When AI Agents Compete for Budget, Which Ones Deserve More?
Most AI agents are easy to justify one at a time.
A customer-service agent cuts handling time. A coding agent helps developers. Another supports procurement, finance or analytical work. Each has a sponsor. Each can point to activity. Some can show faster work, better service or lower effort.
Then Finance asks a different question: If there is another pound to invest, which agent gets it?
Live agent counts, usage and token spend do not answer that.
At the experimental stage, that is fine. Organizations need to learn what agents can do, where they fail, what controls they need and which workflows are worth redesigning. Some duplication and uncertainty are part of that learning.
The economics matter more when agents start competing for operating budget.
The token bill is visible. The full cost often is not. Cloud infrastructure, licenses, orchestration, governance, cyber controls, human oversight, process redesign, training, failure recovery and remediation can sit in different budgets. A cheap model invoice can therefore make an expensive agent look cheap.
The opposite is also true. An agent with a higher visible cost may deserve more investment if it changes something materially valuable in the business.
EY has put useful structure around this problem.
In its 15 July 2026 paper, Unlocking Agentic Value: a new investment discipline for the agentic era, EY argues that token costs are only part of the cost of running agents. It separates agent economics into seven areas spanning model usage, licences, platform infrastructure, governance, organizational change, expected failure and recovery, and possible future regulatory costs.
The categories matter less than the consequence. Agent economics can sit across several budgets at once, so the person looking at the model invoice may be seeing only part of the exposure.
EY describes this as Agentic FinOps and argues that leadership needs enough visibility to decide which initiatives earn the right to scale based on the value they unlock against their fully loaded cost.
That is the shift: every agent must earn the right to scale. Technical success, usage and low token spend are useful signals. Capital allocation depends on what the agent changes in the business, what it costs end to end, and whether the next pound would earn more elsewhere.
EY puts it neatly when it says agents should be “weighed instead of counted”. It also argues that every agent should have a value metric from day one and should not be approved, scaled, or renewed without a clear measure of what it produces per dollar.
That does not mean reducing every agent to one simplistic ROI number.
A risk-reduction agent and a sales-conversion agent need different evidence. Value may show up as revenue, productivity, speed, quality, lower risk or better decisions. The measure must fit the job the agent is expected to do.
For each material agent, leaders need to know three things: what it is expected to change, what it really costs to run and what evidence would justify putting more scarce capital behind it.
That evidence also needs to establish a credible connection between the agent and the business effect. Rising adoption alongside an improving KPI does not, by itself, show that the agent created the value. Leaders should understand whether the agent contributed to the change, and whether that change is material enough to justify further investment.
Without that discipline, activity can become its own defense. A heavily used agent may have vocal internal support. A technically impressive agent may be strategically fashionable. A low-cost agent may look efficient. None of that tells a CFO whether it is the best use of the next pound.
A mature Finance function, FinOps capability, AI governance process or portfolio office may already be able to make these decisions. If it can see the relevant cost, understand the intended business effect and compare competing uses of capital, there may be no reason to create another management layer around agent economics.
EY proposes a Head of Agent Economics or Agent FinOps Lead as one way to centralize accountability. That is an option, not a universal requirement. The role matters less than the decision discipline.
A practical test is to take the five largest agent initiatives in the organization and ask what is actually known about each one: the fully loaded cost beyond the model invoice, the business effect it is meant to create, the evidence that effect is happening, and the point at which the organization would invest more, redesign the agent or stop it.
If leaders can answer those questions with credible evidence, the economics may already be under control. If they cannot, adding more agents will make the allocation problem harder.
Agentic AI will continue to raise technical questions about models, orchestration, security, and governance. Once significant capital is involved, it also becomes a very ordinary investment question.
If you had to cut 20 percent of your agent portfolio tomorrow, what evidence would tell you which agents to protect?
Sources:
EY, “Unlocking agentic value: a new investment discipline for the agentic era”, 15 July 2026.