A whole company needs four AI workshop tracks. A foundation track every employee sits once. An applied track for the functions that own a recurring deliverable. An approver track for the dozen people who sign off on tools, data and spend. And a facilitator track for the three or four staff who keep it running after the vendor stops answering email.
Four tracks, one company, and a small fixed number of hours to split between them. That last part is why most AI workshop curriculum plans fall over: they are written as a menu of topics rather than a division of a budget, so they grow until nobody can schedule them.
Almost every company that buys AI enablement buys the foundation track, runs it twice, and calls that a program. The sessions go well. Nothing in the quarterly numbers moves, because one track out of four was bought.
The four tracks at a glance
| Track | Who sits it | Contact hours | Room size | What they can do after |
|---|---|---|---|---|
| 1. Foundation fluency | Every employee, once | 1.5 to 3 | Up to 15 per facilitator | Repeat one real task of their own, unaided |
| 2. Applied practice | The 20 to 40% who own a recurring deliverable | 5 to 6, over two sittings | 8 to 12, one function per room | Produce a weekly deliverable faster, with a defensible verification step |
| 3. Approvers | The 8 to 15 who authorize tools, data and spend | 2 to 3 | 6 to 10 | Read a vendor claim, set a data rule, say where recovered hours go |
| 4. Internal facilitators | 3 to 6 volunteers with allocated time | 12 to 16 over several weeks | 3 to 6 | Run track 1 without the vendor in the room |
Cutting to three leaves a hole you can name in advance. Drop track 3 and nobody is allowed to decide anything. Drop track 4 and the capability leaves when the invoice is paid. Drop track 2 and you have awareness with no deliverable attached.
The axis is Monday, not seniority
Vendor ladders tier by difficulty and job title: Essentials, then Strategic, then Advanced, each labeled for professionals, leaders, product managers or engineers. That organizes a catalog well, because a catalog sells one seat at a time. It organizes a company badly, because difficulty tiers assume what separates two employees is how much they know. What separates them is what has to be true on Monday.
A benefits administrator and a CFO may have identical starting knowledge. The administrator has to draft 40 letters without putting a social security number into a chat window. The CFO has to decide whether finance may use the tool on ledger exports at all. Two jobs, two tracks, and neither is a harder version of the other.
Which is why these four cut across every department rather than sitting inside one. Sorting curriculum by function is a separate exercise, and the role-based view of the same problem covers what changes between a legal team and a sales team. The tracks here are the other axis: finance usually has people in all four.
The hour budget nobody does the math on
Every curriculum framework published by a vendor adds tracks. None subtracts. The skipped constraint is that a company has a known, small number of training hours per employee per year, for every subject combined.
ATD’s 2025 State of the Industry report, built on data from 539 organizations, put formal learning at 13.7 hours per employee in 2024, down from 17.4 in 2023. That is the whole budget: compliance, safety, systems, management training, onboarding, and now AI.
Price the four tracks against that ceiling. In a 200-person company:
- Track 1 at 3 hours for all 200: 600 hours
- Track 2 at 6 hours for 60 people across three functions: 360 hours
- Track 3 at 3 hours for 12 approvers: 36 hours
- Track 4 at 16 hours for 4 facilitators: 64 hours
That is 1,060 hours, or 5.3 hours per employee: roughly 39% of the company’s entire annual formal learning budget, on one subject, in one year. Say that number out loud before promising a board everyone will be trained by Q2.
Two things follow. The full build is a one-time cost: in year two, track 1 shrinks to new hires and track 4 does most of the delivery, which more than halves the figure. And if 39% is not available, cut headcount inside tracks rather than cutting tracks. Half the company in track 1 plus a complete track 3 beats everybody in track 1 and no approvers. At larger headcounts the calendar, not the curriculum, becomes the binding constraint, which is the real subject of rolling out in cohorts.
What each track teaches
Track 1: one job, not an overview
Ninety minutes to three hours. One task, demonstrated once on a document the team brought, then repeated by everyone on their own version while a facilitator walks the room. No tool tour, no history of the field, no slide explaining what a token is.
The temptation is to use the one session everybody attends to cover everything. Resist it. Someone who leaves having seen a dozen capabilities does none of them.
