With a $35 million series A, SAPIOM wants to become the financial controller for AI Agents

AI agents can chain together dozens of model calls, use external tools, purchase services and retry the same operation several times without the company knowing exactly how much each assignment has cost. As artificial intelligence budgets face their first audits, a new infrastructure layer, embodied by companies such as SAPIOM, is emerging to attribute, govern and optimize this spending. 

When companies ask, “How many agents ran yesterday, and how much did they cost?”, the answer is often reduced to a single figure representing a number of tokens, without anyone knowing exactly which agents generated the expense, which tasks succeeded, where executions stopped or how many times individual steps were retried.

This is what the economic dimension of many agentic AI projects has been reduced to: the company knows how much was charged, sometimes with admirable precision down to the last token, but still struggles to establish what it actually purchased.

This difficulty may appear secondary while agents remain confined to demonstrations or experiments. It becomes much more serious when they operate continuously, use several models, call search engines, execute code, access databases or purchase services without the slightest human intervention. Software then becomes an autonomous consumer of usage-based resources, where every decision carries a cost, and every error does too.

Tokens no longer tell the full story

Although the price per million tokens remains the benchmark metric of generative AI economics, it provides a poor account of an agent’s real cost. A seemingly simple assignment may begin with a call to interpret the request, continue with documentary research, bring in a second model to analyze the results, invoke a specialist tool, and then return to the initial model to produce the final response. If one of these stages fails, the agent can revise its plan and start again. It is easy to see where this leads.

On top of this consumption come reasoning tokens, context retention, vector databases, storage, execution environments, third-party APIs and, in some cases, payments made to access data or content. As ever, the least visible costs are often the most difficult to control.

Where a conventional application will generally return an error, an agent can attempt to resolve it, which is precisely one of its strengths. But this software persistence comes at a cost, and every new attempt adds another line to the bill.

The cost that matters is therefore no longer that of an isolated call and its immediate output, but that of a successfully completed task. A model priced ten times lower delivers no savings if it must be queried fifteen times. Conversely, a frontier model can remain economically justified when an accurate answer on the first attempt avoids a succession of costly failures.

The right equation must therefore compare the total cost of models, tools, retries and controls with the number of assignments actually completed. The result is less elegant than a price per million tokens, but infinitely more useful when calculating a margin.

Giving every execution an economic identity

Founded eleven months ago by Ilan Zerbib, SAPIOM presents itself as production infrastructure for AI agents. The startup says it has processed more than 270 million transactions and now handles over 100,000 agent runs per day.

One clarification is required: the term “transaction” may refer to a model call, the use of a tool, an execution step or a payment operation, and should not be confused with a complete assignment. SAPIOM’s value proposition lies less in the volume announced than in the way it intends to make that volume intelligible.

Each execution is assigned an identifier linking its various stages, the models involved, the tools called, their cost, duration and outcome, as well as the number of attempts required. When an assignment fails, the company can locate the relevant stage, understand what the agent attempted and observe how it responded.

This traceability makes it possible to move from a bill that merely indicates the volume of resources consumed on a platform, whether ANTHROPIC, OPENAI or GOOGLE, to a more detailed analysis identifying the product, customer, process or cost center responsible. That expense can then be compared with revenue, savings or human time recovered. Token volume should be treated as one technical data point among others, rather than as an indicator of economic performance.

Controlling expenditure before it occurs

Observability only solves part of the problem. A dashboard may explain that an agent has just exceeded its budget, but the money has already been spent. SAPIOM therefore aims to intervene at the moment the agent attempts to act, before calling a model, using a tool or purchasing a service.

A company can define the agent’s mandate in advance: which models it may use, which providers it may send data to, the maximum cost of an assignment, the tools it may access and the operations requiring human approval.

The distinction from a conventional monitoring tool is fundamental. “Observability tells you what happened after the agent acted; SAPIOM determines what it is allowed to do before it acts,” explains founder Ilan Zerbib. The startup wants to turn every potential expense into a controlled decision.

This logic can be applied across several levels of autonomy. A low-cost, reversible action may be executed freely; spending above a certain amount may require approval; access to sensitive data may be restricted to specific models or regions; and an abnormal succession of failures may trigger an automatic shutdown. The agent becomes more autonomous, but that autonomy remains bounded by a budget, permissions and escalation rules.

