Did you want to save with AI? Here comes payment per job done

When OpenAI released its now indispensable ChatGPT to the public at the end of November 2022, the whole world went crazy for this tool. In February 2023, it also introduced its subscription model, and today users can choose between the simplest free model, then ChatGPT Go, ChatGPT Plus, and the Pro version. There are, of course, options for business users: Business and Enterprise. But now the subscription model is changing. More and more AI tools are introducing payment based on actual usage alongside fixed fees, meaning that the bill no longer depends solely on how many employees use the software, but also on how much work artificial intelligence does for them.

Thus, artificial intelligence will now be paid like employees, per job done, which is interesting because large global companies have invested millions in it in recent years precisely to automate work and reduce the number of employees, and thus costs.

Change of model

One example of those changing the payment model is Atlassian, the company behind the Jira and Confluence software. This company announced a new payment model at the beginning of September that will come into effect on December 3, which relates, among other things, to the AI functions of its Rovo system, automation, and AI agents. Users will have certain limits with their existing subscription, and additional usage will be charged based on consumption.

This is an important change for the business software segment as the traditional SaaS model has mainly been based on ‘per user’ payment for decades. A company buys a certain number of licenses, knows how much it pays each month, and can relatively easily predict costs. So, a classic subscription like for Netflix. But artificial intelligence has complicated the calculation. Namely, one employee may use AI two or three times a day, while another may use it practically continuously for coding or data analysis. And if one company pays for the software per user for those two employees, both cost the same. But if the tool is paid based on usage, the difference is indeed significant.

This is precisely the logic behind the change in the billing method implemented by Atlassian. For some of its AI functions, it will use the Rovo credit system. They measure the usage of AI functions and access to additional context from the so-called Teamwork Graph. Simpler AI functions have a certain credit consumption, while for more advanced functions, the number of credits may depend on the complexity of the request and the computing capacity required for its solution. But now the real question is how much all this can cost. Atlassian has set the price for additional Rovo credits at one cent per credit, so if we look at just this example, a monthly package of two thousand credits will cost twenty dollars, and the price of the largest available package of ten million credits will go up to one hundred thousand dollars per month.

Growing trend

An even more interesting example is AI agents in customer support. Atlassian will charge for their usage based on the number of cases the agent resolves independently without the need for an employee to take over the conversation. The price will be one dollar per resolved case, starting from the first resolved case. If a company has ten thousand customer inquiries that the AI agent resolves independently, the bill amounts to ten thousand dollars. For larger quantities, the price may drop, so for more than fifty thousand resolutions, it will cost 95 cents, and for more than two hundred thousand resolutions, 85 cents per case.

And Atlassian is not the only one thinking this way. The consulting firm McKinsey recently published an analysis stating that business software is increasingly shifting from user-based billing to consumption-based billing, or payment based on the number of inquiries, computing resources used, and results achieved. McKinsey also conducted a review of the growth in the number of companies applying some form of consumption-based billing, and as they state, their number has more than doubled from 2015 to 2024. According to this consulting firm, software companies that successfully transition to usage-based billing and the results they achieve can better leverage the value that artificial intelligence brings and stand out in the rapidly growing AI market.

In any case, other companies offering artificial intelligence tools are heading in the same direction. OpenAI, for example, still charges business users for ChatGPT based on the number of users, but it has also upgraded this model to include consumption-based billing. ChatGPT Business includes a certain amount of usage in the license price, and when the user exceeds that limit, the company can continue to use advanced features by spending credits from a shared fund. Enterprise users can also purchase a shared credit fund, and consumption can be billed based on the number of messages, tasks, generated content, or used tokens, depending on the function.

Credits and tokens

Additionally, OpenAI charges a fixed number of credits per usage for some functions. One interaction with the agent consumes thirty credits, in-depth research fifty credits, and image generation five credits. For tools that perform more complex tasks, such as Codex and Workspace agents, consumption depends on the number of input and output tokens. In other words, not every use of artificial intelligence is equally expensive; rather, the bill increases with the amount and complexity of the work. A similar model is being developed by Anthropic with its Claude. Its business users can still purchase licenses per user, but Claude Code for Team and Enterprise users is charged based on consumption. Anthropic applies this model even for the most demanding models: for example, companies using certain models through the Enterprise consumption model pay based on the number of processed tokens. For the latest models, the price is calculated separately for input and output tokens, that is, for the amount of data the model processes and the amount of content it produces.

Google, on the other hand, applies a similar logic in Gemini. Google Workspace users still pay for licenses per user, but the company is simultaneously introducing usage limits for advanced features of Gemini. How much users can use depends on the function and package they have, and for certain functions, Google directly measures the number of uses. For instance, business users in certain packages have monthly quotas for some functions in Google Sheets, while more advanced options have separate limits.

Microsoft goes a step further with Copilot. With Microsoft 365 Copilot, which is still sold via a per-user subscription, the company has introduced Copilot Credits, a shared currency for using AI agents. Each activity the agent performs consumes a certain number of credits, and the amount depends on the complexity of the task. Organizations can purchase credits in advance or pay based on actual consumption. Microsoft currently lists the price for the consumption-based payment model at one cent per Copilot credit.

New hybrid model

Thus, a new hybrid model is emerging. Companies no longer necessarily choose between per-employee subscriptions and consumption-based payments, but increasingly pay for both: a base price for access to the tool, and then additionally for the amount of work that artificial intelligence performs.

This is important because the way artificial intelligence is used significantly differs from how traditional business programs were used. An employee may have one license for business software and use it as often as they want. An AI agent, however, can open documents, analyze data, write reports, send messages, check databases, and repeat the entire process hundreds of times on their behalf. Each of these tasks means additional consumption of computing resources for the AI producer. Therefore, the price is increasingly tied to what artificial intelligence actually does, whether it is analyzing a large number of tables, solving problems, or some other form of usage.