The company that bought 400 seats usually has less AI capability per person than the company that bought 25. Not because the big buyer did something careless, but because breadth and depth compete for the same attention, the same budget cycle, and the same four weeks of organizational patience. This is the most expensive of the AI enablement problems, and the only one that gets worse the more you spend on it. The mechanism is structural, it repeats across companies that have nothing else in common, and once you can name the five parts you can reverse most of them inside a quarter.
What the AI enablement paradox actually claims
The weak version of this claim is that license count and capability are unrelated, which nobody disputes and nobody acts on. The strong version is that they trade off, and that is the one worth arguing.
BCG put the cleanest evidence behind it. Its fourth annual AI at Work survey, published June 2026 across close to 12,000 frontline employees, managers and leaders in more than a dozen markets, found that employees with clear direction outperform those with greater access to AI tools, and that an explicit plan raised impact even where access was limited. Read that sentence as a procurement instruction and it says the opposite of what most 2026 budgets did.
The distribution tells the same story from the other end. Section’s AI Proficiency Report tested 5,026 US knowledge workers on hands-on tasks rather than asking them how they felt, and found 5.5 percent at practitioner or expert level against 73.5 percent stuck as experimenters and 20.9 percent as novices. A workforce of permanent experimenters is precisely what an unaccompanied license rollout produces. Everyone has the tool. Almost nobody has a method.
The purchase order closes when the seats light up
Procurement is built to finish. A software purchase has a defined end state: contract signed, seats provisioned, admin console green. The moment that state arrives, the project is done by every definition the company uses to track projects.
Nothing in that document describes a capability. There is no acceptance criterion that reads “sales can draft a first-pass proposal in nine minutes,” so nothing after provisioning is scheduled, assigned, or reviewed. The work that would close the distance between access and skill is exactly the work with no line in the purchase order, which is why it reliably does not happen. AI enablement names that missing work, and the reason the term had to be invented at all is that the invoice does not contain it.
Smaller rollouts escape this by accident. Twenty-five seats is too small to feel like a project, so it stays an open question, and open questions get revisited. Four hundred seats feels like a delivery, and deliveries get closed.
Universal access destroys the group you could have taught
Give a tool to nine people and you have a cohort. You know their names, you can put them in a room, you can compare what one found to what another missed, and six weeks later you can ask each of them what changed. Give it to 400 and you have a population: a thing you describe with averages and address with an email.
The teaching unit is the cohort, not the population. In the sessions we run, the usable work almost always comes from people watching a colleague who does their exact job solve a problem they both recognize, and that only happens when the room is small enough for the problems to overlap. Section’s numbers make the same point from the data side: workers at organizations that ran agent and automation training scored 44 out of 100 against 25 at organizations that ran none, and workers whose managers expected and required AI use scored 1.5 times higher. Both conditions are easy at nine people and expensive at 400.
Broad access also removes the signal of who needed help. When a narrow group has the tool, the people struggling are visible by the second week. When everyone has it, struggle looks like silence, and silence looks like adoption.
The recurring line is set before the one-time line exists
Seat costs are recurring, per-person, and quoted during a renewal conversation with a vendor who has a number ready. Enablement costs are one-time, per-cohort, and quoted by a party nobody has called yet. The two numbers are set in the wrong order, and the first one eats the room.
By the time anyone asks what training would cost, the recurring commitment is already signed for twelve months and the answer is “next fiscal year.” The imbalance is not usually a matter of a few percent either. A full-company seat commitment is a five- or six-figure annual line depending on headcount, while hands-on fluency workshops sit in a market range of roughly $2,000 to $15,000 and are paid once. The cheaper line is the one that gets deferred, because it was never on the same page as the expensive one. The license-versus-enablement arithmetic is worth running before the renewal rather than after it.
BCG found 72 percent of employees saying expectations for their skills had shifted while only 36 percent felt they had received adequate upskilling. That is not a training-budget failure. It is a sequencing failure.
