Back to Guides
Enterprise Software 15 min read

The case for hands-on AI training over curated slide decks

The case for hands-on AI training over curated slide decks

Search for hands on ai training and every result agrees with you. General Assembly’s enterprise pages state it outright: skills are developed through practice, not theory. NobleProg’s catalog says its instructor-led live courses “demonstrate through hands-on practice.” BrainStation promises applied prompting on research, writing, and analysis. Nobody sells a corporate AI session by saying attendees will watch.

So the word has stopped doing any work at the point of purchase. You cannot pick a session by looking for “hands-on”: the session that wasted your team’s Tuesday said it too.

The replacement is a number a proposal can disclose before you sign: how many minutes each attendee spends with their own hands on their own keyboard, in accounts your company controls, on work they did last week. Everything else is overhead. Some of it is needed. Most sessions never count it.

The claim is unanimous, so it tells you nothing

Run the search as a buyer and the top results split in two: vendor catalog pages, which all assert practice over theory, and vendor blog posts arguing that understanding a tool is not the same as being able to use it. The second group is correct and also selling the first.

Ask an AI assistant the same question and the answer changes shape: a reading list for one person learning alone, from DeepLearning.AI’s course library through Google’s AI Professional Certificate to a 23-week university certificate. Useful, and aimed at a different buyer. None of it helps a VP of Operations pick between three proposals on her desk.

The market is loud about the conclusion and silent about the test. That silence costs money, because the gap between a genuinely hands-on session and a deck with two exercises bolted on is invisible in any brochure, and the price range is the same either way. Market rates run roughly $2,000 to $15,000 for a session. Both products sell inside that band.

What the learning research supports, and what it does not

Start with the honest version: no controlled study compares slide-based AI training against hands-on AI training on later on-the-job use. If a vendor cites one, ask for it. What exists is the adjacent literature, and it points one way with some force.

The cleanest evidence comes from physics, not software. In Physical Experience Enhances Science Learning (Psychological Science, 2015), Kontra, Lyons, Fischer, and Beilock ran three lab studies plus a randomized field experiment in a college physics class. One group handled a rig of two bicycle wheels spinning on a shared axle and felt the torque fight back; the other watched. The group that had held the wheels scored higher on later quizzes, and brain imaging showed why: when they reasoned about angular momentum afterward, sensorimotor regions fired, and that activation predicted the score gain. The body had become part of the knowledge.

That is a narrow finding about one physical concept, and mapping it straight onto prompting would be a stretch. The mechanism is the useful part. Doing a thing lays down a different kind of memory than watching a thing, and the watching kind fades faster.

The transfer-of-training literature adds the commercial case. Baldwin and Ford’s 1988 review named the “transfer problem” that still defines the field: training content routinely fails to reach the job. The Association for Talent Development’s summary of the research is blunt, if rounded: less than 30 percent of learning gets applied on the job. Saks and Belcourt’s survey data traces the decay, with reported application falling from 62 percent right after the event to 44 percent at six months and 34 percent at a year. All three are directional rather than precise, and the direction is the point. A day of training is not a day of capability, and most of the leakage happens after everyone goes home.

One more reason this matters more for AI than for the average compliance course. A Fortune report on LinkedIn survey data in August 2025 found 51 percent of professionals called the intensity and frequency of AI training demands excessive, and a third said they felt embarrassed by how little they understood. Your team is not neutral here. They have sat through sessions already, and a second passive one costs you credit you will need later.

Interactive is not the same as hands-on

Here is the distinction most agendas blur. A session can be fully interactive while no attendee ever opens a tool.

Polls are interactive. Breakout talk about where AI might help the department is interactive. So is a facilitator building something impressive on the projector while the room watches, and that is the most common substitute for practice in this category, because it is easier to run and it demos well. It also produces the failure described in a trade-press account of one corporate class: the attendees had already made the charts being demoed on screen. The facilitator was performing a skill the room already had, for people whose real questions were further upstream.

A test you can run on any agenda. For each block, ask what is on each attendee’s screen when it ends. If the answer is “whatever the facilitator put there,” it is a demo. If it is “something they made from their own file, that they could not have made at nine this morning,” it is practice. Both belong in a session. Only one is what you pay a premium for.

