Train one department first. For almost any company between 50 and 500 people that is the right default, and the advice recommending it never says what it costs: facilitation is priced per cohort, so covering the same headcount in six sittings runs roughly half again to double the fee of one company-wide program. The premium is not waste. It buys the right to stop after the first department, which is worth more than the money while nobody can yet show the training changes anyone’s Monday.
Search AI training for one team versus the whole company and you get answers about scaling an AI system: data pipelines, model governance, pilot purgatory. Useful if you are rolling an agent across a warehouse network. You are buying facilitated sessions for people who already hold licenses, and your real constraints are a budget you can spend once and a calendar five functional VPs have to agree to. Whether this capability should be bought at all, or built inside, is the larger question worked through in our guide to buying versus building AI enablement capability; the reason a license count stopped standing in for readiness is argued in the case against buying seats and calling it capability.
What you are actually choosing
You are choosing a unit of purchase, not a rollout plan.
A facilitated session is quoted per group, per sitting. Split 120 people into six groups of 20 and you have bought six engagements, each with its own scheduling and its own facilitator day. Put all 120 through two large sittings and you have bought one. Nothing about the teaching changes. The invoice does, and so does the point at which you can walk away.
Most vendor evaluation checklists stop at curriculum, references, and price, and not one of them asks what the second cohort costs. Which provider is strongest is a separate question, worked through in our read on the best AI training companies for corporate teams.
Companies get stuck at exactly this step. McKinsey’s State of AI in 2025 found 88% of respondents reporting regular AI use in at least one business function, up from 78% a year earlier. Getting one function going is close to solved.
Going wider is not. In that same survey, about a third of organizations said they had begun to scale their AI programs and only 39% reported EBIT impact at the enterprise level. The gap between one function working and the company seeing it in the numbers is where most AI budgets quietly die, and it opens at the sequencing step rather than in the first session.
Base rates at your size also sit lower than the headlines suggest. In Census Bureau Business Trends and Outlook Survey data for the period ending May 3, 2026, 32% of firms with 100 to 249 employees reported using AI in any business function, against 37% at firms of 250 or more. Two-thirds of companies your size are not there yet, which is worth saying to a board before a deadline gets set out of fear alone.
Three things that cannot be sequenced
Stage the sessions. Do not stage these.
The tool decision. Pick the assistant your company standardizes on, ChatGPT, Claude, Gemini, or Microsoft Copilot, before cohort one sits down. Training people on a product you replace in February writes off the payroll hours, not just the fee.
Safe-use expectations. Everyone needs to know what may be pasted into a chat window on day one, because the people you have not trained yet are already using these tools on their own accounts. Staging the rules behind the sessions means running a company where the untrained are the unsupervised.
The measurement baseline. Capture a before on one real, repeated task across every department you intend to train. A baseline taken only inside the first group produces a result nobody outside it accepts.
Most advice folds all three into “phases”. That is how companies end up with a trained marketing team, a policy memo nobody read, and no number to put in front of a CFO.
What each sequence costs
Take a 120-person company and a half-day session, four hours in the room. Market rates for a private AI enablement session run roughly $2,000 to $15,000 per engagement. Payroll runs about $56 an hour for a $90,000 salary once you add roughly 30% employer burden across a 2,080-hour year.
| Company-wide at once | Department first, six cohorts of 20 | |
|---|---|---|
| Facilitation fee | One engagement, upper half of the range | Six engagements, lower half each |
| Realistic fee total | $12,000 to $15,000 | $18,000 to $28,000 |
| Payroll hours consumed | 480 | 480, spread over two quarters |
| Payroll cost of those hours | About $27,000 | About $27,000 |
| Calendar to full coverage | Two to three weeks | One to two quarters |
| Cost of being wrong | Whole fee plus 480 hours | One fee plus 80 hours |
The payroll column does not move: those hours follow headcount, not scheduling, and payroll is the larger number either way. The fee column moves by roughly $6,000 to $15,000, and that difference is the whole financial case for either sequence.
