Two numbers sit on the same desk every renewal season. One is an LLM subscription that bills whether anyone opens the app or not. The other is an AI enablement budget somebody keeps deferring, because the seats were supposed to be the answer. The return comes from neither line alone, and the arithmetic that settles which to fund next is short enough for a napkin.
What each line on the invoice buys
The subscription buys access and a legal posture: a tenant where confidential work stays out of model training under the vendor’s terms, an admin console, single sign-on, retention controls, a usage dashboard. Real goods, and non-negotiable for most firms of 50 to 500 people once anyone has pasted a contract into a chat window.
What it does not buy is the habit of using the thing on real work. No console setting produces that. The dashboard shows you exactly how many people are ignoring what you bought, which is information, not a fix.
An enablement budget buys the other half: a facilitated session where a team works on its own documents, its own reporting cycle, its own ugly recurring task, until each person leaves with two or three prompts they will use on Monday. The two lines get treated as substitutes when they are complements, a confusion we take apart in the broader case for treating license spend and team capability as two different problems.
The seat price is not the price per user
Start with published numbers. OpenAI lists ChatGPT Business at $20 per user per month billed annually, $25 billed monthly, two-seat minimum. Microsoft 365 Copilot lists at $30 per user per month on an annual commitment, an add-on sitting on top of a qualifying Microsoft 365 license rather than replacing it. ChatGPT Enterprise has no published price: it is sold as a prepaid annual contract, and third-party reporting of 2026 deals clusters around $45 to $75 per seat per month with roughly a 150-seat floor. Treat that band as reported deal data, not a vendor price, and get your own quote before it goes in a model.
Now the division nobody does. Take a 180-person company with 150 seats:
| Line | Nominal per seat, per year | 150 seats, per year |
|---|---|---|
| ChatGPT Business, annual | $240 | $36,000 |
| Microsoft 365 Copilot, annual | $360 | $54,000 (plus base licenses) |
| ChatGPT Enterprise, reported midpoint | $720 | $108,000 |
Then apply the largest public read on how many seat holders use their seat. Recon Analytics surveyed more than 150,000 US paid AI subscribers between July 2025 and January 2026 and published it under the title licenses do not equal adoption. Among people with workplace access, ChatGPT converts 83.1% into active users, Copilot 35.8%, and Gemini a near-identical 34.0%.
Divide the invoice by users instead of seats and the price changes:
- 150 Copilot seats at 35.8% conversion is 54 active people. $54,000 over 54 is $1,000 per active user per year against a $360 sticker.
- 150 ChatGPT Enterprise seats at the reported $60 midpoint, using the 83.1% Recon measured for ChatGPT as a whole rather than the Enterprise tier alone, is 125 active people. $108,000 over 125 is $864 per active user against $720.
That spread is the whole argument. On the weaker-converting tool you pay nearly three times list price per person who gets anything from it, and no amount of procurement skill touches that multiple, because it is not a pricing problem.
Where the return actually sits
Return on an AI budget is a product of three terms, not a sum:
seats paid for × share of seat holders doing real work in the tool × hours each one gets back
Money buys the first term, cleanly and instantly. The second and third have to be built, and they are where every dollar of return lives, because a seat nobody opens multiplies the whole expression to zero.
BCG’s 2026 AI at Work survey of close to 12,000 employees and managers across more than a dozen markets sizes the second term. Among people who got more than five hours of training, 79% were regular AI users, against 67% below five hours. Only 36% say they have had adequate upskilling at all. Five hours is a long morning, well inside the 16.7 formal learning hours per employee that ATD’s 2026 State of the Industry report recorded for 2025 across 340 organizations.
One caution on that 12-point gap: it is an association across a survey population, not a projection you can bolt onto your tenant. It says the lever exists and roughly how long the handle is.
The dormant seats, in dollars, and the limit of that argument
Do the multiplication the adoption roundups leave out. In the Copilot case, 96 of 150 seats are dormant. At $360 each, that is $34,560 a year producing nothing: recurring, already approved, invisible on any report your leadership team reads. In the Enterprise case, 25 dormant seats at the reported midpoint is $18,000.
