As artificial intelligence agents spread across enterprise operations, spend management platforms are racing to fill a gap that traditional finance systems were never built to handle.
Ramp, a financial operations platform, raised $750 million at a $44 billion valuation this month, nearly tripling its worth in a year, TechCrunch reported June 4. The company is betting that AI consumption, billed by the token and fluctuating with every prompt and agent action, has become a cost category that most enterprise finance teams can’t track, allocate or control.
Corporate spend historically ran on two pillars, people and vendors, Ramp Co-Founder and CEO Eric Glyman said in a June 4 company blog post. In the past two years, a third pillar, intelligence, has been added. It’s billed per token and invisible to the systems built to manage cost. Ramp pulls usage data from providers including Anthropic, Cursor, Gemini and OpenAI into a single dashboard, letting finance teams tie spending to teams, projects and use cases.
The Budget Gap Agents Created
The problem Ramp is targeting didn’t exist at scale two years ago. AI providers initially priced access on flat subscription terms. As agentic models moved into coding, customer service, research and procurement, usage-based billing became standard. Every step an agent takes runs a meter.
The consequences hit enterprises before finance teams could adjust. Uber burned through its full 2026 AI budget in four months after encouraging engineers to adopt agentic coding tools without usage limits, TechCrunch reported June 2.
Now, the company has capped monthly token spending at $1,500 per employee per tool, a policy that required building an internal dashboard to make individual consumption visible, Bloomberg reported June 2.
The scale of the shift is reflected in OpenAI’s own usage data. Average reasoning token consumption per enterprise organization climbed roughly 320 times over the past 12 months, as more intelligent models moved into production workloads. Unlike seat-based software-as-a-service (SaaS) pricing, token consumption is granular and happens in real time, but it crosses department lines in ways that traditional ERP and expense systems have no mechanism to attribute.
AI invoices arrive as dense records of token counts, model tiers and throughput metrics that finance teams can’t map back to business activity. A cost spike may reflect a successful product launch, an inefficient prompt or an unmonitored automation loop. Without attribution, finance has no way to distinguish between the three.
Stripe Moves to Own the Billing Layer
Ramp isn’t the only platform reconfiguring around AI consumption. In January, Stripe completed its acquisition of Metronome, adding a metering engine already in use at OpenAI, Anthropic and Nvidia to its billing stack.
Open-source billing provider Lago said in a Feb. 14 company blog post that Stripe’s 2018 billing architecture was built for subscriptions with pre-aggregated usage data and couldn’t handle real-time event streaming at AI inference scale, making acquisition faster than a rebuild.
Stripe CEO Patrick Collison said in December that metered pricing is the native business model for the AI era and the revenue shift it drives could equal or exceed the transition to SaaS.
The platform now handles credit burndown, outcome-based billing and multidimensional metering for AI infrastructure companies with complex, usage-heavy product catalogs, according to a Jan. 23 company blog post.
Stripe’s move addresses the sell side. It helps AI companies bill customers accurately for consumption. Ramp targets the buy side, where enterprises are trying to understand what they spend, with whom, and whether it returns value.
A New Control Layer Takes Shape
Middleware platforms are building around the same gap.
OpenRouter, which routes enterprise requests across more than 400 AI models through a single API, reported traffic of 25 trillion tokens per week, a five-fold increase in six months, Tech Startups reported May 26. The company raised $113 million in May, led by CapitalG, Alphabet’s independent growth fund. Enterprises use the platform to push routine tasks to lower-cost models and reserve frontier models for higher-stakes work, converting model selection into a cost-control decision.
Per-token rates have fallen. In his blog post, Ramp’s Glyman said there has been a drop from $60 per million output tokens at GPT-4’s 2023 launch to roughly 40 cents today for comparable performance.
But agentic workloads burn more tokens per task than single-turn queries, and enterprises running multiple agents across multiple providers still have no consolidated view of where that spending goes.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
