Your AI agents are getting cheaper. Your business might not be.

For years, technology financial management has helped organisations understand their technology expenditure, allocate costs and improve the value they receive from their investments. Established FinOps practices provide much of the foundation needed to manage AIconsumption.

The challenge is that agentic AI introduces greater variability into the cost of getting work done. Unlike a conventional process with relatively predictable execution requirements, an AI agent may take several different routes to complete the same task, consuming different amounts of resources along the way.

One customer request might require a straightforward response. Another could involve several rounds of reasoning, multiple system interactions and an escalation to a human employee. Both may appear in a management report as one completed transaction. Yet the cost of delivering them could be materially different.

In August 2026, Gartner predicted that inference costs per agentic workflow would increase more than fivefold through 2028, as AI applications became more complex. It is a forecast, not evidence that every organisation will experience the same increase. But it highlights an important point that is well worth noting: cheaper AI resources do not automatically make the work they perform cheaper.

Token price is only one part of AI workload economics. Context size, model choice, model calls, retries and escalation help explain why the cost of an accepted outcome changes; they do not establish whether that outcome is worth the consumption required to produce it. An executive looking only at token prices or aggregate AI expenditure could therefore miss what is happening inside the business process.

Consider a bank using an AI agent to support customer onboarding. The agent retrieves documents, validates information, identifies missing details and prepares an application for review. The bank could measure the agent's token consumption and the cost of each modelinteraction. That would provide useful information about technical efficiency.

But imagine two versions of the same agent. The first uses a relatively inexpensive model. It performs the initial checks quickly, but struggles with incomplete applications. Employees frequently have to correct its work, and customers sometimes need to resubmit information.

The second uses a more expensive model. It costs more per interaction, but completes more applications correctly, requires less human intervention and produces fewer avoidable delays. Which is cheaper?

That question cannot be answered from the model invoice. The bank needs to understand what it actually costs to complete an application that meets the required standard.

That includes AI inference, supporting system costs, retries, human review, exception handling and any rework required to get the application right. It also includes an appropriate share of the costs of maintaining and operating the capability. Only then can management make a meaningful comparison.

The more expensive model might produce a lower cost per successfully completed application. Or its additional intelligence might make little difference, leaving the cheaper model as the more economical choice. The answer depends on the work. This is why the full cost of producing an acceptable business outcome should be measured, not just the cost of the AI used to produce it.

But there is an important distinction. A lower cost per completed outcome does not automatically mean that the business is better off. The bank still needs to compare the AI-enabled process with how the work was previously performed, understand what has improved and determine whether those improvements justify the investment. And that is exactly where the realeconomic conversation begins.

An agent completing a task is not necessarily the same as the business accepting its work. A claims-processing agent may produce a recommendation, but was it accurate? Did it comply with therelevant policy?

Was the decision approved where required? Did the customer receive an acceptable outcome?

These distinctions affect the economics. An organisation could improve its cost per completed task, while increasing incorrect decisions, downstream rework or customer complaints. The apparent efficiency gain would then be misleading. Before scaling an agent, business owners and technology teams need to agree on what constitutes an accepted outcome, which exceptions require human intervention and what standard of performance must be maintained.

They also need to compare equivalent work. An agent handling straightforward requests should not appear more economical simply because another is dealing with more difficult cases. Without that consistency, it becomes difficult to compare different models, agents or workflow designs reliably.

Once the organisation has defined the outcome, it can make more informed choices about how the work should be performed. A simple request may be adequately handled by a smaller, less expensive model. A complex request involving financial exposure or regulatory requirements may justify additional reasoning, stronger controls or humanapproval.

Some tasks may not need an AI agent at all. A conventional rules-based process could be cheaper, more predictable and entirely sufficient. The objective should be to use the resources appropriate to the work, rather than deploy the most sophisticated AI capability wherever possible.

But the economic question does not end with selecting the right model or reducing the cost of a completed task. Consider an agent that reduces the time employees spend processing applications by 30%. That could represent a meaningful improvement in productivity.

But what happens to the capacity that has been released? If the organisation continues to incur the same employment costs and the additional capacity remains unused, that improvement has not automatically become a financial saving. The value might instead come from serving more customers, improving service quality or absorbing future growth without a proportional increase in staffing.

Those may all be worthwhile outcomes. But they are different forms of value, and the business needs to understand which benefits are actually being realised. This changes the conversation between the CFO, COO and technology leadership. Instead of focusing only on whether AI is becoming cheaper to operate, they need to understand whether the process is performing better, what that improvement is worth and whether the full cost of achieving it makes economic sense.

Technology teams can explain what an agent costs to run and where that cost comes from. They can identify inefficient model usage, excessive retries, expensive system interactions and opportunities to improve the underlying architecture. What they cannot determine alone, iswhether the resulting business outcome justifies the expense.

That requires someone who understands the economics of the process being performed. An agent may cost more to operate, but materially reduce customer abandonment, improve recoveries or enable additional profitable transactions. Its increased consumption could be entirely justified.

Equally, an agent may perform a task impressively, without producing any meaningful operational or financial benefit. That agent may not deserve further investment. The business owner needs to be accountable for the outcome, supported by technology, finance and risk teams.

This does not require creating a separate financial governance structure for every agent. Existing financial management and operational accountability can be extended to include AI consumption, the performance of AI-enabled processes and the benefits those processes produce. The important change is connecting those measures so that someone can make an informed business decision.