Start with the least technical staff. Brynjolfsson, Li and Raymond studied 5,172 customer support agents after an AI assistant arrived and found a roughly 15% average rise in issues resolved per hour, concentrated among the less experienced. The people with the least tooling background had the most to gain, and format choices for non-technical rooms goes deeper on that session.
Track 2: the function that owns a recurring deliverable
Five or six contact hours split across two sittings two to three weeks apart, one function per room, 8 to 12 people. This track produces the measurable change, and it is the one most often skipped because it costs the most hours per head.
Pick the room by deliverable, not by enthusiasm. The finance team that turns ledger exports into leadership commentary every Thursday beats the marketing team that is merely keen: a named, recurring deliverable gives the session a target and you something to measure.
Verification belongs here, taught inside the task. What to check before sending, which fields never get pasted, what a model does when it does not know. As a policy slide that is theater. Taught on the commentary they drafted 20 minutes ago, it sticks.
Track 3: the approvers get a different subject, not a simpler one
Two to three hours with the 8 to 15 people who authorize tools, data handling and spend. The common mistake is running track 1 for this group with better catering.
Their subject is different: how to read a vendor claim, what to permit on which data, what to measure, and where recovered hours go. That last one matters more than it sounds. BCG’s AI at Work 2026 survey of 11,749 workers across 14 countries found 66% get limited or no guidance on what to do with the time AI saves them, and only 36% feel adequately upskilled, a share unchanged from the 2025 edition. Unclaimed hours get absorbed back into the inbox, and the gain vanishes from every measure you own.
They still need hands on a keyboard, on their own work: a real vendor proposal, their own board deck. An approver who has never watched a model invent a citation will approve things they should not. Running an executive session that is not condescending is mostly a matter of choosing their material.
Track 4: the three people who keep it alive
Twelve to 16 hours over several weeks for three to six people, who need allocated time, not goodwill. They learn to run track 1 and host the follow-up clinic, and they inherit the exercise set, the prompts and the data rules.
This is the only track that changes the slope of the cost curve. Without it, every new cohort and every new hire is another purchase order. What a train-the-trainer program costs and who owns the curriculum once the engagement ends are both contract questions, settled before the first session.
Buy them in this order
Not the order they are numbered in.
- Track 3 first. Until the approvers have set the data rules and named the tools, every other track teaches people to do something they are not sure they are allowed to do.
- Track 2 second, in one function only. One room, one deliverable, two sittings. This is your proof and your source material: the real documents and real failures from this room make track 1 concrete instead of generic.
- Track 1 third, company-wide. Now the session uses artifacts from a function the room recognizes, and the facilitator can answer “are we allowed to do this” without hedging.
- Track 4 last, but contracted first. Pick facilitators from the people who were visibly good in track 2. You cannot identify them before you have watched a room.
The intuitive order is the reverse: foundation training for everybody, then specialization. It produces 200 people with an interesting afternoon behind them, no decision, no deliverable and no owner, which is the main subject of the case for building capability rather than buying licenses.
What the catalog sells instead
Published formats and prices from two providers a buyer meets early, checked against their own pages at the time of writing.
| Provider | Published structure | Published price |
|---|---|---|
| Improving | Five tiers (Essentials, Strategic, Advanced, Certification, Custom). Essentials is a 4-hour unit sold in job-title variants: Professionals, Leaders, Product Managers, Engineers, Teams | $495 per student for a public Essentials class, $795 for certification. Strategic, Advanced and Custom quoted on request |
| UMBC Training | 12-month pass to unlimited instructor-led 4-hour workshops, from a schedule of 24 titles | $2,500 for one student, $1,875 each at 25, $1,250 each at 100 |
Both are well built for what they are. Neither is a company curriculum. They are seat ladders, priced per head and tiered by difficulty, and they leave you the two hardest decisions: who belongs in which track, and what you buy first.
Notice what nobody stocks: no catalog item for “two hours with our 12 approvers, using our own vendor proposal,” and none for “train four of our staff to deliver this.” Those get commissioned, which is the honest version of the custom versus off-the-shelf argument. Private sessions run $2,000 to $15,000 at current market rates.
Spacing beats a longer agenda
When budget forces a cut, the instinct is to compress track 2’s two sittings into one long day. Wrong cut, and the learning research is unusually clear about why.