Breadth turns a list of names into a rate
At 25 seats a leader knows, by name, who uses the tool well. At 400 that knowledge is impossible, so it is replaced by a percentage, and a percentage hides the one thing you needed to see.
Microsoft’s own Copilot usage report documentation is admirably precise about what the percentage contains. Enabled Users counts license holders. Active Users counts enabled users who performed one intentional action on an AI feature in the selected window, and the docs specify that opening the Copilot pane in Word does not count while submitting a prompt does. Active users rate divides the second by the first. The window is selectable at 7, 28, 90 or 180 days.
One prompt in six months, therefore, reads as active. That is not a flaw in the instrument, it is the instrument working as documented, and the consequences for reporting are covered in more depth in why usage dashboards overstate capability. The point here is narrower: the bigger the rollout, the more decisions you make from a rate, and a rate cannot distinguish a team that redesigned its weekly reporting from a team where one person tried the summarize button in May.
A tool nobody asked for has no advocate inside the building
Software that arrives in a bundle has a different social life than software someone fought for. A requested tool comes with a person who staked their standing on it, who will answer questions in the hallway and defend it in a meeting. A bundled tool arrives with an IT announcement and belongs to no one.
This is where the ownership problem that shows up on every list of AI enablement problems actually originates. It is not that nobody can be found to own it. It is that broad distribution produces no natural owner, because the people who would have become advocates never had to choose. Section’s testing found workers at organizations with a full-time head of AI averaging 42 out of 100 against 33 without one, and only 5.5 percent of the whole sample reaching practitioner level. Someone has to be accountable, and at 50 to 500 people that job lands on a person who already has one.
BCG’s version: only 28 percent of employees saw a strong connection between what leaders said about AI and what the organization did. Breadth without an advocate is how that gap opens.
Where buying breadth is the right call
The honest counterweight: narrow rollouts have their own failure mode, and it is worse in one specific way.
A nine-person pilot concentrates the company’s AI capability in nine people. Some of them will leave, and the method walks out with them unless somebody wrote it down. That is the trap described in the four levels of AI capability: capability that lives in individuals rather than in a documented method is a liability dressed as a success.
Breadth is also correct when the tool is genuinely ambient, when a compliance position requires a sanctioned alternative to whatever people are already pasting into consumer accounts, or when the per-seat cost is low enough that the question is not worth a meeting. Section’s data supports the sanctioned-access case directly: workers with clear, sanctioned tool access averaged 43 out of 100 against 25 with no sanctioned tools.
So the recommendation is not “buy fewer seats.” It is: stop treating seat count as the variable you optimize, and start treating the trained cohort as the thing you buy seats for.
Three divisions that locate your AI enablement problem
Three numbers, all available from the admin console and the general ledger, no vendor required.
| Division | What to compute | What it tells you |
|---|---|---|
| Seats per trained person | Total assigned seats ÷ people who have completed hands-on training on their own work | Above 20, you bought distribution and skipped capability |
| License dollars per enablement dollar | Annual seat spend ÷ total spent on training, documentation and facilitation | Above 20 to 1, the recurring line ate the one-time line |
| Best-team to worst-team spread | Active-user rate of your highest team minus your lowest, on a 28-day window | Under 15 points, your tenant average is hiding nothing. Over 40, it is hiding everything |
The cut points are our own rules of thumb from running these engagements, not published benchmarks, so treat them as a place to start an argument rather than a standard to pass. The third row is the most useful and the least computed, because it requires segmenting a number most companies only ever view in aggregate. Run it by department before the next renewal conversation. A 50-point spread is not an adoption problem, it is a direct map of where the method exists and where it does not.
How to spend the next dollar
Most corporate AI enablement budgets are built in the order the vendors quote them. Reverse that sequence. Decide which cohort you are building first, buy that cohort hands-on training on its own documents and recurring deliverables, measure a tested baseline rather than an activity rate, then widen the seat count behind the method once a method exists to widen. That ordering is what the AI enablement playbook lays out step by step, and the underlying case for why access and capability are different purchases is made in the argument against buying licenses and calling it enablement.