Whether that premium is affordable depends on a budget line that is tightening. The Association for Talent Development’s 2026 State of the Industry report put direct learning expenditure at $846 per employee in 2025, down from $1,254 in 2024, while average formal learning hours rose from 13.7 to 16.7. More hours, less money. ATD samples a few hundred organizations per edition, so read that as directional, but the direction turns an unpriced fee premium into a real problem. The payroll half of this arithmetic is what also makes free courses expensive, laid out in our comparison of free AI training against paid workshops.
The premium buys an exit
Do this part on a napkin before signing.
A company-wide buy that turns out wrong costs the full fee plus 480 payroll hours, call it $40,000 of real resource, plus a reputation hit that makes the second attempt harder to fund. One wrong cohort costs a fee plus 80 hours, roughly $8,000, and that lesson is cheap enough to spend.
So the question is not which option is cheaper. It is whether you would put better than one-in-four odds on this being the wrong program for your company. If you would, the premium is the cheapest insurance available. If you have already run a session that worked, have a named owner, and know which task you are trying to change, you are paying for an exit you will not take.
Take the insurance once: run cohort one at a premium, then buy the rest as one block at the better rate.
Pick the department whose results will be believed
The first cohort has one job beyond its own productivity: produce evidence the next department accepts.
That rules out the two picks companies make by reflex. Starting with the executive team feels like sponsorship, but executives skip the follow-up work, and a result built on slide decks reads as irrelevant below the VP line. Starting with marketing is tempting because marketing already uses the tools, which is exactly why the before-and-after is small and everyone else reads it as “of course, they’re marketing.”
Pick on three criteria instead:
- Text-heavy work with reviewable output. Proposals, reports, case notes, support replies. The gain has to show up in an artifact somebody else can read, not in how the room felt afterward.
- A manager who will hold the follow-up. One person who runs the 30-day retest and reports the number whether or not it flatters them.
- Peers who respect them. The department whose working habits others copy. Credibility travels along lines that already exist, and a deck cannot create them.
Operations, finance, and client service usually score well on all three. Sales scores well on the third and badly on the second, because follow-up competes with quota.
Where department-first breaks
Four failure modes, in the order they usually show up.
Cohort two never gets scheduled. The quarter turns, the champion inherits a new priority, and the program ends with one trained department and a budget line that reads as waste. The fix is clerical: put the second date in the first agreement.
The material goes stale. In the sessions we run, the method and the exercise structure carry between departments. The examples do not, because they are built on one department’s real documents, and finance does not learn much from a reworked marketing brief. A rebuild is cheap at six weeks and close to a redesign at six months.
Nobody outside the room can see the result. A win locked inside one department is a rumor, not evidence. Make a 20-minute internal readout, delivered by the department rather than the vendor, a deliverable of cohort one.
Per-cohort drift. New facilitator, drifting content, no comparable measurement, and by cohort four you cannot tell whether the program works or the groups got easier. Hold the facilitator and the baseline task constant.
When company-wide at once is right
Buy it in one block when any of these describe you:
| Your situation | Why one block wins |
|---|---|
| Under about 80 people | You are one or two cohorts anyway; staging adds calendar and no information |
| A compliance date to clear | You need safe-use training on record for everyone, and partial coverage fails the test |
| One dominant job family | If most of your headcount does the same work, the second cohort teaches you nothing new |
| A tool switch or reorg already landing | Workflows are being rewritten regardless; one disruption beats six |
| Strong sponsorship and a fixed date | Attendance is the binding constraint, and an owned company date is the only way to get it |
The last line is underrated. A half day the CEO schedules and attends fills the room; six sessions negotiated with six VPs get whatever each quarter allows.
Four clauses that make cohort two easy
The recommended sequence only works if the second purchase is not a fresh negotiation. Settle these in the first agreement, while the vendor still wants the deal:
- A per-cohort rate, held, with an expiry. Price cohort one, then name the rate for cohorts two through six and the date it lapses. Two quarters is a fair window. Without it you renegotiate from zero each time, and your own half-finished program becomes the vendor’s strongest card.
- A named facilitator with substitution rights. Name the person delivering cohort one and require written notice plus your approval for a substitute. Continuity is most of what you are buying and the first thing that quietly changes, one of the differences between the provider types compared in our look at freelance trainers against enablement agencies.
