traces to a 1982 paper and was found unsupported in 2011. A buyer who checks numbers should start there.
- What the primary data supports: 36% of workers feel adequately trained (BCG, 2025), 16% of AI users strongly agree their employer’s AI tools are useful for their work (Gallup, 2025), and executives underestimate real usage by about three times (McKinsey, 2025).
- The proposal predicts the outcome. Fixed agenda before discovery, sample-document exercises, per-head pricing with no follow-up, and satisfaction as the success metric are the reliable tells.
- Five artifacts separate capability from theater: written procedures, a task-labeled prompt set, a baseline taken in the room, a safe-use sheet, and one named owner with a 30-day date.
Not sure whether the session you are being quoted will change anything? We run a free AI-readiness call: a working conversation about which of your recurring tasks a hands-on session could move, and what to measure at day 60. Book a time.
Arthur Wandzel