I got on a call last quarter with an AI company that had just deployed 200 agents across their support org. Their infrastructure bill tripled in the first month. Revenue from those same customers? Flat. Not a single dollar of increase.
I wish I could say that was unusual. It’s not. I talk to AI companies every week, and the same story keeps showing up.Â
An IDC study found that 92% of businesses running agentic AI experience cost overruns. Even worse, 71% can’t pinpoint where the money is going. That’s not a budgeting mistake. That’s a billing model that no longer fits the product it’s supposed to charge for.
Agents consume resources like nothing we’ve billed for before
Here’s the thing. A SaaS user logs in, clicks around, runs a report, logs out. You know roughly what that session costs to serve. You charge per seat. The math works.
An AI agent operates nothing like that. You give it one instruction, and it fires off 30 API calls, chains five tools together, retries three times, and burns through tokens that nobody budgeted for. According to Axios, agentic workflows cost 5 to 25 times more than a standard chatbot interaction. A classification task that costs $0.01 in a chat interface can run $0.10 to $0.50 when an agent handles it.
And the costs don’t scale in a straight line. In multi-agent systems, each tool call adds context. Each sub-agent response feeds back into the orchestrator. Token costs compound rather than add up. One developer tracked their agent costs for three months and found that two thirds of all spend went to API tokens. A busy day cost $8. A quiet Sunday cost $1.50. Same agent. Same job.
Per-seat pricing ties your revenue to headcount. Agent costs tie to autonomous workloads that scale independently of how many people you employ. Those two things cannot coexist in the same billing model. Usage-based pricing is the only structure that aligns what you charge with what your infrastructure actually spends. I wrote a deeper breakdown of how to pick the right pricing model for AI agents if you’re already dealing with this.
The biggest names in tech are already scrambling
I’m not the only one saying this. Look at what the largest companies in the industry did over the past year.
OpenAI moved Codex, their agentic coding tool with over 2 million weekly users, from a mandatory $25 per user per month plan to pay-as-you-go pricing. The reason was blunt: agentic AI costs far more to run than chatbots, and flat pricing meant OpenAI absorbed the compute costs of their heaviest users. If OpenAI can’t make per-seat pricing work for agents, I don’t know who can.
Salesforce tried three different pricing models for Agentforce in under a year. They started at $2 per conversation, then added Flex Credits at $0.10 per action, and they still offer per-user licensing at $125 a month. Agentforce hit $540 million in annual recurring revenue with 330% year-over-year growth. The demand is clearly there. But three pricing models running simultaneously tells you that nobody has cracked the right one yet.
Anthropic shifted to usage-based pay-as-you-go pricing for third-party tools that connect to Claude. Flat subscriptions couldn’t absorb the cost swings.
Intercom took a completely different path with their Fin AI agent. They priced it at $0.99 per resolved outcome. Not per conversation. Not per seat. Per actual result. That model drove a 40% jump in adoption with stable margins inside six months.
Forrester’s numbers confirm the broader shift. Seat-based pricing dropped from 21% to 15% of SaaS market share in a single year. That decline is picking up speed.
What usage-based billing for agents actually requires
I build billing infrastructure for AI companies, so I see the gaps up close. Traditional billing systems were designed for predictable, subscription-shaped revenue. Agents break that assumption in four specific ways.
First, you need real-time event metering. Agents fire API calls, consume tokens, and invoke tools at unpredictable rates. Your billing system needs to ingest millions of these events without losing data or adding latency to the product. Most billing platforms batch usage overnight. That falls apart when an agent burns through a customer’s entire credit balance in 20 minutes.
Second, you need credits and spending controls that actually work in real time. Prepaid credit models are becoming the standard because they give customers a clear budget ceiling. But those credits require real-time balance tracking, auto top-ups that trigger instantly when a threshold is hit, and alerts that fire before the balance runs dry. Monthly reconciliation is useless for agents that run 24 hours a day.
Third, you need usage limits you can change without shipping code. If enforcing a spending cap requires an engineering deployment, you’re already behind. Agents don’t stop working at 5 PM. You need daily, monthly, and custom caps per customer that your team can adjust on the fly.
Fourth, you need hybrid pricing that handles real complexity. The market is landing on a base fee plus consumption. Salesforce already runs this way. Most AI companies will follow. That means your billing system has to handle subscriptions, usage-based charges, and credit drawdowns at the same time, for the same customer, on the same invoice.
This is the problem I’m building Flexprice to solve: an open-source usage-based billing platform for AI and agentic companies.
The 40% that won’t make it
Gartner predicts that over 40% of agentic AI projects will get cancelled by the end of 2027. The top reason? Escalating costs and inadequate controls.
Most of those projects won’t die because the AI didn’t perform. They’ll die because nobody figured out how to meter, price, and bill for what the agents actually consumed. The company I mentioned at the top, the one that tripled their infra costs? Their agents worked great. The billing just couldn’t keep up.
Gartner estimates that AI agents will handle $15 trillion in B2B purchases within three years. That money flows through billing infrastructure. If yours can’t handle usage-based billing, someone else will.
Manish Choudhary is the founder of Flexprice, an usage-based billing platform built for AI and agentic companies.
