Author
TL;DR: An enterprise license agreement is a licensing decision, not a discount depth. It replaces a counted metric with an uncounted grant across a defined population for a defined term, which means the counting moves to the boundary: who is inside the population, what “unlimited” is quantified over, and what happens when the term ends. For decades the construct was safe because headcount is a bounded metric. Products carrying generative AI broke that boundedness, and AI agents sit outside a population defined in human terms even when the grant says unlimited. The census, not the discount, decides how the deal ages.
The request lands on the largest renewal of the year. Procurement wants one number for the whole company, with the counting removed. The deal team produces it: tally current usage, project growth, apply the deepest discount the approval matrix allows. Everyone in the room prices the number. The agreement that follows also carries three definitions, and nobody prices those, because nobody reads them as pricing.
They are the pricing. An enterprise license agreement conveys broad or unlimited use of software across a defined population for a defined term. The fee is arithmetic performed on those definitions. When the definitions are inherited from a template, the arithmetic is precise and the deal is not.
An Enterprise License Agreement Replaces the Count With a Census
An enterprise license agreement is a license grant conveying broad or unlimited use of software across a defined population for a defined term. Vendor catalogs use several names for the construct: enterprise agreement, site license, organization-wide license. Under every label, the four dimensions a software license grant settles are still settled. Assignment widens from named individuals to an organization. Boundedness moves from a meter to a boundary. Duration becomes the only counter still running.
The metric has not been removed. It has been relocated. A counted agreement measures usage inside the population; an enterprise agreement counts the population itself, once at signature and again at renewal. That is a census.
The industry files the ELA conversation under pricing, as the deepest row of the discount schedule. Backwards. The discount is output; the census is the first of the three decisions behind every software price, remade at account scale, usually by whoever drafted the template. In an enterprise agreement, the census is the contract.
The Three Definitions That Are the Deal
Who Is Inside the Population
“Enterprise” reads as self-evident until an event tests it. Employees, presumably. The offshore contractors who log in every day? The subsidiary consolidated last year but operating on its own systems? The company your customer acquires in month seven, doubling the census the fee was computed on? Divestiture asks the same question in reverse: whether the grant follows the unit that leaves. Each answer is a licensing decision. The assignment variations that decide who may hold a seat do not disappear under an enterprise grant; they compress into the population definition. And when an affiliate embeds the product or passes it onward rather than using it, the question exits the ELA entirely and becomes a channel or OEM grant, denominated in someone else’s units.
What “Unlimited” Is Quantified Over
Unlimited use of what? The products named at signature, every edition of them, the capability shipped in month nine? A vendor that adds generative features mid-term has changed its own cost structure inside a grant that may or may not include the change. Whether new capability rides inside the enterprise grant or sells beside it is a packaging boundary, and it needs drawing before the term starts. Drawn afterward, every line reads as taking away something the customer believes they already own.
A customer of ours ran into the sharpest version of this. Their agreement was written for an on-premise product, and it carved out that any new product substantially similar in capability would fall inside the unlimited grant. That clause was reasonable on the day it was signed, because everything that could plausibly be built then carried the same cost shape. Then the cloud edition shipped, on an entirely different cost architecture, and the customer argued the obvious: it is substantially similar, so our unlimited use extends to it.
The clause did the job it was written to do. What it could not do was anticipate that capability similarity and cost similarity would stop travelling together. A future-products provision quantified over what the software does will keep working right up until someone ships the same capability on a cost base that behaves nothing like the old one, which is precisely what the cloud did to on-premise and what generative features are doing now to everything. If the provision has to exist, quantify it over the delivery and cost model rather than over the capability, and say plainly which side of the line a new cost architecture lands on.
What Happens at the Term Boundary
An enterprise agreement defers the counting rather than eliminating it, which makes the re-count the central commercial event of the construct: who takes it, from which system, and what both parties agreed the number would do once known. Silence on the boundary does not defuse that conversation; it reschedules it as an argument with the date left blank.
The widest gap we have seen came from a client who priced by territory. At signature the country was carved into four regions, northeast, southeast, northwest and southwest, so the count was four and the fee followed it. A new chief revenue officer arrived, consolidated the map into two regions, north and south, and asked for the licence fee to be halved. Usage had not fallen. It was still growing. The population running the software was the same population; only the label counting it had changed, and the customer had unilateral control of that label.
That is what a population definition risks whenever it is quantified over an organising construct rather than over something the vendor can observe. Headcount, seats and named users move for business reasons. Territories, divisions, business units and cost centres move because someone redrew a chart.
Populations are also managed downward on purpose, and vendors sometimes coach it. Buying a marketing platform years ago, we were asked how many contacts we held; the salesperson encouraged us to pick a low number, then noted that the count would be inspected a year later and suggested we clear contacts out shortly before the inspection. At an HRIS company we worked with, customers routinely trimmed their monthly employee counts just before the billing date. Both are usually described as customers gaming the metric. They are better read as a metric that can be gamed, which is a design problem the vendor owns.