Here is where vendor pages overreach, so let us not. Suppose a program moves conversion by something like BCG’s 12 points. On 150 seats at $360, that is 18 seats switched from dormant to used, worth $6,480 a year. Real, recurring, and nowhere near enough to carry a program in year one.
So reclaimed license spend is not the business case. It is what makes the conversation start in the right place, because it turns an abstract ask into a number finance already has in a system they trust. The case rests on hours returned, with attendee time in the cost base, a discipline worth borrowing in full from the way to underwrite this budget so a CFO signs it.
What the program costs, at market rates: facilitated sessions run roughly $2,000 to $15,000 each depending on format, and a multi-cohort rollout works out to $150 to $400 per trained person, so $22,500 to $60,000 for 150 people. Against a learning budget that already exists, the ask is smaller than it sounds. At ATD’s $846 per employee, a 180-person company already spends around $150,000 a year on workplace learning. This is a reallocation inside that number, not a new line.
One line recurs forever, the other is paid once
The asymmetry between the two budgets decides more than the per-user math does.
A seat subscription is permanent and grows with headcount. It is usually prepaid annually, often with a seat floor, so the year you discover your conversion rate is the year you cannot act on it. Over three years, 150 Copilot seats is $162,000 before a single new hire.
A program is mostly one-time against that stream, with a small recurring tail for new joiners. So the real comparison is not program versus seats. It is one-time spend that raises the second and third terms against recurring spend that only raises the first. Where conversion is low the choice is obvious, on the same reasoning that separates a capability program from a multi-year change effort in the difference between AI transformation and enablement.
Why nobody sets the ratio between your license and AI enablement budgets
The organizational reason the money sits in the wrong line is dull and fixable.
The seat bill lives with IT or procurement, renewed by someone measured on unit cost and security posture. The capability spend lives with L&D, or with whichever functional VP got handed AI as a second job, on a different calendar and approval path. Neither owner sees the ratio, so nobody sets it. Procurement optimizes seat price, L&D optimizes cost per attendee, and the figure that would tell you something belongs to no one.
The fix is one meeting and one number. Pull the active-user count from the admin console, divide annual contract value by it, and put that figure in front of whoever signs both. In rooms we facilitate, what changes the budget conversation is almost never a slide about model capability. It is someone from finance seeing the per-active-user figure for the first time.
What the AI enablement your vendor includes actually covers
Vendor enablement is useful and is not the same product as the AI enablement services a specialist sells. Microsoft Learn, OpenAI’s published guides, and the webinars bundled with an enterprise agreement teach the product surface well: where the features are, what the admin controls do, how to set up a shared workspace.
What they do not do is sit with your eight-person finance team while it rebuilds the monthly close memo in a model, argue over which parts of that memo must never be drafted by a machine, and leave each person with a prompt that survives next month’s version. Vendor material is written for every customer, so it cannot be about your documents, and yours is the only context in which fluency transfers. That is also why a cheap generic program and a real one look identical in a proposal and nothing alike in a room, a gap worth reading before you compare quotes in what a cheap AI enablement vendor actually costs you.
When buying more licenses is the right call
Three cases where seats are the better next dollar, stated plainly because no seller of either line will tell you.
Conversion is already high and people are rationing. If 80% of seat holders are active and the complaint is usage limits or shared logins, buy seats. The second term is solved and you are constraining the third.
Nobody has a tenant yet. Training a team on consumer accounts is worse than not training them, because the first thing they learn is a habit your general counsel has to unwind. Buy the business tier, then train.
You cannot move the second term. If a previous attempt failed and you have no sponsor, no protected time, and no manager clearing calendars, a program will fail the same way. Cut seats at renewal to match real usage instead. Annual prepayment and seat floors block mid-term correction, so put the renewal date on your calendar now.
How to split the next dollar between seats and AI enablement
A sequence, not a formula.
- Get the active-user count per tool from the admin console, weekly actives rather than anyone who logged in once.
- Divide annual contract value by that count. That is your real unit price. Compare it with the sticker.
