As the agentic era reshapes the AI economics of enterprise technology, organisations have an opportunity to govern run-rate exposure as a growth investment, protecting operating budgets and margin while compounding enterprise value.

In brief

• Agentic AI is shifting enterprise AI from predictable software/labour costs to variable compute consumption, where token costs are only one part of the full cost picture.
• Organisations need agentic FinOps to manage the total cost of agents, including infrastructure, governance, organisational change, failure recovery and emerging regulatory risk.
• Leaders should treat agents as growth investments, with clear ownership, cost controls and value metrics so agents are scaled based on measurable return, not just adoption.

The loudest conversation about AI right now is about what it costs. The more important one is about what it earns. Effective governance treats that spend as a growth investment, directing it toward measurable enterprise value.

Token costs are the most visible line item in an AI budget and are often the first sign that the economics of agentic AI are changing. They are not the full cost, but they matter because they are where inference intensity, model choice, orchestration complexity and upstream compute scarcity first surface, often only after the work is done.

But while tokens are the story today, the real economics of running agentic workflows at scale include infrastructure, operations, people, risk and the hidden cost of engineering around AI’s limitations. Optimising tokens without understanding total cost of ownership is like managing a factory by watching the electricity bill.

The story previously presented to many boards was straightforward: an expensive predictable line of human labour traded for a cheaper predictable line of software. But an agent is not a licence. It lives in an operating stack of compute, models, data pipelines, governance, oversight and redesign work, each of which compounds rapidly. As a result, the tokens AI consumes, serving as the metered proxy for much of that operating stack, become a highly variable operating cost.

Consider a customer service AI assistant built two years ago to answer product questions. A chat that once cost $0.04 may now involve tool retrieval, planning and subagents that turn it into a $1.20 orchestration. Additional costs — such as knowledge-base updates, agent evaluation and human-collaboration design — may not appear on the model vendor’s invoice but still surface to the enterprise.

When a CFO asks what one agent costs, they are often shown a vendor invoice that captures only part of the total financial exposure.

The challenge is that agentic AI shifts enterprise AI from a fixed-cost labour comparison to a dynamic compute consumption model. Boards and CFOs are therefore looking for investment cases that capture the full operating picture, including usage, orchestration, governance, change, risk controls and remediation.

This also changes how AI is budgeted and managed. CFOs need visibility into consumption across use cases. CTOs need to understand inference volume, model mix and retrieval load. CEOs need to assess not only labour savings, but the ongoing technology run rate required to deliver them.

According to a Gartner poll, more than 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. These outcomes are not inevitable. They are more likely where organisations scale agents before establishing agentic FinOps to manage cost, risk, usage and value together.

How to break down the cost of an agent

Most companies only include costs 1-3 in their agentic investments and business cases, with costs 4-7 often emerging later in the lifecycle as agents scale.

1. Tokens and API calls

The cost that shows up directly on the invoice; i.e., input and output token volume, model selection, reasoning intensity and retries. Agentic workflows can consume hundreds of thousands of tokens in a single session, much more than the hundreds needed to support a more traditional generative AI chat experience.

2. Subscriptions and licenses

The fixed commitments made before any agent runs; i.e., model provider contracts, SaaS and Orchestration platform licences, committed-use agreements with cloud and AI vendors.

3. Platform infrastructure

The compute and services that keep agents running but never appear on the model vendor’s invoice; i.e., orchestration runtimes, sub-agent steps, application environments. This typically lands on the cloud consumption bill and gets filed as infrastructure. As agents scale, GPU availability, power and architectural choices become inputs to manage rather than assumptions to inherit. This is a theme that the next papers in this series will return to.

4. Governance burden

The incremental investment required to keep agents safe, auditable and compliant; i.e., guardrails, cyber protections, new attestation procedures and human-in-the-loop reviews. These vary depending on the nature of the industry regulatory landscape and often manifest themselves as different teams doing “work” against the initiative. This cost compounds as agents scale out but becomes more predictable when designed in from the start.

Most companies only include costs 1-3 in their agentic investments and business cases, with costs 4-7 often emerging later in the lifecycle as agents scale.

5. Organisational change

The cost of the transformation to reorganise the operating model around the new agentic workflow; i.e. change management, workforce retraining, role-redesign, human-in-the-loop architecture, labour relations, and in some industries, union negotiations. These investments are front-loaded against each workflow an agent enters, but then often recur with every major model upgrade or capability change.

6. Expected failure and recovery

The probabilistic cost of agent failures, including hallucination remediation, back testing, reputational damage and poorly governed agents without kill switches. These costs are zero until they are significant, and most organisations are only beginning to plan for expected loss or insurance. Black swan failures, such as a coding assistant deleting a core database, are too catastrophic to price and must be designed out entirely, via hard-coded boundaries, write-protected environments and human-in-the-loop gates for irreversible actions.

7. Potential AI taxes for agents (speculative)

The regulatory cost that does not exist yet but has already been signalled. Various governments and regulatory bodies are exploring AI-specific reporting obligations and potential tax on AI agents as a result of job loss. This carries a major question mark today. Price volatility, capital intensity and concentration risk could ultimately reshape how AI is governed, priced and consumed, a shift a future paper in this series will examine in depth.

The practice of agentic FinOps does three things at once:

• It allows enterprises to estimate the costs, including lag.
• It assigns an owner to each row before the spend, not after the invoice.
• It gives leadership the basis to decide which initiatives earn the right to scale based on the value they unlock against their fully loaded cost. 

The table below is our working framework for agentic FinOps, illustrating how agent costs are organised and evaluated.

The key challenge: The total cost of an agent, at most organisations, is structurally invisible until designed for visibility. This is not because the costs are hard to capture, but because they surface irregularly across uncoordinated budgets.

The capital cycle where agents get weighed instead of counted

The era that exploration is giving way to is one that rewards discipline. Investment numbers are only meaningful when tied to value. Every agent should have a value metric on day one: output, revenue, productivity, speed, quality, risk reduction or decision improvement it is expected to produce. No agent should be approved, scaled or renewed without a clear measure of what it produces per dollar.

That requirement moves the discussion from spending to return. It gives the Head of Agent Economics a practical basis to shift investment away from agents that consume resources and toward agents that create compounding value over time.

The next AI cycle will not be won by companies with the most agents, the best models or the largest spend. It will be won by those who treat agent capacity as capital and allocate it deliberately, like any other investment driving growth, margin and enterprise value.

What’s next in the EY Total Cost of Agents series

This whitepaper is the first edition of the EY Total Cost of Agents series. Other papers in this series will examine:

• The path to break-even and the pivot from productivity gains to growth
• The strategic management of compute as a critical supply chain asset
• The policy and regulatory landscape that will shape how AI is governed, priced and consumed in the years ahead.