AI brings back the licensing trilemma

English · Nederlands

Three cracked stone blocks standing in water, marked simple, flexible and predictable.

Most AI price lists I review manage two out of three: simple, flexible, predictable. Almost none manage all three. In 2011 I wrote about how the cloud forced software licensing to change. Reading it again in 2026, I could swap “cloud” for “AI” and send it to a client.

What I saw then

In 2011 every software vendor had to balance three conflicting demands in its pricing. Simple enough to explain. Flexible enough to fit different business models and use cases. Predictable enough that the price tracked the expected benefit and stayed consistent over time. The usual result was legal documents nobody read; Microsoft’s product use rights ran to 138 pages.

The cloud exposed the gap. At a Microsoft conference you could walk 20 metres from the on-premise booth to the cloud booth and find the same functionality at a far simpler price. Meanwhile pricing per server or processor core stopped making sense in virtualised environments. My question was how long vendors would allow both models to coexist, and what the licensing specialists, PC-Ware and Insight among them, would do next: merge with integrators or retreat into niches.

What is the same now

The trilemma is back, and each AI pricing model picks a different corner of it.

Seats are simple and predictable, but not flexible. Microsoft launched Microsoft 365 Copilot at $30 per user per month (Axios). Easy to budget, easy to sell through existing agreements. It strains when usage differs a hundredfold between users, and it has nothing to say when an agent does the work instead of a person.

Usage and credits are flexible, but not predictable. In June 2025 Cursor changed its $20 Pro plan from 500 fast requests to $20 of usage billed at API rates. Some users ran out after a handful of prompts; others were charged for overage they had not expected. The CEO apologised publicly and the company offered refunds (TechCrunch). The reason given was that newer models use more tokens per request. The cost was real. The customer still could not forecast it.

Outcomes are simple to explain and track value, but they are rigid. One definition of a resolved case or a completed task rarely fits every customer, and the definition becomes the contract.

Credits deserve a special mention. They are often presented as the simple option. In practice they move the complexity of a 138-page use rights document into a rate card: so many credits per action, per model, per feature. The customer still cannot work out what a year will cost.

The transition question is back too. When Broadcom took over VMware, it stopped selling perpetual licences and support renewals for more than a dozen products in December 2023 and moved customers to subscriptions (Redmond). That is one answer to my 2011 question about transition periods: none. It gives the vendor a clean book and the customer a hard deadline. Expect AI vendors to face the same choice when they retire seat licences for agent or outcome pricing.

And the partners. PC-Ware, renamed Comparex, was bought by SoftwareONE in a deal announced in 2018 to create a licensing specialist of roughly 5,500 people (ChannelWeb). So the answer then was consolidation, not niches. With AI the reseller question is sharper. A partner earns a predictable margin on a seat. On metered usage, that margin becomes as volatile as the customer’s bill, and on outcomes it may be hard to define at all.

What is different

In 2011 the new model was the simpler one. Cloud subscriptions replaced core counts and use rights with a monthly fee. In 2026 the direction has reversed: seats are the simple model, and the AI models replacing them are more complex. Vendors are asking customers to give up predictability, not gain it.

The cost base behaves differently as well. Cloud unit prices changed slowly, and mostly in one direction. AI inference costs fall fast per unit, but vary per task, and agents can multiply them. A vendor that promises a fixed price per unit takes a margin risk; a vendor that passes cost straight through hands the risk to the customer.

Buyers are better prepared. After a decade of cloud bills, many enterprises have FinOps teams and procurement rules for usage-based software. They will ask for caps, commitments and forecasts before they sign.

What to do about it

Choose which corner you give up, on purpose. Score your model from one to five on simple, flexible and predictable, ideally with a customer in the room. Decide which one you sacrifice and say so in your sales story.
Combine a predictable base with a value unit. A platform fee or seat base, an included allowance of the unit customers care about, and a spend ceiling the customer controls. Cursor’s lesson is that surprise overage costs more goodwill than it earns.
Design the bridge before the destination. For existing contracts, set how long old and new models coexist, a conversion rate from seats to the new unit, and a date. Publish it.
Model partner economics before launch. Calculate what a reseller earns on a typical deal under the old and new model. If it drops, give partners a role they can be paid for, such as consumption management or implementation.
Never change a meter without a simulation. Before switching, run every customer’s last quarter through the new model and send them the result. Then switch.

The same three demands, a harder trade-off

In 2011 I argued that the cloud should make licensing simpler for customers, if vendors allowed it to. In AI the risk is the opposite: that vendors use the new technology to make pricing harder to understand. The companies that keep at least two of the three demands intact, and are honest about the third, will sell faster.