Back to Guides
Enterprise Software 14 min read

Why most corporate AI workshops are forgettable

Why most corporate AI workshops are forgettable

Roughly half the reasons your AI training did nothing are not things the vendor did. They are decisions made in procurement and scheduling, weeks before anyone walked into the room: how many people you put in it, whether they all did the same job, and whether anyone owned the month afterward. The other half belong to the vendor. Only one of those two lists is yours to fix before you book again.

So start by being precise about what failed. A session is forgettable when it left nothing behind: no document anyone needed the following week, no habit anyone could name, no person who owned what happened next. People in the room may well have enjoyed it. The feedback form came back positive and the work came back unchanged.

The gap that makes a forgettable session likely

Three measurements, taken independently, describe the same hole.

Stanford HAI’s 2025 AI Index reports that 78% of organizations used AI in 2024, up from 55% the year before. Microsoft and LinkedIn’s 2024 Work Trend Index, published May 8 2024 from a survey of roughly 31,000 knowledge workers across 31 markets, found that 75% of them use AI at work and that 39% of those users had received any AI training from their employer. BCG’s From Potential to Profit, published January 15 2025, found that fewer than one-third of companies had upskilled even a quarter of their workforce, and that about one in four executives said their company had created significant value from AI.

Tools arrived first and in volume. Capability did not follow. The same BCG work puts a number on where effort should go: a 10-20-70 split, with 10% on algorithms, 20% on data and technology, and 70% on people, processes and cultural change.

Set a single half-day session against a 70% allocation and the arithmetic answers itself. The session was asked to carry the whole people-and-process side of an AI program on its own, once, with no structure around it. Most forgettable sessions are not badly delivered. They are correctly delivered into a system with nowhere to put them — which is the argument we make at length in the case that buying licenses is not the same as building capability.

What “forgettable” means, precisely

Vague complaints are hard to act on, so use a definition with three checks. A session was forgettable if, thirty days later:

  1. No artifact survives. Nobody is using a prompt, a template, a checklist or a reviewed document that came out of the room.
  2. No habit has a name. Ask three participants what they now do differently and you get a general endorsement of AI rather than a task.
  3. No owner exists. When you ask who is following up, the answer is a function rather than a person.

All three are observable without a survey, and none of them require measuring the one thing most reports measure. Satisfaction scores stay high through every failure mode below, because a session can be clear, well-paced and genuinely interesting while producing nothing transferable. That disconnect is why the question worth asking six months out is a different one, which we take up in what actually sticks six months after a session.

Hold the three checks in mind, because each failure mode below fails a specific one.

Three failure modes you were sold

1. The session taught the tool instead of the work

The clearest marker: participants leave able to define prompt engineering and unable to say what to run on the variance report sitting in their inbox. Content organized around how a model works produces understanding that has nowhere to land. Content organized around a task the team already does produces a changed task.

This is the most common finding in critiques of bad AI training, and the most commonly misdiagnosed. The problem is not that the material was too technical. A room of operations managers can follow an explanation of context windows. The problem is that nothing in the session required them to open their own files, which is the case made in training on last week’s real work rather than toy prompts. Fails check one: no artifact.

2. The deck was built before anyone asked about you

Catalog material gets delivered as if it were custom. The tell is available before you sign: a session built on your team’s work has to ask for your team’s work. If the pre-work note asks participants to bring a laptop and nothing else, the material cannot be about their documents, because the vendor has not seen any.

The cost of this one is larger than the fee, and the arithmetic is worked through in the real cost of a workshop built from a generic slide deck. Fails check one.

3. The room watched

A session that runs as 80% presentation and 20% questions is a talk. Talks are a legitimate format when labeled as briefings, and several formats on the market are briefings wearing a different name. Hands-on AI training means minutes on the clock in which participants type against their own material and someone corrects what comes back. Everything else is exposure.

Watching a competent facilitator produce a good output teaches you that good outputs are possible. It does not teach you to produce one. Fails check two: no nameable habit.

