Your AI Agents Can Spend Money. Is the Board Watching?

As AI moves from answering questions to doing work, boards need to understand a new governance challenge: not how many tokens the company consumes, but what authority AI agents have, what resources they consume, and what value they create.

For many companies, the economics of artificial intelligence initially looked familiar. Buy software, purchase licenses, give employees access, put the expense into the technology budget, and measure adoption. AI agents are beginning to change that equation.

Unlike a chatbot that responds to a question, an AI agent can reason through a problem, retrieve information, use tools, take actions, retry approaches, and execute multi-step workflows. Those activities consume computing resources and can generate costs.

An AI agent does not need a corporate credit card to “spend money.” But as companies deploy more agents and give them increasingly complex work, those systems can consume corporate resources while pursuing the objectives they have been given.

That creates a new corporate governance question: What are AI agents authorized to do, how much corporate resource can they consume, and what economic value is the company receiving in return?

From AI That Talks to AI That Works

The difference between a chatbot and an AI agent might sound technical. Economically, however, it can be significant.

Traditional generative AI might involve an employee asking a question and receiving an answer. An agentic system can undertake a sequence of activities to accomplish an objective, potentially consuming additional resources every time it reasons, retrieves information, interacts with tools, or attempts another step.

Research from Microsoft Research illustrates how dramatically those economics can differ in some workloads. In a 2026 study involving eight frontier language models and agentic coding tasks, researchers found that the agentic tasks studied consumed 1,000 times more tokens than code reasoning and code chat. Runs involving the same task could also differ by as much as 30 times in total token usage, and greater token consumption did not necessarily translate into greater accuracy. [microsoft.com]

Those numbers should not be generalized to every agent or business process. The research concerned agentic coding tasks specifically. But the findings illustrate a broader management problem: agentic AI consumption can be large and highly variable, and more AI consumption does not necessarily produce more value.

A token, put simply, is a unit of data processed by an AI model. Directors don’t need to understand the mechanics of tokenization. They do need to understand the underlying economic principle: AI work consumes computational resources, and multi-step agentic work can substantially increase that consumption.

From Seats to Consumption

That creates an important change in how companies may need to think about enterprise software economics.

Historically, one simplified way of thinking about software expenditure was:

Employees × licenses = relatively predictable software expense

Agentic AI adds another economic model:

AI work performed × resources consumed = potentially variable expense

That second equation is not an industry pricing formula. It is simply a way of illustrating the change.

You can already see elements of this shift in Microsoft’s approach. Microsoft describes Copilot Credits as a common currency in a usage-based model across supported AI experiences. Its guidance specifically says that as AI workloads become more advanced, particularly in multi-step, agent-driven scenarios, the work required to complete a task can vary significantly. The number of credits consumed for a response, action, or operation depends on the complexity of the task performed. [cdn-dynmed...rosoft.com]

Importantly, this does not mean subscriptions are disappearing. Microsoft’s own guidance describes its usage-based approach as complementing subscription-based services. [microsoft.com]

For boards, the broader point matters more than one vendor’s pricing model. Enterprise AI expenses can increasingly be influenced not only by how many people have access to AI, but also by what AI systems are doing, how frequently they are doing it, and how much computational work those activities require.

The AI Token Paradox

That leads to one of the most counterintuitive aspects of AI economics:

AI can become cheaper and still become more expensive to the enterprise.

Accenture, citing projections from Goldman Sachs, says token use could increase 24 times by 2030, while cost per token could decline 60% to 70% annually. Its broader point is that falling token prices combined with rapidly increasing usage can still drive total enterprise spending higher. [newsroom.a...enture.com]

The paradox is straightforward:

Unit cost of AI ↓AI consumption ↑Total enterprise AI spending can ↑

There is nothing inherently wrong with that. If a company’s AI spending doubles while the economic value created increases fivefold, spending more could be an excellent allocation of corporate resources.

The important governance question is therefore not whether AI spending is rising. It is whether the value created is rising with it.

The Real Question Is Value

This is where the economics become particularly important for directors.

Accenture’s 2026 research, based on a survey of 750 senior executives across 17 countries and interviews with 15 technology and finance leaders at Fortune 500 companies, found that four out of five executives said AI was delivering measurable outcomes. Yet less than one dollar in five of enterprise token spending could be traced to a quantified financial outcome, such as revenue influenced, cost avoided, or productivity converted into a financial measure. [accenture.com]

That goes directly to the governance issue. The problem is not necessarily that companies are spending too much on AI. Companies should be prepared to increase AI investment when the returns justify doing so. The problem arises when management cannot clearly connect increasing AI consumption with measurable economic value.

For boards, that is fundamentally a capital-allocation question.

The Token Bill Isn’t the Whole Bill

There is another reason boards should not become overly focused on tokens: the token invoice does not capture the full economic cost of deploying AI.

EY identifies seven categories involved in the cost of AI agents: consumption, platform, infrastructure, governance, workforce, failure and recovery, and compliance. In separate 2026 analysis, EY estimated that the full enterprise cost of AI may be roughly three times the token invoice when infrastructure, software, governance, organizational change, expected failure, and regulatory costs are included. [ey.com] [ey.com]

The implication is important. A board focused exclusively on token costs might be looking at the most visible expense while missing the broader economics of deploying agentic AI.