The same logic applies to the material. A room practicing on a sample marketing brief learns to work on sample marketing briefs. Running the exercises on the team’s own documents is what makes the skill transfer, and it is a preparation question rather than an engineering one.

The keyboard-time fraction

Take the agenda, add up the minutes in which each attendee works alone on their own task, and divide by total session minutes. That fraction is the most honest single number about a workshop proposal, and you can compute it from the document a vendor already sent you.

Worked example. A four-hour session, 240 minutes: 30 minutes of welcome and AI landscape framing, 40 minutes of capability demo, 25 minutes of safe-use policy, 45 minutes of group talk and use-case brainstorm, a 20-minute break, 60 minutes split into three hands-on exercises, 20 minutes of wrap-up. The keyboard fraction is 60 over 240, or 25 percent. That is a curated deck with exercises attached, and it is what a large share of the market ships.

A session built for transfer inverts it. Framing drops to 15 minutes, the demo becomes a four-minute walkthrough of the first exercise rather than a showcase, the brainstorm disappears because the use cases came out of pre-work, and practice expands to about 150 minutes across four tasks, each ending with the attendee holding a reusable prompt and a written step. That is 62 percent, and everyone leaves tired in a different way.

Agenda element Deck-led session Practice-led session
Framing and landscape 30 min 15 min
Facilitator demo 40 min 4 min, tied to exercise one
Safe-use discussion 25 min, standalone Inside each exercise
Use-case brainstorm 45 min Done as pre-work
Independent practice 60 min 150 min
Keyboard-time fraction 25% 62%

Two cautions, because a metric with no limits becomes a stick to beat vendors with. A fraction of 100 percent is not the goal: a session with no framing and no safe-use guidance produces confident people doing unsafe things. Somewhere above half is the band worth paying for. And the fraction says nothing about what happens inside those minutes: practice on the wrong task, unsupervised, at a cohort size where nobody can look over shoulders, still fails. It is a filter, not a verdict.

Four things that quietly eat the practice minutes

A proposal can promise 60 percent and deliver 20 on the day. Four mechanisms account for most of that loss, and all four are preventable beforehand.

Setup. Accounts, logins, access approvals, and the one person whose IT policy blocks the tool. Thirty minutes of a half-day session can vanish here, and it comes out of practice, never out of the deck. Settle access in pre-work and confirm it two days out.

Cohort size. Published vendor guidance clusters around 10 to 12 attendees per facilitator, and that number is not about comfort. It is the limit at which someone can see a screen going wrong and step in before the attendee decides the tool does not work. Above roughly 20, individual correction stops and the session reverts to a presentation whatever the agenda says. For 200 staff, that means cohorts, not a bigger room.

The mixed room. Put a daily power user and a first-time user in the same exercise and the facilitator has to choose who to lose. Both get a watered-down session. Splitting by current usage rather than by department costs one short pre-work survey and buys back more practice minutes than any other change.

Demo creep. This one happens live. A good question gets a good answer, the facilitator shows rather than tells, and the room watches for eleven minutes. It is the most forgivable failure here, and the most common. A session built to resist it has the facilitator answering at the attendee’s keyboard, not at the projector.

What a deck is still the right purchase for

A piece that cannot name its own limits is a pitch, so here are the cases where the curated deck wins outright.

When the goal is awareness across a large group on a fixed date, a well-built talk to 150 people is the right instrument, and a hands-on session for the same audience is a scheduling fantasy. When the content is policy, a clear deck plus a written one-pager beats practice, because the deliverable is a rule people can look up rather than a skill. A board that wants to know what the company is doing about AI needs 40 minutes and a briefing, not a lab.

What those cases share is that the goal is informing people, not changing what they do on Monday. The failure here is not that decks exist. It is that decks get sold and bought as capability building, at a price that includes the payroll cost of the room, and then nobody checks.

How to get the number out of a proposal

Most vendors will not volunteer a keyboard fraction, and the silence is rarely evasion. Nobody asks. Five questions pull it out, and how they are answered is itself the signal.

