AI agents are changing the math on software spending
As vendors mix seats with credits, actions and consumption-based charges, CIOs face new questions about how to compare, forecast and fund software costs.
For many enterprise software purchases, the calculation was familiar: Count the employees who needed access and buy the corresponding licenses.
Agentic AI is beginning to complicate that equation.
Microsoft, for example, pairs fixed Copilot user subscriptions for everyday AI tasks with usage-based billing for more advanced agentic work. Separately, its agent management and governance platform, Agent 365, is licensed per user, not per agent. Microsoft says these different billing approaches can help organizations plan and manage costs as agents take on more complex work.
Those approaches point to a model in which seats aren't disappearing so much as being supplemented. CIOs may pay separately for human access and oversight and for the work agents perform.
Fixed Copilot user subscriptions are generally purchased by IT and provided across the organization, according to Microsoft. But the company suggests usage-based costs for more advanced agentic work may be better funded by individual business units, which can weigh consumption against the value of the work.
"The seat is not completely disappearing; it is being unbundled from execution," said George Brocklehurst, managing vice president at Gartner.
As agents perform more of the work, Brocklehurst said, the human seat increasingly represents the people responsible for orchestrating, governing and signing off on that work rather than the total volume of software activity. That leaves CIOs managing a growing mix of seats, credits, actions and consumption within the same software portfolio.
Other enterprise software vendors are mixing per-user and usage-based pricing. Salesforce uses Flex Credits to price Agentforce actions and also offers per-user options. SAP, meanwhile, uses AI Units to meter certain AI capabilities.
For CIOs, the result isn't simply a transition from seats to consumption. It could mean a portfolio of multiple units that don't map neatly onto one another.
The same agent may be cheap when the workflow follows the happy path and expensive when it encounters the enterprise as it actually is.
George Brocklehurst, managing vice president at Gartner
What a workflow really costs
Brocklehurst said CIOs should resist comparing products by unit price and instead compare them by the cost of completing the work. Two agents, for example, could advertise the same price per action but require vastly different numbers of steps, retries, tool calls and human interventions to complete the same workflow in an enterprise environment.
"The same agent may be cheap when the workflow follows the happy path and expensive when it encounters the enterprise as it actually is," he said.
Rather than focus on the advertised price of a credit or action, CIOs should consider the expected cost of successfully completing a workflow, how much that cost might vary and who bears the cost when execution doesn't go as planned, Brocklehurst said.
Tracy Woo, principal analyst at Forrester, said the move toward hybrid pricing models is already underway, with vendors increasingly combining fixed fees with consumption-based pricing.
Woo pointed to Anthropic's Claude Code pricing as an example. Under newer Claude Enterprise plans, customers pay a per-user fee for access plus separate charges for AI usage, including Claude Code.
"The shift to consumption-based pricing is in the future for every model and AI SaaS provider, at least in part," she said.
AI investments are a portfolio decision.
Tracy Woo, principal analyst at Forrester
That shift also complicates product comparisons. CIOs aren't necessarily choosing between two products measured using the same economic unit, Woo said. Vendors also differ in how much pricing detail they disclose, particularly when AI capabilities are embedded in existing software products.
"AI investments are a portfolio decision," Woo said.
Who pays as agent activity grows?
As more software execution shifts to agents, headcount could also become a less reliable predictor of software spending.
"Seat count will increasingly measure the human control layer, not the total volume of software execution," Brocklehurst said.
Rather than forecasting demand primarily around the number of users, CIOs may also need to account for factors such as workflow volume, execution frequency, complexity, autonomy and exception rates, he said. Simply replacing the number of users with the number of agents wouldn't solve the problem because agents themselves aren't consistent units of work.
That shift also raises questions about who pays for variable agent spending. Brocklehurst said central IT should remain responsible for the platform, enterprise limits and risk controls. At the same time, business units should own the consumption their workflows generate and the value they expect from that spending.
"Traditional software budgets fund access. Agent budgets delegate spending authority to software," he said.
The mix of seats, consumption and other pricing units may be more than a temporary step toward replacing per-user licensing. Those units can reflect different elements vendors are trying to price, from human access and oversight to the amount of work an agent performs.
"The market may never find a single replacement for the seat because it is trying to price several different things at once: human authority, machine work and vendor confidence in the outcome," Brocklehurst said.
Liz Hughes is an award-winning editor and writer covering AI and emerging technology and the former editor of AI Business and IoT World Today.
