Every enterprise AI enablement playbook ranking for this topic assumes a company four times your size. A governance board. A learning team with its own headcount. An LMS somebody already administers. A roadmap measured in fiscal years. Delete those four and about half the playbook stops working, while the other half gets cheaper and faster.
Nobody publishes which half is which, because the expensive half is where the margin lives. So here it is, with the vendor thresholds and the size data that explain why the two playbooks separate at all.
What the enterprise AI enablement playbook quietly assumes
Correlation One’s enterprise playbook is designed to run from 50 learners to 20,000, with four modules, a prompting framework and a maturity model. The teaching content holds up. The architecture around it is built for the 20,000 end of that range: cohort waves, facilitator cascades, regional scheduling, a maturity ladder that needs enough teams to sit on different rungs.
Five assumptions run through almost every page in this category:
- A dedicated owner. Someone with AI in their job title and hours protected for it.
- A governance body. A council that meets, reviews use cases, and issues approvals.
- A delivery channel. A learning management system, a curriculum team, and somebody who chases completions.
- A long horizon. Phase one in Q1, measurable capability by Q4 next year.
- Buying power. Enough seats that vendors discount and bend terms.
At 50 to 500 people you have none of the first four and a damaging half-version of the fifth. The work still has to happen, but the structure that carries it cannot be the same, and copying it anyway is how these budgets get wasted. The underlying argument for treating capability rather than seat count as the deliverable is made in full in the case against buying licenses and calling it progress.
The size data behind the divergence
Three published datasets, none of them vendor marketing, show the split is real.
The Census Bureau’s Business Trends and Outlook Survey, published May 26, 2026, breaks AI use out by employment band. For the period ending May 3, 2026, 37% of firms with at least 250 employees reported using AI in their operations, against 32% of firms with 100 to 249 employees and under 20% of firms with four or fewer.
McKinsey’s state of AI survey for 2026, fielded May 4 to June 8 with 1,719 respondents across 97 nations, puts a second number on it: 54% of organizations with at least $1 billion in annual revenue report scaling AI across the enterprise, against about one-third of smaller organizations. Same tools, same prices, a 20-point gap in whether anything scaled.
The third dataset says the gap is not about enthusiasm. BCG’s AI at Work study from June 3, 2026, surveying close to 12,000 employees, managers and leaders, found 74% of frontline employees now call themselves AI users, and 72% say the skills expected of them have shifted. Only 36% feel they got adequate upskilling, a number that has not moved since 2025.
Your people are using the tools and know the bar moved. The structure around them is missing, and what the market sells is sized for a company four times larger.
Which half of the playbook transfers down
The cut that matters separates practices that are good because they are right from practices that are good because you can afford them. The first group transfers down intact and usually costs nothing. The second group breaks.
| Enterprise practice | At 50 to 500 people | Why |
|---|---|---|
| Baseline the work before training | Keep, non-negotiable | Two hours of timing real tasks, and the only thing that makes an after-number mean anything |
| Build exercises on your own documents | Keep | Works better at small scale, because one cohort can share one real workload |
| Brief the managers who assign the work | Keep | Six managers in a room beats 300 on a webinar. This is where small size is an advantage |
| Write the method down per task | Keep | Your version of the enterprise playbook’s curriculum, at one page per task |
| Role-based curricula | Trim to two or three roles | Nine tracks needs nine audiences. You have three that matter |
| Named internal champions | Keep, but one, not a network | A champion network is a staffing model. At your size it is one person with four protected hours |
| Governance council | Replace with a one-page policy and a named decision-maker | A five-person council in a 180-person company is a meeting, not governance |
| LMS rollout and completion tracking | Cut | Attendance data about 200 people you could ask directly |
| Train-the-trainer cascade | Cut | Cascades exist to cover headcount you do not have |
| 18-month phased roadmap | Cut to one quarter | Sponsor, budget and tool stack all change before month 12 |
| Maturity model | Cut | Built for reporting across business units. You have one |
Four clean keeps, three things to shrink, four to cut, and the keeps are the cheap ones. In the sessions we run, what separates a team still using the method six weeks later from one that drifted is almost never the teaching; it is whether a baseline existed and whether the managers were in the loop. Both sit on the keep list and neither costs real money.