Cepeda, Pashler, Vul, Wixted and Rohrer synthesized 839 assessments of distributed practice across 317 experiments in 184 articles in Psychological Bulletin in 2006. Their central result: the gap between sessions and the delay before the test interact, so the longer you need learning to last, the wider the best gap becomes. Massing material into one block reliably loses to spreading it out.
The second appointment is also where the useful questions live. On day one nobody has a question worth asking yet. By week two they have three: the prompt that worked twice and failed the third time, the summary that invented a number, the task that needed two tools. A longer first session cannot answer those, and the session design playbook works through both sittings.
How to tell each track worked
Each track has its own success test, and none is a confidence survey. People rate their own new skills generously, so a rating sheet cannot tell a session that worked from one that felt good.
- Track 1: three weeks later, count how many attendees repeated the task unaided. A number from a short check-in, not a rating.
- Track 2: the named deliverable’s clock time, before and after, on work that already existed. The only track that yields a defensible number.
- Track 3: an artifact. A data rule written down, a line in next year’s budget, a measure added to an existing report. No artifact, no result.
- Track 4: an internal facilitator runs a track 1 session with the vendor present only to watch.
Track 2 is the one a CFO will ask about, which is why the first room should be picked for how measurable its deliverable is. Proving capability moved covers the baselines to capture before the first session rather than after.
Where this framework breaks
Track 4 as volunteer work. Facilitators with no allocated hours quietly stop facilitating by month three.
Running track 2 on sample data. If the sandbox question is unsettled, exercises get built on fictional documents and the session teaches a demo. Settle it in track 3, which is partly why track 3 goes first.
No second appointment. Book track 2’s clinic before the first sitting runs. It never gets booked later.
Frequently asked questions
How many AI workshop tracks does a company actually need?
Four: foundation fluency for everyone, applied practice for the functions with a recurring deliverable, a separate track for the people who approve tools and spend, and a facilitator track so delivery survives the vendor leaving.
Can we start with just the foundation track?
You can, and it is the most common way programs stall. One session for everybody, with no approver decisions behind it, leaves people unsure what they are permitted to do. If budget covers one track this year, run the approver track plus one applied room instead.
What is the difference between the approver track and an executive briefing?
A briefing informs. The approver track produces an artifact: a data rule, a budget line, a measure. If nothing is written down by the end, it was a briefing.
What does a four-track curriculum cost?
Private facilitated sessions run about $2,000 to $15,000, priced per room rather than per head, so the total tracks how many rooms need their own material. Catalog classes publish lower unit prices, around $495 for a half day, but they teach generic material and leave each attendee to translate it alone.
Do technical staff need their own track?
They need their own room inside the applied track, not a fifth track. Engineers already read documentation and experiment unprompted. What they lack is verification and data-handling discipline, taught on their code and their tickets.
Who should be in the facilitator track?
People you watched do well in an applied session, picked after it rather than nominated before. Enthusiasm predicts volunteering. Credibility and patience for repetition predict teaching.
How often do we rerun this?
Foundation belongs in onboarding rather than an annual event, and folding it into the first two weeks is cheaper than a yearly refresher. Applied rooms are worth revisiting yearly as the work changes. Approvers need an hour whenever the tools or rules change.
Will a vendor sell us all four?
Some will, and most catalogs cannot, because their product is a seat and two of these tracks are commissioned rooms. Ask which of the four a proposal covers. A vendor that bids only on foundation sessions and calls that a curriculum is telling you something useful.
Key takeaways
- Four tracks cover a company: foundation fluency for all staff, applied practice for the functions with recurring deliverables, a distinct track for approvers, and a facilitator track so the capability stays in-house.
- The useful axis is what each group must be able to do next Monday, not how senior or technical they are. Catalog tiers sort by difficulty because that is how seats are sold.
- Do the hour arithmetic. A full build for 200 people runs about 1,060 hours, roughly 39% of a year’s formal learning budget at ATD’s 13.7 hours per employee. Cut headcount inside tracks before cutting tracks.
- Buy in the order approvers, one applied room, company-wide foundation, then facilitators. The intuitive reverse produces attendance and no decisions.
- Measure each track differently: repeat counts, deliverable clock time, a written artifact, a session run without the vendor.
If you are deciding which track to fund first, a free AI-readiness call is a short conversation about your headcount, your recurring deliverables and who approves what. You will leave with a track order, whether or not you run it with us.
Arthur Wandzel
