How AI is rewriting the business model of software

Imagine you no longer buy software but work. You no longer pay for access to a tool, but for the job once it is done. It is one of the biggest shifts in the software industry since Software-as-a-Service (SaaS) arrived.

Until recently, software was billed per user, per 'seat'. As long as a program was a tool that helped a person work faster or better, that made sense: more users, more licences, more revenue. But AI models are getting so good that they do the work themselves. And once the AI replaces the employee, you end up with fewer users while consuming more AI compute. At that point the old model no longer holds. So the bill shifts from access to outcome: you no longer pay for a seat, but for the work once it is done. Venture firm Bessemer calls it the "AI pricing pivot".

Why the per-user model no longer works

Under the classic SaaS model, you pay a fixed amount per user per month. A company with fifty people in customer service buys fifty licences. Put an AI agent in place that takes over the work of thirty of them, and you are left with only twenty people behind a screen. Under the old model that means thirty fewer licences, so considerably less revenue for the vendor while the costs go up (AI tokens), while the customer gets the same work done.

No business model survives that arithmetic. The vendor is in effect rewarded when its own product underperforms, because only then do all those people stay necessary. That obviously does not work. The price has to match the value, not be based on access to a tool. And with an AI agent that value is, for example, a resolved customer query, a fixed bug or an automated workflow.

GitHub Copilot starts the meter

On June 1, 2026, GitHub Copilot, the AI assistant for programmers, overhauled its pricing model. You still pay a fixed amount, but the heavy AI work now draws from a pool of credits, and once that runs out a meter starts running. Only the simple autocompletions as you type stay unlimited. GitHub's own reasoning: a quick question in a chat window and a session in which the AI spends hours autonomously reworking a codebase used to cost the same, and that is not sustainable.

Other development environments where developers write and test their code, such as Cursor, are moving the same way with hybrid models: a fixed amount per developer plus a usage component for heavy tasks. The fact that several major tools are shifting in that direction around the same time is a clear signal. Unlimited use for a fixed price no longer fits the cost of these models.

Devin: an AI engineer you hire by working time

You see the shift most sharply with Devin, from the American company Cognition. Devin is not an assistant that types along with you, but an autonomous software engineer. You give it a task, and Devin plans, writes, tests and delivers on its own.

Devin bills in Agent Compute Units, its own unit of measurement that bundles all the resources the system uses: not just the AI itself, but also compute and bandwidth. According to user estimates, one such unit works out to roughly fifteen minutes of Devin working time, which comes to about $8 to $9 per hour. It depends heavily on what you have it do, so treat this as a ballpark figure and not a guarantee.

This shift, paying for working time instead of access per user, changes how customers look at an AI tool. You no longer compare Devin with a software licence, but with an hourly wage. The question then becomes whether a task is cheaper with AI than with a junior developer, or perhaps even a more experienced offshore developer.

Customer service: paying per resolved conversation

A completely different corner, but the same movement. With its AI agent Fin, Intercom offers a simple billing model: $0.99 per outcome. Such an outcome is a fully resolved conversation, or a clean handover to a human further along in the workflow, or a successful handover to another AI agent. If it does not work out, you pay nothing. So you only pay when something has genuinely been resolved.

Sierra, the company founded by Bret Taylor (former co-CEO of Salesforce and now chair of OpenAI's board), has built its entire business on this idea. No seats, you only pay when a case is resolved. The promise is as simple as it is bold: we only earn once your customer's issue has actually been solved. Because Sierra does not lean on licences, it has no conflicting interest either. If their agent gets better and you need fewer people, that is good news for them.

Even Salesforce and SAP are moving along

As always, it was the innovative AI startups that first experimented with new revenue models. Now the giants that ran on per-user licences for decades are following too.

Salesforce now bills its AI, Agentforce, as "digital labor": you buy credits or conversations, and every task the agent handles consumes a slice of that budget. And Salesforce took a bigger step. On June 15, 2026, the company announced it was acquiring Fin for roughly $3.6 billion, exactly the customer service agent from above. The acquisition has been announced but not yet closed, and the price of $0.99 per outcome is unchanged for now.

SAP is shifting its AI agents to usage-based pricing too. Instead of counting how many employees log in, the company is moving to billing based on the volume of automatically processed tasks and transactions, through a credit it calls AI Units. In March 2026, SAP chief executive Christian Klein told Bloomberg that it would be foolish to keep billing per subscription, because AI automates so much work. But that transition has to be handled carefully, because the customer's bill goes up and that leads to difficult conversations with those customers. The large software companies are therefore adding usage-based models slowly, alongside their existing licences and seats, rather than switching over all at once.

The flip side, and an unexpected benefit

The three examples illustrate the broader trend: from seat to salary. Still, I want to add a critical note. The moment you pay per unit of work, software no longer competes with other software, but with the labour market: with temp workers, offshore teams and juniors. That makes the comparison fairer, but also harsher, and with a bigger impact on the job market.

There is also a modest upside. When the bill shifts from a fixed monthly amount to paying per use, organisations are forced to choose more deliberately what they do and do not let an AI do. Unlimited experimentation feels less noncommittal when a meter is running that can send the bill up sharply. Fewer unnecessary runs means less compute demanded. That certainly does not solve AI's environmental impact, but it does reduce the amount of energy and infrastructure deployed without clear value. In this way the new business model achieves something good intentions rarely manage: thrift.

Frequently asked questions

What exactly is outcome-based pricing?

With outcome-based pricing you do not pay for access to software or for the number of users, but per result achieved: a resolved customer query, a completed task, an automated workflow. If it does not work, you usually pay nothing. The price then follows the value the software delivers.

Why does the per-user SaaS model no longer work for AI?

Because an AI agent does not log in the way a person does. The better the agent works, the fewer human users a company needs, while the value delivered stays the same or even rises. Under the old model the vendor's revenue then falls while its costs rise (AI tokens) and the customer gets more work done. That is not sustainable.

Is paying per use cheaper?

Not necessarily. It does tie costs more fairly to actual use. It also makes them less predictable. A single heavy job can drive the bill up considerably. For light use, usage-based pricing is often favourable; for intensive use, a fixed amount can work out cheaper.

This article is an adaptation of my conversation with John van Schagen on the BNR podcast Baanbrekende Businessmodellen (July 3, 2026). Listen to the eleven-minute episode via BNR (in Dutch).