Three failure modes you bought

4. The room was too big to practice in

Sixty people in one session looks efficient on a calendar and is the format in which practice is impossible. One facilitator cannot see sixty screens, so nobody’s output gets corrected, so the session reverts to presentation regardless of what the agenda promised. Twelve to fifteen people is the band where a facilitator can reach everyone, and some providers publish a working figure of around ten participants per instructor. We have written out the arithmetic behind that band in why twelve beats forty.

Cohort size is a procurement decision. No vendor can deliver practice into a room you sized for presentation.

5. Nobody in the room had the same job

Mixed audiences fail quietly. A session pitched to span a skeptical finance manager and an engineer who has been using these tools daily for a year lands in the middle and is useful to neither. The engineer disengages in the first twenty minutes; the finance manager concludes the material was for technical staff.

Streaming by role, or at minimum by current fluency, is the fix, and it is something you specify when booking rather than something a facilitator can improvise. Four tracks for four groups is more work than one session for forty, and it is the difference between a program and an event. The teaching problem a single mixed room creates is examined in the mixed-room problem.

6. The thirty days after the session belonged to nobody

A session is a start condition, not a deliverable. If no named person owns the month that follows, nothing enforces the new behavior and the default reasserts itself. Not because people forgot, but because reverting costs nothing and nobody notices.

This item changes the outcome more than any other on the list, and it is the one most often left out of a scope of work entirely. What that follow-up structure needs to contain is set out in the follow-up architecture a good session depends on.

Failure mode Whose decision Fix point
Taught the tool, not the work Vendor Curriculum review before signing
Catalog deck sold as custom Vendor The pre-work request
The room watched Vendor Hands-on minutes in the agenda
Cohort too large Buyer Booking
Mixed job functions, one track Buyer Booking
No owner for the next 30 days Buyer Scope of work

Three of six are yours. That is the actionable half.

The symptom calendar: day 3, day 10, day 30

Each failure mode shows a different symptom, and they show up on a schedule. Checking on this cadence tells you which mode you hit, which is more useful than knowing the session underperformed.

Day 3. Ask two participants to show you something they produced with an assistant since the session. If they describe an intention rather than opening a file, the session taught the tool rather than the work. Day 3 is also when a too-large cohort reveals itself: in a room of sixty, most people never got a single piece of their own output looked at, and they will say so plainly if you ask.

Day 10. Look for reuse. Is anyone running the same prompt a second time on a new input? First use is curiosity; second use is a habit forming. Silence at day 10 usually points to content pitched at the wrong altitude or a room that only watched.

Day 30. Look for spread. Has anyone who was not in the room been shown something by someone who was? Spread is the only signal that distinguishes a capability from an enthusiasm, and its absence at thirty days almost always traces to the sixth failure mode rather than anything that happened during delivery.

One warning about the measurement itself. License dashboards and seat-activity reports will usually look fine through all three checkpoints, because opening an assistant is not the same as using it on work that matters. Activity is the easiest thing to count and the least informative.

What a session that sticks leaves behind

Invert the three checks and you get a specification you can hand to a vendor.

  • An artifact per participant. Something each person built on their own material, saved somewhere they will find it, and would be embarrassed to have produced by hand again. A reviewed prompt for a recurring report counts. A certificate does not.
  • One named habit per role. Not a tool, not a capability statement: a task someone now does differently, stated in the words they use for the task.
  • One owner and three dates. A person responsible for the following month, with checkpoints already in the calendar before the session happens.

None of these need a large budget. They need to be in the scope of work, and they are usually absent from it because nobody asked. What a session should produce, and what to require in writing, is set out in what an AI workshop should actually produce and in our playbook for designing a session that changes Monday morning.

