Most AI enablement programs are designed for 200 people and quoted to 40. Part of what they contain scales down untouched, part inverts, and part is weight you pay for and never use.
One fact does the sorting: a facilitator’s day costs what it costs whether eight people or eighty sit in the room. Fixed delivery cost over a smaller group pushes every per-head number the wrong way. So a mid-market proposal reads as expensive at 40 heads, and the fix is to change what you buy rather than ask for a discount.
The capability gap by company size is measured
The Census Bureau asks firms whether they use AI to produce goods or services. Its May 2026 release from the Business Trends and Outlook Survey, reference date May 3, 2026, put the national rate at 19.8%: 37% at firms of 250 employees or more, 32% at 100 to 249, and under 20% below 20 employees.
The second finding matters more. Use rose among firms with at least 20 employees over the preceding months and did not change significantly below 20. The gap by size is widening on its own, so a 40-person company weighing this decides against a trend line, not in a vacuum.
What scales down in an AI enablement program without modification
Four parts of a standard program are size-neutral. They cost about the same per person at 40 as at 400.
| Element | Why size does not change it |
|---|---|
| Scheduled practice on your own documents | One person, one real task, with help. Group size changes the room, not the work |
| A named internal owner | Someone with standing to change how work gets done. Easier to name at 40 people, where everyone can see who it is |
| Before-and-after measurement | A smaller denominator reads cleaner, and you can inspect every output, which a 400-person company cannot |
| Tool licenses | Team tiers are built for small groups. Claude Team lists at $20 per seat per month billed annually with no seat minimum published; Coursera publishes Teams at $399 per user per year for 2 to 499 users |
Licenses are the one place where small companies are structurally advantaged: per-seat products are priced linearly and sold self-serve from a handful of users. The enterprise tier behind them does not scale down. Claude Enterprise carries a 20-seat minimum, and Microsoft 365 Copilot sits at $30 per user per month on annual commitment on top of a qualifying base license. Read the seat floors before the per-seat rate.
Measurement is the quiet advantage. At 40 people you can read the work product of everyone who attended, which no mid-market program can buy. The sequence from rollout to measured capability is the same at any size, and our AI enablement playbook sets it out in order.
What inverts below 50 people
Three line items do not shrink. They reverse.
Per-cohort pricing. A fixed fee divided across fewer heads produces a higher cost per head, and the benchmarks already show the gradient. Training magazine’s 2025 Training Industry Report, its 44th annual survey of US corporate training spend, put per-learner spend at $1,091 for small companies against $468 for large ones. That survey covers only organizations of 100 employees or more, so the 2.3x spread appears before your band begins, and nothing below 100 heads reverses it. Which structure you are quoted therefore matters more here than anywhere, and the three structures and what each one pays your vendor for set the per-head figure before any negotiation does.
Role-group curricula. Separate sessions for sales, finance, and operations are the correct mid-market answer. At 40 people you have nine functions with four people each, and nine cohorts of four means nine facilitator days to cover one company. The substitute is one mixed-function room with exercises that differ by role inside the session, which is a delivery decision rather than a budget one.
Phased waves. Mid-market programs run a pilot, then a second wave, then the rest. Phasing needs a bench: people who absorb the first wave’s lessons and carry them forward. Under 50 people there is no bench, and the whole company fits in one room.
The AI enablement arithmetic nobody publishes
Reduce every option to cost per person per supervised practice hour, for a 40-person company.
| Option | Spend | Scheduled practice hours | Cost per person-hour |
|---|---|---|---|
| Catalog seats for everyone, at a published $399 per user per year | about $16,000 yearly | 0 scheduled | undefined, recurring |
| One facilitated day for the whole company, at the $3,000 to $8,000 market day rate | $3,000 to $8,000 once | about 240 person-hours | $13 to $33 |
| Nine role-specific days | nine times the day rate | about 240 person-hours | nine times as much, same hours |
The middle row is the only one with a defined denominator. ATD’s 2026 State of the Industry report, drawing on 2025 data from 340 organizations, put average direct learning spend at $846 per employee and formal learning hours at 16.7 per year, or roughly $51 per formal learning hour across every subject a company trains on. One facilitated day for a whole small company lands below that, the opposite of what its per-head figure suggests.
