The pricing page is the most human artefact in a software company. It was designed to be read by a person weighing three columns, nudged by a highlighted middle tier, reassured by a list of ticks, and eventually persuaded to click. Every assumption in it, from the seat as the unit of value to the annual contract as the unit of commitment, rests on a buyer who reads, compares, hesitates and decides. That buyer is increasingly not the one arriving. More purchasing is being delegated to software acting on a person’s behalf: an agent tasked with finding a tool for a job, provisioning it, using it and dropping it when the job is done. That buyer does not read the pricing page. It reads the API, and what it finds there was never designed for it.
The shift is not that humans stop buying software. It is that a growing share of the deals start, expand and end without a person ever looking at a page, and the companies whose pricing can accommodate that will win business their competitors cannot even quote. Stripe’s own reporting on how AI companies monetise points at the same direction: usage-based and hybrid models growing, pricing changed far more often than the old annual cycle, and the commercial machinery becoming something engineering owns rather than something marketing publishes.
When the buyer is a program
A machine buying software wants different things from a person. It wants metering, so that it can pay for what it used and no more. It wants limits it can read, so that it knows before it starts whether the job fits inside the plan. It wants terms it can parse, which means machine-readable rather than a PDF. And it wants to sign up, use, expand and cancel programmatically, without a sales call, a demo or a support ticket. A product that offers only a seat-based tier with an annual contract and a talk-to-sales button for anything larger is, from the agent’s point of view, not for sale.
Seats are the assumption that breaks first. A seat prices a human’s access, and it made sense when the work was done by humans. When an agent does the work of what would have been five people’s tools, a five-seat plan either drastically undercharges for the value or absurdly overcharges for a single account, and neither is stable. Companies are quietly discovering that their seat counts no longer describe anything real, and the ones that keep pricing on them are pricing on a fiction.
The annual contract is the second assumption to go. A person signing for a year is buying predictability and a discount, and both made sense when switching tools was a project. An agent that provisions a tool for a task has no use for a year, and a company that only sells years has priced itself out of every short, intense job, which is increasingly where the interesting demand lives. Usage-based and hybrid models are growing for exactly this reason: they describe what the buyer actually consumed, and a machine buyer cares about nothing else.
Build the plumbing before you need it
The response is not to guess the right agentic pricing model and publish it. Nobody knows the right model yet, and the companies that have looked closest say so. The response is to build the infrastructure that lets pricing change quickly, so that the company can try a model, measure it, and adjust in weeks rather than in a two-quarter engineering project. That means metering built in from the start, even for features that are not metered today. It means plans and limits held as data rather than hard-coded. It means billing that can handle usage, credits, caps and tiers in combination, because the winning model is likely to be a combination. And it means the ability to run two pricing models at once, for different cohorts, without the codebase knowing the difference.
A product that offers only a seat-based tier with an annual contract and a talk-to-sales button for anything larger is, from the agent’s point of view, not for sale.
This is unglamorous work and it is easy to defer, because a pricing page can be changed in an afternoon while pricing infrastructure takes a quarter. But the page is not the product. The plumbing is. A company that can change price weekly will learn what agent buyers value faster than one that can change it yearly, and the learning is the asset, not any particular price.
Testing without breaking promises
The reason pricing changes are slow in most companies is not technical. It is fear of the existing customers, who signed up under one set of terms and will not enjoy discovering another. That fear is legitimate and it is also manageable, provided the plumbing supports it. New models can be offered to new customers first, with existing accounts grandfathered on their current terms until they choose to move. Changes can be trialled on a cohort small enough that a mistake is affordable. And every change can be reversible, which is the property that makes experimentation safe: a price that can be rolled back is a price that can be tried.
There is also a part of the relationship that should stay human, and it is worth naming so that the push towards programmatic buying does not flatten it. The security review, the procurement conversation for a large account, the moment a customer is deciding whether to trust the company with something important: these are not steps an agent should complete alone, and a product that tries to remove the person from them will lose the deals where the person mattered. Machine-readable pricing is for the deals that never needed a conversation. The others still do.
What must not happen is a silent change to terms a customer, human or machine, has built on. An agent that provisioned a tool under one limit and finds a different one mid-job will fail its task and its owner will draw conclusions. The same discipline that governs rate limits governs prices: announce, explain, give a path, and never surprise.
Signals that the model is drifting
A pricing model does not fail all at once. It drifts, and the drift shows up in a few places worth watching. Support tickets that ask how to do something programmatically that the product assumes a person will do by hand. Accounts with one seat and enormous usage, or many seats and none. Customers who churn not because they were unhappy but because the job finished and there was no way to pause. Sales conversations that keep stalling on the same question about limits or metering. And, most tellingly, deals lost to a competitor whose pricing an agent could simply consume.
Each of these is a message from a buyer the pricing page was not written for, and the messages arrive quietly, one ticket and one lost deal at a time. The companies that hear it early will rebuild the plumbing while it is still cheap. The ones that hear it late will discover that their pricing was a set of assumptions about a person clicking a button, and that the person has stopped showing up.