Two questions separate a sound population metric from a fragile one. Can the customer move the number without changing anything about their business, and can you see the number move. Contacts and month-end employee counts fail both. Named users with assigned credentials fail the second less often, because the assignment leaves a trace in your own system.
The most fragile version charges over a population while assuming some average share of it will use the product. Nobody is average. Every customer’s real share differs from the assumption, and the ones below it arrive at renewal with a well-prepared argument that they are paying for people who never signed in. The assumption never appears in the contract, and the whole model rests on it.
Can Your ELA Survive a Workforce Composition Audit?
When a contractor, affiliate, or subsidiary event hits, your population definition either holds or collapses. Find out which of your licensing, packaging, and pricing decisions leaves you exposed before a renewal forces the answer.
Why Unlimited Was Safe for Decades
The enterprise grant was a sound trade for most of the industry’s history, and the reason is structural. A flat commitment is only as risky as the metric beneath it is unbounded. Headcount is a bounded metric: it moves at the pace organizations hire, both sides can watch it move, and a person working a day produces at most a day of usage. Physical location is bounded the same way, and more slowly. A new site means approvals, land, construction, and staffing, so a vendor priced by location can see the expansion coming for a year or more before it arrives on an invoice. Both metrics share the property that made the construct safe: the customer cannot move them quickly, and the vendor can watch them move. Unlimited use by a bounded population is a ceiling with a generous name. The vendor traded expansion revenue for commitment. The customer paid for predictability. Both sides could see the maximum from where they signed.
What Generative Capability Broke
Generative capability broke the half of the trade nobody was watching. When the product carries LLM inference, cost scales with consumption rather than with the census, and consumption per member is no longer capped by human working hours. One member of the population can direct more compute in an afternoon than a full department consumed under the old product. The paper stayed the same while the metric under it stopped being bounded. A flat grant over a bounded metric is sustainable by construction. A flat grant over an unbounded one is margin exposure with a signature on it. Generative AI ended the boundedness that made unlimited enterprise grants safe.
Why an Unbounded Grant Hardens
The other half of the risk is that these grants are difficult to reverse. A client of ours brought in a chief revenue officer who converted the customer base to all-you-can-eat licenses and beat every sales goal that year. Years later the company worked out that its largest accounts, paying millions, should have been paying tens of millions against what they were actually consuming. Every attempt to correct the provision met the same answer: the customers threatened litigation rather than give the term back. Those accounts are still enterprise accounts, and there is no expansion path left in them. The only way to grow them now is to sell something else entirely.
An unbounded grant is not only a pricing error, it is one that hardens. The customer holds a contractual right and will defend it, the account stops growing at the moment the term is signed rather than the moment anyone notices, and the correction that would fix it is the one thing the paper forbids.
An Agent Is Not a Member of the Census
The population definition faces one more test, and it is the one arriving in deals now. An enterprise agreement defines its population in human terms: employees, contractors, named affiliates. An AI agent operating against the software is not a member of that population. Its usage does not sit inside the unlimited grant at a discount, because unlimited was quantified over people and the agent is not one. It sits outside the grant altogether.
That reframes the renewal. A customer running agents against an enterprise-licensed product is not consuming harder inside their rights; they are running usage the agreement never contemplated. The productive posture is the one that applies when agents collide with seat-counted agreements: a design conversation about licensing what is actually running, never a compliance confrontation. The boundary cuts the other way as well. Where the metric itself captures the agentic work and the inference cost underneath it, per conversation or per resolution, the construct contains the agent and no census question arises. The exposure belongs to grants denominated in people.
Tests to Run Against Your Own Enterprise Agreements
Can you state who is inside the population without opening the agreement? If nobody on your side can, the widest plausible reading is the one in force.
If your customer acquired a company tomorrow, does the grant answer what happens to the census? An account team improvising that answer mid-term is pricing under pressure.
Which capability shipped this year sits inside the enterprise grant, and which sells beside it? If the line was never drawn, the grant drew it for you, at unlimited.
If consumption doubled while headcount held flat, which line of the agreement responds? A construct priced on people has no line that moves with inference.
Is anything consuming the product that is not a person, and does the population definition contemplate it?
In the enterprise agreements we review, the population definition is routinely the section that has traveled furthest from the deal it was written for: inherited from a template, stretched by acquisitions, and read closely for the first time when something breaks. For a read on your own paper, talk to a pricing expert and describe who your largest grant covers.
An enterprise license agreement is the licensing model rewritten at account scale, and it ages exactly as well as its census. Take the census deliberately and the construct delivers what both sides bought: predictability. Inherit it from a template and the biggest account on the board runs on definitions nobody chose. Talk to an expert and describe what your enterprise agreements define today.