The top row is where small companies lose money quietly. Catalog seats are cheap per head and scale down cleanly on price, so they look like the obvious answer. The function that makes them work does not: a mid-market buyer has someone who assigns courses, chases completion, and curates a path. Under 50 people nobody has those hours, so the subscription becomes access that renews. Why optional access does not build capability is the argument in the case for funding enablement as its own line.
For dollar figures across program shapes and team sizes, the companion piece on what an AI enablement program costs carries the ranges.
What is dead weight at this size
Strike these before you sign.
- A paid readiness assessment. Two hours of asking people what they do all day tells you more at 40 heads. A scored diagnostic earns its fee when the org is too large to see at once.
- Governance scaffolding beyond one page. You need a written policy on what data goes into which tool, and a named person to ask. Committees and tiered approval paths assume a headcount you do not have.
- An open-ended monthly retainer. Reinforcement has a natural end. Buy expiring facilitation hours, not a standing invoice.
- A learning platform nobody can run. The platform is not the cost. The unstaffed hours are.
A vendor who agrees to strike all four is reading your situation. One who insists they are all required is selling a template.
What AI enablement for a small company should contain
Four parts, in order.
- Preparation against your own documents. Your proposals, client updates, reporting pack. Generic exercises teach a tool; your own material changes next week’s work. Ask how many quoted hours sit here, because this is the line vendors cut first when a quote is squeezed.
- One mixed-function session with per-role exercises. Everyone in one room, shared ground rules, then exercises split by what each person produces. The sales lead and the controller learn the same mechanics on different documents.
- A named owner with hours set aside, before you sign. Not a volunteer. Someone with standing in the operating rhythm and explicit time, which here usually means the COO, chief of staff, or an operations lead who gives up something else.
- One follow-up block, four to six weeks out. Long enough for habits to have formed or failed, short enough to fix them. Priced in hours, expiring by default.
That is the whole shape: the mid-market version with role segmentation collapsed into one room and phasing removed.
The bench problem, which is the real constraint
The clause that protects a mid-market buyer is handover: the provider trains two named employees to deliver the core material and hands over exercises and facilitation notes in editable form. Good clause, and worth asking for. At 40 people it hits a wall, because two employees with spare hours do not exist.
Be honest about which case you are in:
- One owner with real hours. Ask for editable materials and one trained owner rather than two. The clause still works.
- One owner with no hours. Fix that before spending anything. A program landing in a company with no set-aside time fades once the sessions end, at full price.
- No owner. Do not buy yet. No vendor, price, or structure survives this, and fixing it costs nothing.
A provider who asks who will run this afterward, before quoting, is reading the risk correctly. Our twelve rules for choosing an enablement partner name the artifact to demand behind each promise.
A decision table by headcount band
| Headcount | Buy | Skip |
|---|---|---|
| Under 10 | Team licenses, plus a half-day on your own documents | Everything else. The founder is the owner and the program is one morning |
| 10 to 25 | One facilitated day, whole company, per-role exercises in the room, one follow-up block | Role-split cohorts, catalog subscriptions, any retainer |
| 25 to 50 | One or two days split by broad function, one follow-up block, a one-page use policy | Phased waves, paid assessments, governance scaffolding |
| 50 and above | The mid-market shape: role-group cohorts, a measured pilot, a second wave | Assuming the small-company playbook still fits |
The 50-person line is not arbitrary. It is roughly where a single function first holds enough people to fill a cohort, which is what role-based delivery needs. Below it, segmentation costs more than it returns.
