The short answer
A seat is a proxy. It was a good one for twenty years, because software only produced value when a human sat in front of it, so counting humans was a fair way to count value. AI agents remove the human from the middle of the task. The work still happens, the output goes up, and the number of seats stays where it is or falls. Revenue and value part company, and both sides start to feel it.
Vendors feel it first as a growth problem: their best customers automate a department and then ask to cut licences. Buyers feel it second as a fairness problem: they are billed for chairs nobody sits in, while the useful work is done by something that never logs in.
Where the seat model actually breaks
Three things go wrong at once. First, the unit stops correlating: one seat might now trigger ten thousand agent runs, or none. Second, the expansion path dies: land-and-expand assumed more users next year, and agent-led teams hire fewer of them. Third, cost of goods stops being flat. Traditional software cost roughly the same to serve whether a user did five things or five hundred. Inference does not. Every agent run has a real, variable cost behind it, so a fixed per-seat fee can turn your heaviest users into your least profitable ones.
That last point is why this is not a marketing fashion. It is a margin problem. Pricing that ignores variable cost eventually gets corrected by the P&L, usually at renewal, usually badly.
Building an agent and unsure how to charge for it?
We build agents, portals and automations in house, with metering and reporting designed alongside the feature. Fixed monthly pricing, published deliverables.
Monthly + VAT. Six month proving period, then rolling monthly. No 12-month lock-in.
The four models replacing per-seat
Usage
- What you pay for
- Per token, minute, run or document processed
- Where it fits
- Early tools with wildly different customer volumes
- What to watch
- Budgets become unpredictable and teams under-use the tool to stay cheap
Per action
- What you pay for
- Per completed unit of work: ticket resolved, invoice matched, lead enriched
- Where it fits
- Agents that finish a clean, countable job
- What to watch
- Define what counts as complete, and who pays when the agent is wrong
Outcome
- What you pay for
- A share of a measurable result: booked meeting, recovered payment, closed claim
- Where it fits
- Agents sitting directly on revenue
- What to watch
- Attribution arguments, and a vendor incentive to over-trigger the event
Hybrid
- What you pay for
- Fixed platform fee, plus a bundled allowance, then transparent overage
- Where it fits
- Most serious buyers and most serious vendors
- What to watch
- Check the allowance matches real volume, not the demo volume
Most credible pricing in 2026 is hybrid: a fixed platform fee that keeps the lights on and the roadmap funded, a bundled allowance generous enough that daily use is never a decision, and transparent overage above it. That structure gives the vendor predictable revenue and gives the buyer a number they can put in a budget.
The adoption tax nobody prices in
Pure metered pricing has a quiet failure mode. If every run shows up on an invoice, people stop running it. Teams batch work, skip the agent for small jobs, or route around it entirely to protect their budget line. Six months later the tool has low usage, weak internal advocacy and no renewal case, and everyone blames the technology rather than the meter.
The fix is not complicated. Bundle the expected volume into a fixed monthly price so ordinary use costs nothing extra to decide on, then charge for genuine excess. Predictability is a feature, and it is the one buyers pay a premium for.
What to ask before you renew
Whatever the vendor calls the model, get six things in writing. The exact unit that triggers a charge and where it is logged. The definition of a completed action, including part-completions. What happens commercially when the agent is wrong, and who reworks it. A hard spend cap or alert threshold you control. The overage rate, in the contract rather than a support article. And the notice period, because usage pricing with a twelve month lock-in is the worst of both worlds.
If a vendor cannot answer those, the pricing is not a model, it is a hope. That is the same test we apply to marketing and software fees, which is why every price we publish is monthly, plus VAT, and attached to named deliverables.
If you are the one doing the pricing
Building an agent into your own product or portal changes the commercial design, not just the code. Work out the variable cost of a single run before you name a price. Decide which unit your customer already counts, because a unit they recognise sells far faster than one you invent. Keep a floor so your revenue survives a quiet month. And instrument the thing properly from day one, since you cannot bill per action, or defend a bill, without a clean event log.
We build this kind of thing in house: agents, internal portals and automations, with the metering and reporting designed alongside the feature rather than bolted on after the first billing dispute. That work sits under Custom Software and, where the question is what to automate first, Digital & AI Consulting.
Questions we get asked
Seat pricing assumes value scales with the number of people logging in. An AI agent does the work without a login, so the software can triple its output while the seat count stays flat or falls. The vendor's revenue stops tracking the value delivered, and the buyer starts paying for chairs instead of work.
Four models, usually combined. Usage pricing charges per token, minute or run. Per-action pricing charges per completed unit of work, such as a ticket resolved or an invoice reconciled. Outcome pricing charges a share of a measurable result, such as a booked meeting or recovered payment. Hybrid pricing puts a platform fee underneath any of those so the vendor still has predictable revenue.
It is better when the outcome is measurable, attributable and audited by both sides. It is worse when the definition is loose, because you end up arguing over attribution every month. Ask for the exact event that triggers a charge, where it is logged and what happens when the agent gets it wrong.
Model a floor and a ceiling. Take last quarter's real volume of the task the agent replaces, price it at the vendor's per-action rate, then add a cap or an alert at the level you are willing to spend. Any vendor unwilling to give you a spend cap is asking you to sign a blank cheque.
It can. If every extra run costs money, teams quietly stop using the tool, and you lose the compounding value you bought it for. The fix is a bundled allowance at a fixed monthly price, with overage priced transparently, so day-to-day use is never a decision.
We price marketing and software subscriptions as a fixed monthly fee for a fixed, published list of deliverables, plus VAT. Every subscription runs an initial 6-month proving period, then rolls monthly. No mystery hours, no per-seat maths, and no surprise usage bill at the end of the month.
Where to go next
If you want the fixed-price version of software and marketing, the pricing page shows what lands each month, and the subscription page explains the terms in full. Building an agent and unsure how to charge for it, ask us and we will tell you straight.
Written by
Sam, co-founder, growth and paid media
Runs growth and paid media at Mesmerise Marketing. Ten years of paid and organic campaigns for challenger brands across DTC, B2B services and hospitality.
Meet the team