AI Agents Introduce Two Kinds of Authority

This is where the issue moves beyond technology spending and becomes a corporate governance question.

Companies have long established rules around delegated authority. Executives have budgets. Employees have approval limits. Certain transactions require additional authorization. Important decisions are escalated.

Agentic AI raises the question of how some of those principles should apply when software begins performing increasingly consequential work.

Boards should think about two different forms of authority.

Economic Authority

How much corporate resource is an AI agent permitted to consume?

The operational answer belongs to management, but the governance principle is familiar. Organizations already put boundaries around how human decision-makers can deploy corporate resources. As autonomous systems undertake more work, management should consider appropriate limits, controls, and escalation mechanisms around material agentic activities.

Decision Authority

What is an AI agent permitted to do on behalf of the company?

This could ultimately be more consequential than token consumption itself. An agent that summarizes documents raises one level of governance concern. An agent that can interact with business systems or take actions creates another.

The questions should sound familiar to directors: What authority has been delegated? Where are the limits? What requires human approval? When is escalation required? How is performance monitored? And who remains accountable for the outcome?

The technology is new. The underlying governance principles are not.

Management Manages AI. The Board Oversees the Governance.

This distinction is critical. Management’s responsibility is to select models, design workflows, manage AI consumption, establish controls around agents, monitor performance and costs, and determine how AI is deployed across the organization. Those are management and operating decisions.

The board’s responsibility is different. Directors should determine whether management has established an appropriate governance framework, whether material AI investments are producing sufficient value, whether significant risks are being managed, and whether the authority being delegated to autonomous systems is consistent with the company’s strategy and risk appetite.

The board should not become an AI operations committee. Nor should directors be debating how many tokens an individual workflow ought to consume.

Instead, the board should be able to determine whether management understands what significant AI systems are authorized to do, what resources they consume, what risks they create, and what economic value the company receives in return.

Please Don’t Put Token Consumption on the Board Dashboard

As tokens become increasingly measurable, companies could make an understandable mistake: add token consumption to the board dashboard and ask directors whether the number is too high.

Token consumption alone says very little about business performance. A billion tokens that generate substantial incremental economic value could represent excellent economics. The same billion tokens consumed by inefficient workflows that contribute little business value could represent waste.

BCG similarly argues that companies should focus on cost per outcome, rather than simply token expenditure. Its proposed return-on-AI approach relates economic return to the costs of human intelligence and tokens used to produce business outcomes. [bcg.com], [web-assets.bcg.com]

So instead of asking,“How many tokens did we use?”, directors should ask:

What business outcome did we receive from the AI resources we consumed, and was the return sufficient to justify that consumption?

In my view, token consumption may be a useful management metric. The economic productivity of AI is the more important governance concern.

Four Questions Boards Should Ask

Directors don’t need to approve every AI agent or interrogate individual token invoices. They do need confidence that management has established an appropriate governance structure around materially important uses of agentic AI.

1. Where are we giving AI agents meaningful authority?

Directors should understand which financially, strategically, or operationally important activities are moving from AI assistance toward agentic execution. The objective isn’t to receive an inventory of every AI application in the company. It is to have visibility into areas where agentic AI could materially affect the business.

2. What are the boundaries of that authority?

Management should be able to explain both decision authority and economic authority. What can important agents do? What resources can they consume? Where is human authorization required? What circumstances require escalation?

3. What does an AI-enabled business outcome actually cost?

Management should move the discussion beyond cost per token and toward the cost of producing a meaningful business result. That allows directors to distinguish AI activity from AI productivity.

4. Can management demonstrate value as AI scales?

If AI consumption doubles, triples, or eventually increases tenfold, can management demonstrate corresponding increases in productivity, revenue, margin improvement, risk reduction, or other measurable outcomes?

This question becomes particularly relevant given Accenture’s finding that less than one dollar in five of enterprise token spending in its research could be traced to a quantified financial outcome. [accenture.com]

Together, these questions keep directors at the appropriate governance level. The board isn’t choosing models or determining how many tokens a workflow should consume. It is overseeing whether management has the framework, controls, accountability, and measurements necessary to govern AI responsibly as it scales.

AI Is Becoming an Economic System

Generative AI initially looked like another productivity tool. Agentic AI increasingly looks like a scalable input into how companies perform work.

That changes the board conversation. The first phase of AI governance largely focused on who could use AI, what information could be given to it, and what risks accompanied its use. Agentic AI introduces another set of questions: What can AI do? What authority has management delegated to it? What resources can it consume? What happens when it scales? And can management demonstrate sufficient economic value in return?

Those are no longer simply technology questions. They are corporate governance questions.

The Question I Would Put in Every Boardroom

Boards don’t need to count tokens. They need to know whether management understands what significant AI agents are doing, what authority they have, what resources they consume, and what value they create.

One question cuts through much of the technical complexity:

If our AI consumption increases tenfold, can management demonstrate that the economic value created by that consumption is increasing with it?

If the answer is yes, increased AI expenditure may be an investment worth accelerating. If management cannot answer the question, the problem isn’t necessarily that the company is spending too much on AI. The deeper problem may be that management does not yet know whether its increased consumption of machine intelligence is producing sufficient economic value.

The company may have an AI strategy. It may not yet have an AI economic governance strategy.