What to change before you book the next one

Four changes, in order of effect:

  1. Name the owner first. Before scope, before vendor, before date. If no one will take the thirty days, a session is premature and the money is better held.
  2. Cut the cohort and raise the count. Three sessions of fifteen cost three times the facilitator hours of one session of forty-five. They are also the only arrangement in which anyone practices, which is why the cheaper option is the one that produced nothing.
  3. Require the pre-work request in the proposal. Read the exact words participants will be sent. A vendor who cannot say what they need from your team in advance is going to deliver a catalog session.
  4. Pick the tasks before the tools. Two or three recurring pieces of work, named by the people who do them. The question of which assistant to teach is downstream of that and smaller than it looks.

Market pricing for hands-on sessions of this kind runs roughly $2,000 to $15,000, and the spread inside that range has more to do with preparation on your material than with hours delivered. A quote at the bottom of the band with no pre-work request is not a bargain; it is a different product.

Frequently asked questions

What does “bad AI training” usually mean in practice?

A session that was competently delivered and produced no artifact, no named habit and no owner for the following month. It is rarely about delivery quality. The three strongest predictors are cohort size, whether participants worked on their own documents, and whether anyone owned the thirty days afterward.

How do I know whether the session was bad or whether my team was the problem?

Check at day 3 for output, day 10 for reuse and day 30 for spread. If nobody produced anything in the first three days, the content was wrong. If people produced once and stopped, the follow-up was missing. If the output never left the original group, nobody owned the rollout.

Why were the feedback scores high if the session failed?

Feedback measures the experience of the session. A clear, well-paced session that gave participants nothing to carry out of the room scores well and changes nothing. Measure output at day 3 instead and the picture separates.

Is a longer session the fix?

No. Doubling the hours on the same material roughly doubles the payroll sitting in the room and does not change whose work is on the screen. Length is not the variable that failed.

How large should a cohort be?

Twelve to fifteen people per facilitator for hands-on work; some providers publish a working figure of around ten per instructor. Above roughly twenty-five, practice stops being possible and the session becomes a presentation whatever the agenda says.

What should the pre-work email ask for?

Two or three real pieces of recurring work per participant: a report they write, a document they review, an email pattern they repeat. If it asks only for a laptop and a charged battery, the material was built before anyone looked at your team.

Can one session cover a whole company?

No, and attempting it produces the mixed-room failure. Different roles do different work, so they need different tasks on screen. Stream the group by role or by current fluency and run shorter sessions more often.

Who should own AI enablement if we have no one in the role?

Whoever can require the behavior, not whoever is most interested in the technology. In a company of 50 to 500 people this is usually a VP of Operations, a Chief of Staff or a Head of L&D, and it needs to be a named person with the next month in their calendar rather than a committee.

Does the choice of assistant matter?

Less than the choice of task. Teaching transferable skill on whichever assistant your team already has a license for outlasts teaching one product’s menus, and the products change faster than the skills do.

How soon should we re-run a session that failed?

Not until the owner, the cohort split and the task list exist. Re-running the same session into the same conditions produces the same result and costs the credibility you need to ask a third time.

Key takeaways

  • A forgettable session fails three observable checks at thirty days: no surviving artifact, no nameable habit, no owner.
  • Three of the six common failure modes are procurement and scheduling decisions, not delivery failures. Cohort size, room composition and ownership of the following month are all set before the vendor arrives.
  • The structural context: 78% of organizations used AI in 2024 (Stanford HAI), 39% of workers using AI at work had received employer training (Microsoft/LinkedIn, May 2024), and fewer than a third of companies had upskilled a quarter of their workforce (BCG, January 2025).
  • BCG’s 10-20-70 split puts 70% of AI effort on people, processes and culture. One session cannot carry that allocation alone.
  • Diagnose on a calendar: output at day 3, reuse at day 10, spread at day 30. License-activity dashboards will look acceptable through all three.
  • Name the owner before you name the vendor. Without that, a second session repeats the first.

If a session has already come and gone with nothing to show for it, the next decision is narrower than booking another one. A free AI-readiness call will work out which of the six modes you hit, which tasks the next session should be built around, and who in your company is in a position to own the month that follows.

Last Updated: Nov 19, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles