Your Software Bill Stopped Being Predictable. Here's What Changed

Disclosure, up front

We sell automation on fixed prices and monthly retainers, which is the model this article treats most kindly. Read section 05 with that in mind. The pricing-model figures are third-party industry surveys with the usual caveats, and where a number comes from a company selling billing software we say so.

Software used to be a predictable line item. You paid per person per month, you knew what next year cost, and the only surprise was the annual increase.

That arrangement is quietly ending, and not because vendors got greedy. It is ending because per-seat pricing does not work when the thing doing the work is not a person. An agent does not log in, does not hold a licence, and can complete a thousand tasks while an employee completes one. There is no seat to charge for.

What replaces it moves the risk from the vendor to you. Here is what changed, what the models actually are, and the three things worth negotiating before you sign.

The short version

Usage-based pricing went from about 30% of SaaS companies in 2019 to roughly 85% by 2024, and pure per-seat fell from 21% to 15% of offerings in a single recent year.
Hybrid is now the most common model — a base fee plus overage — at around 41% adoption.
Your bill is now a function of how much you use it, which is the same thing as saying it is no longer predictable.
The unit is the whole negotiation. Per token, per conversation, per ticket, per resolution: these are wildly different bets.

01The six models, and who each one favours

AI pricing models and where the risk sits
Model	You pay for	Who carries the risk
Per seat	Each named user	You, mildly. Predictable, but you pay for seats that go unused.
Per token / usage	Volume consumed	You, entirely. A busy month or a runaway automation is your bill.
Per conversation	Each interaction started	You, and note a conversation can end without resolving anything.
Per ticket	Each item handled	Shared, though "handled" is doing a lot of work in that sentence.
Per resolution (outcome)	Only successful outcomes	The vendor — the best alignment, if you agree what "resolved" means.
Hybrid (base + overage)	A floor, then usage above it	Shared below the cap, yours above it. Now the most common.

Model categories and adoption figures from 2026 industry pricing surveys, several published by billing and monetisation vendors with an interest in the topic — treat the percentages as directional. The risk column is our assessment, not from those sources.

Read the right-hand column rather than the price. Two vendors quoting similar monthly figures on different models are offering materially different deals, and the difference only becomes visible in your busiest month.

02Outcome pricing is the interesting one, and the definition is everything

Paying only for resolved tickets sounds obviously better, and often is. It is the only model where the vendor loses money when the product underperforms, which puts their incentives where you want them.

The catch is entirely in the definition. What counts as a resolution?

Does a conversation the customer abandoned count?
Does one where the customer said "no, that's not it" and gave up count?
Does an escalation to a human count as resolved by the AI, or not?
Who decides — their system, your records, or a shared log you can both read?

If the vendor's own system decides what counts as a success it also decides your invoice. That is not necessarily bad faith; it is simply a measurement you cannot audit. Get the definition in writing and make sure you can see the underlying events.

The question that reveals the model

Ask: "what does this cost me on our worst month?" Not the average, not the typical customer — your worst month. A vendor who has a good answer has thought about your risk. A vendor who redirects to averages has told you where the exposure sits, which is with you.

03Why this catches small businesses specifically

Large companies have procurement people whose job is exactly this. A small business signs up on a website, and three things follow:

The bill becomes seasonal in a way the budget is not. If usage tracks your busy season, your software cost now peaks precisely when cash is tightest.
An automation error becomes a financial error. Under per-seat, a loop that runs ten thousand times is embarrassing. Under usage pricing it is an invoice. This is the single most common unpleasant surprise we see.
Nobody owns the number. Usage-based costs drift upward without any decision being made, so there is no approval moment where someone notices — the same dynamic behind tier-upgrade sprawl.

04The three things worth negotiating

You have more room here than you think, because these models are new and vendors are still working out what customers will accept.

A hard cap, or an alert you cannot miss. The single most valuable term. Either the service stops at a number you set, or somebody is contacted before you cross it. Ask for a cap; settle for an alert; do not accept neither.
The overage rate, in writing, before you need it. Base rates are advertised and overage rates frequently are not. Overage is where the cost lives.
The definition of the billable unit, with a worked example. Ask them to price a specific month you actually had. Ambiguity in the unit always resolves in the vendor's favour, not through dishonesty but because they wrote it.

Two smaller ones worth asking for: what happens to unused allowance at month end, and whether you can move down a tier as easily as up. Both are commonly one-way.

05What this means for how you buy

Prefer predictable pricing when you can get it, even at a slightly higher expected cost. For a small business, budget certainty is usually worth more than optimising the average. Fixed-price and retainer work has this property, which is a fair thing to note as a bias of ours.
Match the model to the volatility of the work. Steady volume suits per-seat or a flat fee. Genuinely spiky work suits usage pricing, which is the case where it works in your favour.
Instrument before you commit. Run a month, measure your actual volume, then choose. Vendors will quote against their typical customer, which may be nothing like you.
Re-check every renewal. Models are changing fast enough that the deal you signed last year may not be the one on offer now — sometimes better.

06The honest summary

Per-seat pricing is being replaced because agents do not occupy seats, not because anyone is trying to overcharge you. Usage-based models now dominate, hybrid base-plus-overage is the most common arrangement, and the practical consequence is that your software bill moved from a fixed cost to a variable one.

That transfers risk to you, and the transfer is invisible until a busy month or a misconfigured automation. Get a cap or an alert, get the overage rate in writing, and make them define the billable unit against a real month of yours. Then ask what your worst month costs, and listen carefully to whether they answer.

07Common questions

Why is AI software moving away from per-seat pricing?

Because the thing doing the work is often not a person. An agent does not log in, does not hold a licence, and can complete a large number of tasks in the time an employee completes one, so there is no seat to charge for. Usage-based pricing grew from roughly 30% of SaaS companies in 2019 to about 85% by 2024, while pure per-seat pricing fell from 21% to 15% of offerings in a single recent year.

What are the main AI pricing models?

Six are in common use: per seat, per token or usage, per conversation, per ticket, per resolution which is outcome-based, and hybrid meaning a base fee plus overage. Hybrid is now the most common at around 41% adoption. They differ mainly in who carries the risk — per seat is predictable but you pay for unused seats, usage pricing puts a busy month entirely on you, and outcome pricing puts the risk on the vendor if you can agree what counts as a resolution.

Is outcome-based pricing better?

Often, because it is the only model where the vendor loses money when the product underperforms. The catch is entirely in the definition of the outcome. Does an abandoned conversation count? One where the customer said that is not it and gave up? Does escalation to a human count as resolved by the AI? And who decides — their system, your records, or a shared log? If the vendor's system decides what counts as success, it also decides your invoice.

What should I negotiate on a usage-based contract?

Three things. A hard cap where the service stops at a number you set, or failing that an alert somebody actually receives before you cross it — ask for the cap, settle for the alert, accept neither at your peril. The overage rate in writing, since base rates are advertised and overage rates often are not, and overage is where the cost lives. And the definition of the billable unit with a worked example priced against a real month you had.

What is the biggest risk of usage-based AI pricing for a small business?

An automation error becoming a financial error. Under per-seat pricing, a loop that runs ten thousand times is embarrassing; under usage pricing it is an invoice. Two related risks: the bill becomes seasonal in a way the budget is not, peaking when cash is tightest if usage tracks your busy season, and costs drift upward with no approval moment where anyone notices.

How do I compare two vendors on different pricing models?

Not on the monthly figure. Two vendors quoting similar prices on different models are offering materially different deals, and the difference only shows up in your busiest month. Ask each one what it costs on your worst month rather than an average or a typical customer, and price both against a specific real month of your own volume. A vendor with a good answer has thought about your risk; one who redirects to averages has told you the exposure is yours.

Should I avoid usage-based pricing entirely?

No — match the model to the volatility of your work. Steady, predictable volume suits per-seat or a flat fee. Genuinely spiky work is the case where usage pricing works in your favour, because you are not paying for capacity you do not use. For most small businesses, though, budget certainty is worth more than optimising the average expected cost, which is a preference worth stating explicitly rather than assuming.

How often should I revisit AI pricing?

Every renewal, at minimum. Pricing models are changing fast enough that the arrangement you signed last year may not be the one currently on offer, and sometimes the newer deal is better. Also worth asking at renewal what happens to unused allowance at month end, and whether you can move down a tier as easily as up — both are commonly one-way and both are negotiable.

Send us a quote and a real month

Take an AI tool you are considering, or already pay for, and send us the pricing page plus roughly what your volume looked like in your busiest recent month. We will work out what it actually costs you at your peak rather than at their average, and flag whether the billable unit is defined tightly enough to rely on.

Ask for a pricing check

Sources, read 9 September 2026: 2026 industry surveys of SaaS and AI pricing models for the shift from per-seat to usage-based pricing, the 30%-to-85% and 21%-to-15% figures, the six-model taxonomy and the roughly 41% hybrid adoption figure. Several of these are published by billing and monetisation vendors, who have a commercial interest in the direction they describe, so the percentages are directional rather than precise and none is a first-party measurement of the whole market. The risk allocation in the table, the negotiation points and the buying guidance are our own, and we sell on fixed prices and retainers, which is the model section 05 treats most favourably. Related: The Median AI-Using Small Business Now Runs Five AI Tools and Your Software Is Growing Its Own AI Agents.

Hero image from Unsplash, used under the Unsplash License.