1. The Hidden Cost of AI
AI is moving beyond small experiments. Companies are now using it to support finance, research, customer service, and other core business tasks.
But unlike many traditional software tools, some AI services charge based on usage. One key measure is the token: a small piece of text that an AI model processes or produces. The more a system uses AI, the more tokens it may consume, and the higher the bill can become.
That makes costs harder to predict. A team might start with a small trial, then use AI far more often than expected. More complex tasks can also require many rounds of work between the AI model and other tools.
Esker experienced this problem firsthand. According to reporting by CFO Dive, its AI costs ran about four times over budget as usage grew and token limits were repeatedly exceeded.
The problem extends beyond one company. In its 2026 Global Finance AI Trust Index, Esker surveyed 338 finance leaders. Seventy-two percent said their organizations had spent more than planned on AI initiatives over the previous year.
The message for CFOs is clear: approving an AI budget is one thing. Predicting what the business will actually spend is another.
2. What Does AI Add to the Cost of an Employee?
Esker is taking a closer look at how AI costs vary across its business.
Its finance team reviews usage data from AI providers, estimates the monthly cost for different functions, and calculates a running cost per employee. It then uses those figures to help forecast spending for the year ahead.
The reason is simple: not every employee uses AI in the same way.
A person using AI occasionally to draft emails may generate little expense. Someone in research or finance who runs complex tasks throughout the day may use far more.
McDermott has noted that finance and research and development teams are among the groups driving significant usage overages.
If all those costs sit inside one general IT budget, managers may struggle to see which teams are driving spending and whether the extra cost is worth it.
A more complete view of employee costs could look like this:
Salary and bonuses
Benefits and payroll taxes
Software and other work tools
AI usage and computing costs
This is not a rule that every company must follow. Some costs are shared across teams and need to be allocated carefully. But the principle is useful: if AI is helping a team produce work, finance should understand what that AI costs and what the business gets in return.
That can give CFOs a better view of department costs, product margins, and the true cost of delivering work to customers.
3. Why AI Agents Can Make Costs Harder to Predict
The next challenge is the rise of agentic AI.
Traditional AI tools often respond to a single request. AI agents can do much more. They may break a task into steps, use other software, check their work, and repeat the process until the task is complete.
That can make them more useful. It can also make them more expensive.
Futurum Research, in a report produced with AI infrastructure provider QumulusAI, estimates that some agent-based tasks can use 10 to 100 times more tokens than a simple AI request.
The difference matters because more steps can mean more computing use and higher costs.
Consider a simple example. Asking AI to summarize a document may require one request. Asking an AI agent to review hundreds of documents, compare the results, check for errors, and prepare a report could require many requests behind the scenes.
The user may see one finished report. The company may be paying for a much larger amount of AI activity.
Futurum's report also points to a wider change in the market.
AI computing metric | Figure |
Estimated data center inference spending in 2025 | $120 billion |
Forecast spending in 2030 | $885 billion |
Reported growth in agent and reasoning inference in 2026 | 219% year over year |
These figures reflect the report's estimates and forecasts, not guaranteed outcomes. The report was produced with QumulusAI, which has a commercial interest in alternatives to pay-per-use computing.
Its broader point is still worth considering: as AI use grows, companies may need to rethink how they buy computing power.
Paying for each request can make sense when usage is low or unpredictable. For companies with steady, heavy workloads, reserving computing capacity may offer better cost control. The right choice depends on actual usage, pricing, and business needs.
For CFOs, the priority is to understand the costs before they begin to affect margins.
4. Spending More Does Not Always Mean Getting More
Controlling AI costs is only half the problem. The other half is proving that the spending delivers value.
Esker's 2026 Global Finance AI Trust Index highlights the gap:
75% of CFOs reported higher AI software and subscription costs.
65% of CFOs said they had difficulty linking AI use to specific business results.
58% of finance leaders identified productivity as AI's biggest benefit.
66% of finance leaders knew or suspected that employees were using unapproved AI tools.