Four enterprise habits that backfire under 500 people
Some of this is worse than overhead. Four patterns actively cost you ground.
The pilot that never ends. Enterprises run pilots because the blast radius of a bad rollout is thousands of people. At 180 people the pilot is the rollout. A six-month proof of concept burns the sponsor’s attention to produce a finding you could have had in three weeks.
Waiting for the governance to land first. The sequencing advice in enterprise frameworks is policy before training, which is correct when a regulator is reading over your shoulder. At your size it usually means nine months of legal review while people keep pasting client documents into whatever tool they opened. A one-page acceptable-use policy issued in week one, then taught inside the sessions, protects you faster than a framework nobody finished.
Buying the platform before the use case. Enterprise budgets absorb a tool that turns out wrong. Yours does not. Name the task first, then pick what fits it.
Measuring adoption instead of output. Seat use and weekly active users are enterprise metrics because enterprises cannot see individual work. You can. Ask what three tasks changed and look at the output. The measurement side of this, including what to record before session one, sits in our guide to moving from license rollout to measurable team capability.
The procurement asymmetry nobody prices
This is the part no page in either camp mentions, and it explains why enablement often costs more per head at your size than at the companies above or below you. Catalog vendors set their self-serve ceilings well under your headcount. Checked against vendor pages in October 2026:
| Vendor | Self-serve list price | Self-serve seat range | Above that |
|---|---|---|---|
| Coursera Teams | $399 per user per year | 2 to 499 licenses | Contact sales |
| LinkedIn Learning for Teams | $379.88 per license per year | 2 to 20 self-serve, 50 maximum per account | Enterprise sales |
A 120-person company that wants everyone trained is past LinkedIn’s self-serve ceiling six times over. It lands in a quoted, sales-led motion with none of the buying power that gets an enterprise buyer a third off list: the enterprise purchase process and the small-business negotiating position at once.
Two consequences follow. First, published per-seat prices are the wrong basis for your budget, because you will not transact at them in either direction. Second, per-cohort pricing beats per-seat at your size for arithmetic reasons, not philosophical ones.
A cohort price does not scale with headcount, so the 40 people whose work has to change carry the engagement instead of 200 logins carrying it. The three billing shapes, and when each is cheaper, are in our breakdown of per-seat, per-cohort and retainer pricing.
Hands-on programs on a team’s own work are quoted per cohort in a market range of roughly $2,000 to $15,000, with custom automation that follows running $25,000 to $150,000. The full model, including internal hours nobody budgets, sits in our cost breakdown for these programs.
What AI enablement looks like at 50 to 500 people
Compress the enterprise playbook to one quarter and one cohort, and it becomes something a COO can run between board meetings.
Weeks one and two: pick two or three tasks that a named team does every week and that someone complains about. Time them. Write down what today’s version produces. Issue the one-page policy on what may and may not go into a model.
Weeks three and four: run the hands-on sessions on those tasks, with the team’s own files, not a sample dataset about a fictional company. Brief the managers separately in an hour. They need to know what good model-assisted output looks like, because they are the ones who will accept or reject it.
Weeks five and six: nothing scheduled except the work. This is where enterprise programs lose people to the next mandatory module, and where BCG’s other finding bites: 66% of AI users in that study get limited or no guidance on what to do with the time the tools free up. Your version of guidance is a manager saying which task is now done the new way.
Weeks seven and eight: re-time the same tasks. Keep what held, drop what did not, write the method down for the two that worked. Then decide whether the next cohort or a custom automation project earns the following quarter’s budget.
Eight weeks, one owner, two or three tasks, a number at each end. That is the whole program. Which delivery option fits is a separate question, and we compared the six realistic ones in our review of options for companies between 50 and 500 people.