- Above about 2x, the next dollar is enablement. You carry more dormant seats than used ones, and buying more makes the ratio worse.
- Near 1x, the next dollar is seats or scope. The constraint has moved to coverage or to the third term, which is a workflow question rather than a fluency one.
- Size the program against license spend, not a learning benchmark. A planning range of 15% to 25% of annual license and tooling spend sanity-checks a quote, putting a $54,000 seat bill at $8,100 to $13,500 per wave.
- Time three recurring tasks before anyone sits down. Without a baseline you will argue about feelings at the renewal, and feelings get discounted. The full method sits in our playbook for getting from license rollout to measurable team capability.
Pull two numbers, annual contract value and weekly active count, and bring them to a free AI-readiness call. Half an hour is enough to say which line your next dollar belongs in, including the cases where the answer is to cut seats rather than buy anything.
Frequently asked questions
Is AI enablement cheaper than buying more licenses?
Usually, and not close when conversion is low. A program covering 150 people at market rates lands around $22,500 to $60,000 once, while 150 Copilot seats at the published $30 per user per month is $54,000 every year, and the program also raises the value of the seats you already have.
Can I pay for AI enablement out of the software budget?
Yes, and that is usually the easiest path. You are protecting a recurring commitment rather than asking for a new initiative, and the approver who signed the seat contract already owns the problem. The learning budget is the fallback.
What is a reasonable split between license spend and enablement spend?
Plan on 15% to 25% of annual license and tooling spend, then check it against role count. A program for one function costs a fraction of one that has to work for finance, sales, and operations separately.
Does ChatGPT Enterprise include training?
It includes onboarding support and OpenAI’s published guides, which cover the product surface. It does not include a facilitated session on your team’s own documents and recurring tasks. Read the order form closely, because “onboarding” and “enablement” are used loosely by every vendor here.
How many of our seats are probably dormant?
Check rather than guess: the number is in your admin console. For calibration, Recon Analytics found 35.8% of people with Copilot access actively using it against 83.1% for ChatGPT, so a tenant with two thirds of seats quiet is ordinary rather than a crisis unique to you.
Should we just cancel the unused seats instead?
If you have no sponsor and no protected time, yes, and nobody selling either line item will say so. Match seat count to weekly actives at renewal and revisit later. Prepayment and seat minimums make this a renewal-window decision, so put the date on your calendar.
Is the free vendor training enough?
Enough to learn the interface. Not enough to change how a team works, since it cannot reference your documents, your reporting cycle, or the judgment calls your industry requires. Use it for mechanics; spend money on the applied layer.
We already ran a vendor demo day and nothing stuck. What was missing?
Three things, almost always: work on the team’s own files rather than sample data, a named task each person commits to doing differently, and a manager who clears the time. A demo shows what a tool can do. What changes Monday is a person leaving with a prompt and a task.
Does moving everyone onto one tool fix adoption?
Standardizing helps support and billing more than adoption. Recon Analytics found preference varies sharply when people have a choice, so forcing one tool can suppress the usage you want. Pick one for security and billing, then earn the usage.
Key takeaways
- An LLM subscription buys access and a legal posture. An enablement budget buys the habit of using it. Complements, budgeted as substitutes.
- Return is a product of three terms: seats, share of holders doing real work, hours each returns. Money buys the first term only.
- Divide the contract by weekly active users, not seats. At the measured 35.8% Copilot conversion, a $360 seat costs $1,000 per person who uses it.
- Price dormant seats in dollars. On 150 Copilot seats, 96 quiet ones is $34,560 a year of approved, invisible spend.
- Do not oversell reclaimed license spend. A 12-point conversion gain on that invoice is about $6,480 a year, which starts the conversation rather than winning it.
- Seats recur and scale with headcount: $162,000 over three years for 150 Copilot seats, before a single new hire. A program is paid roughly once.
- Nobody owns the ratio, because the seat line sits with IT and the capability line with L&D. Cost per active user fixes that in one meeting.
- Sometimes the answer is fewer seats. With no sponsor and no cleared calendars, cut the count at renewal instead of buying training that fails the same way.
Dirk Jan van Veen, PhD
