Introduction
Six months of agent billing, what customers paid for, and why per-seat AI pricing rarely survives a second renewal.
The problem with per-seat AI
Per-seat pricing was designed for software where a seat corresponds to usage. AI does not work that way. Ten people might have licences and two might use it, or one person might drive enormous value through a single seat.
More importantly, per-seat pricing charges you identically whether the AI works or not. That is a comfortable position for a vendor and an uncomfortable one for a buyer who has watched a pilot fail before.
What makes an outcome measurable
Not every AI task has a clean outcome. Summarising a document is useful but hard to price on. Recovering an overdue invoice is not: either the money arrived or it did not.
The processes that price well on outcomes share three properties — a binary or clearly quantified result, an auditable record of what happened, and a value that both sides agree on in advance.
Where the money actually goes
Across deployments, the pattern is consistent. Collections agents generate the largest absolute value because receivables are large and the alternative is doing nothing. Document processing generates the highest volume of small charges. Screening and scheduling sit in between.
Crucially, customers pay nothing in months where an agent produces nothing — which happens, and which is the point.
Questions worth asking a vendor
- What specific outcome does this agent produce, in one sentence?
- How is that outcome measured, and can I audit the measurement?
- What do I pay in a month where it produces nothing?
- Can I switch it off without affecting the rest of my subscription?
A vendor who cannot answer the first two should not be charging a fixed fee for the third.
The honest limitation
Outcome pricing does not suit every task, and pretending otherwise is a sales tactic. Assistive AI — drafting, summarising, answering questions — genuinely is better bundled as an allowance. The distinction is worth insisting on rather than blurring.
In summary
None of this requires a transformation programme. It requires deciding where each piece of data is created, making sure it is only created once, and letting everything downstream read from that record instead of keeping its own copy.
Meera Krishnan, Marketing Lead at easyto.work.