These results point to a difficult question: companies may know how much they spend on AI without knowing exactly what that spending achieves.
An AI tool may help employees finish tasks faster. But does that time saving reduce costs, allow the team to handle more work, improve accuracy, or increase revenue?
Those are different outcomes. A productivity gain does not automatically translate into a financial gain.
CFOs need to look beyond usage reports and ask what has changed in the business.
For example:
Has the cost of completing a task fallen?
Can the team handle more work without adding staff?
Has AI reduced errors or shortened delays?
Has the extra spending improved revenue or profit?
Without answers, a company risks treating AI adoption as a success simply because more employees are using it.
The goal is not to spend less on AI at any cost. It is to know when spending more makes financial sense.
5. Tracking Costs Is Good. Preventing Waste Is Better.
Companies have several ways to manage AI spending. Two broad approaches are becoming important.
Cost reporting tools | AI request gateways | |
Main purpose | Show where money is going | Help control spending before requests reach AI providers |
How they work | Collect and analyze usage and billing data | Apply rules and limits to AI requests |
Main benefit | Help identify expensive teams, tasks, and unusual usage | Help prevent uncontrolled use |
Examples | Finout, Vantage, CloudZero | LiteLLM, Portkey, Databricks AI Gateway |
Cost reporting tools help finance teams see where the money went. They can show which teams use the most AI, identify unusual spending, and help managers understand the cost of different workloads.
But reporting alone may not stop an expensive task from running.
AI gateways can sit between an application and an AI provider. Depending on how they are configured, they can apply spending limits, control request rates, or block requests that exceed set rules.
Imagine an AI agent running overnight and repeatedly trying to complete a task. Without suitable controls, that activity could generate an unexpectedly large bill. A gateway with the right limits can help reduce that risk.
Many companies may need both approaches: reporting to understand costs and controls to manage them.
The important point is that AI spending should not depend on someone noticing a problem after the bill arrives.
6. Finance Teams Still Need Human Judgment
There is another risk that CFOs cannot ignore: employees may become too comfortable accepting AI answers without checking them.
AI can help teams analyze information, prepare reports, and recommend actions. But it can also produce incorrect answers, miss important details, or make recommendations based on incomplete information.
In finance, those mistakes can matter.
A forecast may look convincing but rely on a weak assumption. A cost-saving recommendation may ignore a risk elsewhere in the business. A hiring decision may seem logical on paper but fail to account for what a team actually needs.
McDermott has warned about the risk of finance teams relying too heavily on AI without developing the judgment needed to challenge its output.
Esker has responded with an internal training approach that uses experienced AI users, known as champions, to help colleagues learn how to use the tools.
The aim is to make adoption easier while encouraging employees to question the answers they receive.
That balance matters. Companies want employees to get more value from AI, but they also need people who can spot mistakes, understand the business context, and take responsibility for decisions.
For CFOs, successful AI adoption is not simply about giving employees better tools. It is about building a team that knows when to trust those tools and when to challenge them.
7. What CFOs Should Take Away
Esker's experience points to a wider change in financial planning.
As AI becomes part of everyday work, CFOs need a clearer view of what it costs, where that cost comes from, and whether the business is getting enough value in return.
Three steps can help:
First, track AI usage where it happens. Understand which teams and tasks drive spending rather than leaving every cost inside a general technology budget.
Second, forecast and control costs. Use actual usage to improve budgets, set sensible limits, and consider different pricing models as workloads grow.
Third, connect spending to business results. Measure whether AI reduces costs, improves output, increases revenue, or delivers another clear benefit. More usage alone is not proof of success.
Esker is a useful example of how one company is responding to the challenge. Its approach will not fit every business, but the underlying issue applies to any company whose employees increasingly depend on AI to get work done.
The cost of a workforce is changing. Finance needs to keep up.
The question for CFOs is no longer just how much their employees cost. It is how much it costs to get the work done, including the AI behind it, and whether that investment is paying off.
Sources