Not every company needs to spend more. Some need to spend the same money in the other order. If you want a second opinion on which of the three divisions above is the binding constraint for your team, a free AI-readiness call is a reasonable place to run the numbers with someone who has watched this go both ways.
Frequently asked questions
Does buying more AI licenses really reduce capability?
Not mechanically, but it reliably crowds out the work that builds capability. Breadth consumes the budget cycle, closes the project on provisioning, replaces named people with a usage rate, and removes the small group you could have taught. BCG’s 2026 survey found employees with clear direction outperforming those with greater tool access, which is the empirical shape of that trade-off.
How many seats should a 200-person company start with?
Start with the cohort, not the seat count. Pick one team of eight to fifteen whose weekly output is measurable, license that team, train it on its own documents, and let the result set the size of the next wave. The seat number is an output of that decision, not an input to it.
Our usage dashboard shows 70 percent active. Is that good?
It depends entirely on the window and the segmentation. One prompt inside the selected period is enough to count as active in Microsoft’s reporting, and the window can be set as wide as 180 days. Pull the same figure on a 28-day window, then split it by team. The spread between your best and worst team is the number that matters.
Why do smaller companies often look more AI capable?
Fewer approval layers, and the person who bought the tool is usually the person using it. A 30-person company has no gap between the budget holder and the practitioner, so the feedback loop that produces a working method closes in days. That advantage thins as soon as the person signing for the tool stops being a daily user of it, and after that it has to be rebuilt deliberately.
Is the problem training or workflow redesign?
Both, in that order. Training without redesigned work produces faster versions of the same output. Redesign without training produces a process nobody can execute. The practical sequence is to train on real deliverables first, because the redesign ideas that stick almost always come out of the people doing the work once they can see what the tool does well.
What does it cost to fix this?
Most AI enablement services quote hands-on fluency training in a market range of roughly $2,000 to $15,000 depending on cohort size and session count, and it is a one-time cost against a recurring license line. Deeper automation work on a specific workflow runs in the $25,000 to $150,000 range. The first number is usually a rounding error against a full-company seat commitment, which is exactly why it is worth computing the ratio.
Should we claw back unused seats?
Reclaim genuinely dormant seats at the 28-day mark for licensing hygiene, but do not treat the recovered spend as the win. The seats were never the asset. Redirect the recovered budget into the cohort you are training rather than back into the general pool, or the paradox simply repeats at a lower price.
Who should own this when nobody’s title says AI?
Whoever signs both the license renewal and the training budget, which at 50 to 500 people usually means the COO, the CFO or a VP of Operations. The two numbers sit in different departments at most companies, which is the real reason nobody has computed the ratio between them.
Does the 95 percent failure statistic apply to us?
Treat it carefully. The widely quoted figure that 95 percent of enterprise AI pilots show no measurable return comes from MIT’s 2025 GenAI Divide work, which rests on 52 executive interviews, surveys of 153 leaders and analysis of 300 public deployments. That is a small, non-random sample. It is directionally useful and it is not a prediction about your company.
Key takeaways
- License breadth and per-person capability compete for the same budget cycle and the same organizational attention. The company with the most seats is frequently the one with the least method.
- Five structural causes: the purchase order closes on provisioning, universal access destroys the teachable cohort, the recurring seat line is set before the one-time training line exists, breadth replaces named people with a rate, and a bundled tool arrives with no internal advocate.
- BCG’s 2026 survey of close to 12,000 employees found clear direction beating greater tool access, and 72 percent reporting shifted skill expectations against 36 percent who felt adequately upskilled.
- Section’s hands-on testing of 5,026 US knowledge workers put 5.5 percent at practitioner level or above and 73.5 percent at experimenter, which is the distribution an unaccompanied rollout produces.
- Run three divisions before your next renewal: seats per trained person, license dollars per enablement dollar, and the active-rate spread between your best and worst team. The third is the one nobody computes and the one that tells you the most.
Arthur Wandzel