Four questions for a vendor whose deck was written for 300 people
- What is your minimum fee and minimum group size, in writing? Almost no provider publishes either, and both exist. A vendor who will not state them is quoting a mid-market program with the headcount swapped.
- How many of these hours are preparation against our documents, and how many are delivery? A fee that cannot be split was never costed.
- Which of these deliverables would you strike for a team our size? A provider with small-company delivery behind them names two or three unprompted.
- Who at our company runs this after you leave, and what do they receive? The answer separates a program from an event.
Four clear answers mean the proposal was written for you, not translated to you.
Frequently asked questions
Is AI enablement worth it for a company under 50 people?
Yes, if one person inside will own it with hours set aside. A single facilitated day for a whole small company sits in the $3,000 to $8,000 market day-rate band and buys roughly 240 person-hours of practice, below the $51 per learning hour implied by ATD’s 2026 figures. Without an internal owner, no price makes it worth it.
Should a small company just buy everyone a subscription instead?
Buy the tool licenses, yes. A catalog learning subscription is a different purchase: seats at a published $399 per user per year cover 40 people for about $16,000 and schedule nothing. Mid-market buyers have someone to assign and curate; under 50 people nobody has those hours.
Why is the quote the same as for a company three times our size?
Because delivery cost is mostly fixed: preparation time and a facilitator’s day do not shrink with headcount. Training magazine’s 2025 survey shows the same gradient across all corporate training, at $1,091 per learner for small companies against $468 for large ones, in a sample that starts at 100 employees. Ask for a smaller scope, not a lower rate.
Do we need role-specific AI training with four people per function?
No. Separate sessions per role need enough people per function to fill a room. Below 50 people, run one mixed-function session with exercises that differ by role inside it. You keep the relevance and drop eight facilitator days.
What should a small-company program cost?
Workshop-scale engagements run roughly $2,000 to $15,000 in the market, depending on days delivered and preparation depth. Automation built afterward is a separate product in the $25,000 to $150,000 range. Treat a blended quote as two quotes.
Who owns this when there is no L&D team?
Someone with standing to change how work gets done and explicit hours to do it: usually the COO, chief of staff, or an operations lead. Name them before signing. A volunteer with no set-aside time is the most common reason a small-company program fades.
Does training include our AI licenses?
No. Seats for ChatGPT, Claude, Gemini, or Microsoft Copilot are contracted with the tool vendor and billed separately. Microsoft 365 Copilot lists at $30 per user per month on annual commitment on top of a qualifying base license, so at 40 seats the license line can exceed a one-day fee.
How soon should we expect to see anything change?
Judge it at four to six weeks, on work product rather than login counts. At this size you can read what people produced, a stronger test than any usage dashboard. If the output has not moved, the follow-up block is where you fix it.
Should we wait until we are bigger?
The federal data argues against it. Census figures put AI use at 37% among firms of 250 or more and under 20% below 20 employees, rising for larger firms and flat for the smallest. Waiting means the gap grows while you do.
Key takeaways
- A facilitator day has a fixed cost, so per-head figures rise as the group shrinks. Training magazine recorded $1,091 per learner at small companies against $468 at large ones, in a sample that begins at 100 employees.
- Scheduled practice on your own documents, a named owner, measurement, and tool licenses scale down untouched. Measurement improves, because you can read everyone’s work.
- Per-cohort fees, role-group curricula, and phased waves invert below 50 people. Collapse the roles into one room and run it once.
- Catalog seats scale down on price and not on value: the person who would assign and curate them does not exist at this size.
- Strike the paid assessment, the governance scaffolding, the open-ended retainer, and any platform nobody can run.
- No owner with real hours means no purchase. Fix that first. It is free.
If you have a mid-market proposal in front of you and a team of 40, bring it to a free AI-readiness call. We will name the line items we would strike at your size, the hours worth paying for, and what one room running on your own documents would cover.
Arthur Wandzel