  1. Minute by minute, what is each attendee doing? A vendor who has run this answers in blocks with timings. One who has not answers in themes.
  2. What do you need from us two weeks out? A session built on your work needs documents, accounts, and a written list of what may not enter the room. No pre-work request means a generic session, whatever the cover page says.
  3. How many facilitators for this headcount? One per 10 to 12 attendees during practice, or an account of how correction happens without it.
  4. What is on each person’s screen when they leave? Listen for named artifacts: prompts labeled by task, a written procedure, a before-and-after on one task they own.
  5. What did you cut to fit our four hours? The answer shows the design priority. If framing survived and practice shrank, you know the product.

Push on all five and the two products separate in ten minutes. What happens after the day is where the real decay sets in, and that belongs in the same conversation: our guide to designing a session that changes Monday morning covers the follow-up, and the working definition of the format and its outputs covers what should leave the room.

The wider argument: buying licenses is not the same as building capability, and the gap between the two is where most AI budgets now sit. AI enablement is the work of closing it. Keyboard time is the cheapest honest proxy for whether a purchase does any of that work.

Frequently asked questions

What counts as hands-on AI training?

Each attendee working on their own in a real AI tool, in an account your company controls, on a task from their own job, with a facilitator close enough to fix mistakes as they happen. Watching a demo and answering polls are not practice, however lively the room feels.

How much of a workshop should be hands-on?

Somewhere above half the scheduled minutes, counted as independent attendee work. A quarter is a deck with exercises attached. Pushing toward 100 percent is its own failure, because framing and safe-use guidance have to come from somewhere.

Is a facilitator demo a waste of time?

No, when it is short and tied to the exercise that follows it. Four minutes showing the first move is useful. Forty minutes of showcase is the most common way a hands-on session becomes a passive one, and it is usually unintentional.

Can hands-on AI training work remotely?

Yes, and screen sharing can make correction easier, since a facilitator sees the screen without walking over. The trade is that you cannot read a room of 15 remote faces, so remote cohorts should run smaller and need a second facilitator above about 10 people.

What if our team is at different skill levels?

Sort the cohorts by what people already do with an assistant, not by which team they sit on. One pre-work question, asking what each person used an assistant for last week, sorts a mixed group well enough, and it buys back more practice time than any other prep step.

Does hands-on training cost more than a lecture?

Usually yes, and the gap is in prep rather than delivery. Building exercises from your documents takes facilitator hours before the day. Two near-identical proposals that differ by several thousand dollars usually differ in prep hours, so ask how many are budgeted.

How many people can be in one hands-on session?

Around 10 to 12 per facilitator, up to roughly 20 with a second facilitator. Past that, nobody is correcting individual screens and the session has become a presentation. Larger headcounts need repeated cohorts.

How do we check afterward whether the practice stuck?

Take a before-and-after on one task per attendee during the session, then ask four weeks later which procedures they still run and what they abandoned. Login counts and satisfaction scores measure attendance and mood, which is not the same question.

Our vendor says their deck is interactive. Is that good enough?

It depends what you are buying. For awareness across a large group, yes. For changing how a team works next week, no: ask what is on each attendee’s screen at the end of each block, and if the answer is always the facilitator’s slide, you are buying a briefing.

Key takeaways

  • Every vendor in this category claims hands-on, so the word cannot be used to choose between them. The claim is unanimous, so it cannot separate one vendor from another.
  • The replacement test is arithmetic: independent attendee working minutes divided by total session minutes. Above half is the band worth a premium; 25 percent is a deck with exercises attached.
  • Interactive and hands-on are different things. Polls, group talk, and an impressive projector demo can fill a session in which nobody makes anything from their own files.
  • Four preventable leaks eat practice minutes: unresolved tool access, cohorts above 20, mixed skill levels in one exercise, and live demo creep.
  • A curated deck is the right purchase for awareness at scale, for policy, and for board briefings, and the wrong one when the goal is different behavior on Monday.
  • The evidence for practice over watching is strong but adjacent. No controlled study compares slide-based and hands-on AI training on later tool use, so ask any vendor citing one for the reference.

Reading proposals and unsure which one is a practice session? We run a free AI-readiness call: a working conversation about which of your processes a hands-on session could change, what the agenda would need to look like, and what should be in your team’s hands when it ends. Book a time.

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