Reserving frontier models for tasks that justify them

The second source of savings lies in model selection. Many agents are developed around the most capable model available at the time they are designed, then continue using it for every operation. Basic classification, data extraction and complex reasoning are consequently billed at the same rate, as though a company systematically called on its most experienced executive to open the mail.

SAPIOM Router aims to select, for every call, the execution path offering the best balance between cost, quality, latency, availability and internal policy. A simple task can be sent to a less expensive open model, while a complex or sensitive operation remains assigned to a frontier model.

The principle appears obvious, but its execution is far less so. Before receiving the response, the router must determine whether a model can deliver the expected level of quality. A poor decision may reduce the price of the initial call while increasing the number of retries or human checks. Optimization therefore does not consist of mechanically choosing the cheapest option, but of estimating the full cost of a satisfactory result.

This capability also depends on evaluations conducted against real workloads. A model that performs exceptionally well on a general benchmark may prove mediocre within a particular business process, while a more modest model may execute a highly structured task with consistent reliability. The more executions, errors and corrections a platform observes, the more it can, in theory, improve its routing decisions.

This is also where SAPIOM hopes to build a defence against the rapid commoditization of this capability. In May 2026, OPENROUTER claimed to process some 25 trillion tokens per week and already offered unified access to more than 400 models. AMAZON has integrated intelligent routing into Bedrock, MICROSOFT is developing it within Foundry, while open-source routers allow companies to build their own arbitration systems. Simply directing requests from one model to another may therefore become a feature rather than a standalone market.

To escape this trap, SAPIOM combines Router with Agent Studio, designed for building agents, and, more importantly, Runtime, which handles execution, state preservation, retries and traceability.

Preventing agents from retrying indefinitely

An agent can also cause consumption to spiral because it does not know when to stop. When an external service fails to respond, a request is poorly formulated or a stage produces an ambiguous result, the system may change its approach, call the tool again, reconstruct its context and resume part of its reasoning.

This capacity to adapt distinguishes an agent from a rigid workflow, but it also opens the door to expensive sequences in which each individual step appears locally reasonable while the overall process produces nothing useful.

The runtime must therefore enforce limits: a maximum number of steps, retry caps, a maximum duration, a budget per assignment, fallback models and escalation to a human after several failures. It must also preserve execution state so that the agent can resume at the point of failure instead of restarting the entire assignment.

Cost management and reliability engineering converge at this point. An intelligent retry can generate greater savings than a reduction in token prices, just as stopping early may prove more valuable than a tenth stubborn attempt.

From FinOps to AgentOps

Many companies are beginning to understand the scale of the problem. In its Global AI Pulse published in June 2026, KPMG reported that only 7% of the 2,145 executives surveyed had established a return on their AI investments. Forty-two percent had only partial visibility into their spending, 23% struggled with usage-based costs and one-third found it difficult to understand token-related pricing structures.

The study also found that organizations with strong cost visibility were five times more likely to report an established ROI. Correlation alone does not prove that a dashboard creates profitability, but it does recall a fairly robust rule of management accounting: what a company cannot attribute is rarely optimized.

An AgentOps discipline is consequently beginning to emerge at the intersection of FinOps, security and software operations. The CTO can no longer manage alone an expense that directly affects margins; the CFO cannot analyze it without understanding the mechanics of execution; the CISO must control access and data; and business teams must define the expected outcome.

Every agent must therefore have an owner, a budget, an objective and termination conditions. Performance indicators should track the average cost of a successful task, the first-attempt success rate, the number of retries, the share of calls sent to frontier models, the proportion of assignments requiring human intervention and, above all, the margin retained after AI consumption.

A new infrastructure layer, or a feature destined for the cloud?

SAPIOM raised $35 million in a Series A led by DRAGONFLY, six months after announcing a $15 million seed round, bringing its reported total funding to $50 million. ACCEL, ANTHROPIC, OKTA VENTURES, MENLO VENTURES and ARRAY VENTURES are among its investors.

Among its challenges, SAPIOM must demonstrate that its control layer can remain independent. AWS AgentCore already brings together runtime, identity, permissions and observability; MICROSOFT Foundry combines hosting, models, access control and agent monitoring; and GOOGLE is integrating agentic payment protocols into its platform. In China, ALIBABA CLOUD and TENCENT are directly embedding these capabilities into their own platforms.