Who owns it when nobody owns it
The enterprise answer is a function. Your answer is a person, and the choice matters more than the vendor.
Pick whoever already owns the process you want changed and give them stated hours, not a title. A COO or chief of staff with four hours a week and the authority to change a workflow beats a committee of eight. The failure at this size is rarely the wrong owner; it is an owner handed the job with no hours carved out of the one they already have.
Hire or retain comes down to how much enablement work the next two years hold. We ran the year-one and two-year numbers in our comparison of hiring an AI lead against retaining a partner.
Frequently asked questions
What is enterprise AI enablement?
Enterprise AI enablement is the practice of aligning strategy, data, tools, workforce skills and governance so AI gets used consistently across a large organization rather than in pockets. At enterprise scale it is a standing function with its own staff, and the AI enablement services sold into that market are priced and structured for one. At 50 to 500 people the same outcome has to come from one owner, one cohort and one quarter, which is why the published playbooks do not transfer whole.
Is mid-market AI enablement just the enterprise version with less money?
No. Four things change at once: the horizon shortens from years to a quarter, every dollar has to return, there is no AI team to hand the work to, and the binding constraint is one person’s calendar rather than cross-silo coordination. Shrinking an enterprise program keeps its overhead and loses the thing that made it work, which was dedicated staff.
Do we need an AI governance council at 200 people?
Not a council. You need a one-page acceptable-use policy that says which tools are approved and what data must never be pasted into them, a named person who decides edge cases, and five minutes of that policy taught inside every session. A standing committee at this size produces meeting notes, not control.
How long should a program take at our size?
One quarter end to end, with the sessions in the middle four weeks. If a proposal spans twelve months, ask what the first eight weeks produce on their own. Anything that cannot show a before-and-after on two or three real tasks inside a quarter is priced on hope.
Why do published per-seat prices not help us budget?
Because your headcount sits above most vendors’ self-serve ceilings and below the volume that earns a discount. Coursera’s self-serve Teams plan runs to 499 licenses and LinkedIn Learning’s to 20 with a 50-license account cap, so a 150-person company negotiates off a sales-led quote with no volume behind it. Use per-cohort math for the group whose work has to change instead.
What should we measure if not adoption?
Two or three named tasks, timed and sampled before the first session and again six to eight weeks after. Usage dashboards tell you people logged in. A re-timed task tells you the work changed. At your size you can look at the output directly, which is the one measurement advantage you hold over a 20,000-person company.
Can we run this without an LMS?
Most companies at this size should. An LMS solves a distribution problem created by headcount. With one or two cohorts the distribution channel is a calendar invite, and the artifacts that matter (a written method per task and a prompt library) live in whatever document tool your team already opens.
Does any of the enterprise playbook survive intact?
Four parts do, and they happen to be the cheap ones: baseline the work before you train, build the exercises on your own documents, brief the managers who assign and accept the work, and write the method down per task. Those four are the engine. Councils, cascades, maturity models and multi-year roadmaps are the chassis built to carry thousands of people.
Our vendor pitched a role-based curriculum across nine functions. Worth it?
Rarely at this size. Nine tracks assumes nine audiences large enough to fill a cohort each. Pick the two or three functions where model-assisted work would change an output someone outside the company notices. Add tracks once those have numbers behind them.
We already bought licenses and training. Is that sunk?
Not at all. License telemetry tells you which teams never logged in. Cut the dead seats, keep self-starters on catalog access, and move the difference to hands-on work with the one or two teams whose output has to change this quarter.
Where to start
Take the framework your vendor sent and mark every line against the four-keep list. If the proposal is mostly councils, cascades, maturity stages and a twelve-month phase plan, you are being sold the chassis of a program whose engine runs in eight weeks.
Then name the tasks, name the owner, and put a number on them before anyone teaches anything. Everything else here is negotiable; that is not.
To pressure-test a proposal with someone who has no seats to sell you, book a free AI-readiness call. Thirty minutes, your own headcount and tasks, and a straight answer on which half of the playbook you should be paying for.
Arthur Wandzel