- Material ownership plus a rebuild at no new design fee. You own the workbook and the exercises, and rebuilding examples on the next department’s documents is delivery work rather than a new design engagement. Price it before you need it.
- A readout as a deliverable. Cohort one ends with a short written result handed to the next department’s manager. Make it a line item or it will not happen.
The sequence we recommend at 50 to 500 people
- Week 0. Settle the tool, publish safe-use expectations company-wide, and baseline one real task across every department you intend to train. Half a day of work that saves a quarter of argument.
- Weeks 1 and 2. Cohort one: 15 to 20 people from the department that scored best on the three criteria above, working on their own documents.
- Weeks 3 to 6. Re-run the baseline task unannounced at 30 days and have the department manager write the readout. This is your purchase decision for everything after it.
- Weeks 6 to 10. Cohorts two and three at the rate you held. If the retest moved nothing, stop and fix the diagnosis rather than buying more sessions.
- Quarter two. The remaining departments as one block at the better rate, plus a reinforcement layer. The argument for pairing a live session with cheap repetition afterward is set out in our retention comparison of platforms and facilitated sessions.
- Ongoing. Re-baseline the same task each quarter: one task, one number, same conditions. That number is what the next budget request rests on.
Expect a quarter before an outsider would notice the behavior change. Anyone promising it in an afternoon is selling an afternoon.
Frequently asked questions
Should we train one department first or the whole company?
One department first, unless you are under about 80 people, facing a compliance date, or already pushing a tool change through everyone’s workflow. The staged version costs more in fees and buys the right to stop before the budget is gone.
How do we choose which department goes first?
Pick the one whose results other departments will believe: text-heavy work with output somebody else can read, a manager who will run the 30-day retest honestly, and peers who copy their habits. Operations, finance, and client service usually fit. The executive team almost never does.
Does training one department at a time cost more?
On fees, by roughly half again to double, because facilitation is priced per cohort and six groups is six engagements. Payroll hours are identical either way and they are the bigger number. Budget the fee premium as the price of an exit.
How long should we wait between departments?
Weeks, not quarters. Four to eight weeks covers a 30-day retest and keeps the material fresh. Past a quarter the examples need rebuilding and the champion has usually moved on.
What if the first department loves it and nobody else wants it?
That is a signal about your evidence, not about the other departments. A readout saying people enjoyed the session sells nothing. Go back to the baseline task, get a number on an artifact, and let the next manager read it.
Can the first department just teach everyone else?
For the mechanics, sometimes. For the method, rarely. It works when one person from cohort one gets explicit hours to run short practice sessions on their own department’s work, and it fails when teaching is added to a full workload. Hours removed matter more than talent.
How many people should be in one session?
15 to 20 for hands-on work. Above about 25, feedback on someone’s own document stops being possible and you have bought a presentation. Covering 120 people in a day is a briefing, and it should be called that.
What does a half-day session actually cost?
Corporate AI training is quoted per engagement, at market rates roughly between $2,000 and $15,000, moving inside that band with group size, length, and how much material is built on your own documents. Then add four hours of payroll per attendee, usually the larger figure.
How do we know cohort one worked before buying cohort two?
Re-run the exact task you baselined, unannounced, at 30 days, scored by someone who is not selling you training. Then count how many people use the tool on that task without being asked.
Key takeaways
- Department-first is the right default at 50 to 500 people, at roughly half again to double the facilitation fee of one company-wide buy. That premium is an exit option, not waste. Keep the gaps between cohorts to weeks, because past a quarter the examples go stale.
- Payroll hours are the same either way and larger than the fee: a 120-person half day is 480 paid hours, about $27,000.
- The tool choice, safe-use expectations, and the measurement baseline go company-wide on day one. Only the sessions get staged.
- Pick the first department for believability rather than enthusiasm or seniority. Its product is evidence the next department accepts.
- Settle four things in the first agreement: a held per-cohort rate with an expiry, a named facilitator, material ownership with a rebuild clause, and a readout as a deliverable.
Working out which department should go first, or whether your company is one of the cases that should buy it in one block? Book a free AI-readiness call and we will look at the work your teams do before anyone recommends a sequence.
Arthur Wandzel
