Here's the lowdown
Pigment found that 83% of 2,000 finance leaders said consumption-based AI costs pushed spending above expectations.
The typical overrun was 11% to 25%, while 17% of respondents said their organizations were more than 50% over budget.
CPA firms are helping clients set token limits, monitor usage, choose cheaper workflows, and decide when a human is still the better answer.
A $20 software seat feels easy to budget. A meter that runs every time an employee asks an AI model to read, reason, or retry does not.
That pricing change is creating a new job for accounting firms. Clients do not just need help buying AI. They need help understanding why the bill moved and whether the work was worth it.
The problem is already widespread. Pigment's Q3 CFO Index surveyed 2,000 CFOs and finance executives across four countries. It found that 83% said consumption-based AI costs had exceeded expectations. Only 13% stayed within budget.
The typical overrun was 11% to 25%. More than one-third went over by more than 25%, and 17% were more than 50% over.
That is not a rounding error. It is a forecasting problem, an operating problem, and increasingly an advisory opportunity.
The price moves with the work
Traditional software pricing made the math simple. Multiply the number of users by the monthly seat price and add a little room for growth.
AI can work differently. Many enterprise products charge for consumption. The bill may depend on how much text a model reads and produces, which model performs the task, how many files it reviews, and how many times an automated workflow calls another tool.
One request can also trigger several hidden steps. An AI agent might search a database, read a contract, compare records, draft an answer, check the answer, and try again. Each step can use more computing power.
That means the same employee can create a small bill one month and a large one the next. A fixed budget built around licenses will miss the meter running underneath them.
Pigment found that 40% of respondents blamed unpredictable usage for making AI costs hard to forecast. The hard part is not merely finding the total. It is explaining what caused it.
Firms are selling what they learned at home
Accounting firms have already faced this problem inside their own businesses. They have tested usage limits, token monitoring, model training, and reviews of high-cost workflows.
Now they are bringing those controls to clients.
Wiss & Co. trains clients to separate user groups, monitor tokens, and teach employees how to work more efficiently. The firm's technology leader said governance and cost monitoring need to be built from the beginning because adding them later is difficult.
LBMC is taking a tool-neutral approach. Its advice may be to use a large language model, build a smaller controlled workflow, or leave the job with a human. The answer depends on the business problem and the economics, not on which AI product has the best demo.
PKF O'Connor Davies starts AI advisory work with governance. That conversation covers which tools are allowed, how teams surface useful ideas, who owns compliance, and how tightly usage should be controlled.
Some clients want a hard token allowance for every employee. Others want more flexibility. There is no universal number because a software developer, a finance analyst, and an occasional user do different work.
Start with the business problem
A token budget can stop a surprise. It cannot prove that AI created value.
That is why Aprio starts with a more basic question: What is the client trying to accomplish?
Maybe a finance team wants to shorten monthly reporting. Maybe employees spend hours finding clauses in contracts. Maybe customer questions pile up faster than staff can answer them.
The workflow comes after the problem. That order matters because an impressive model can still be the wrong tool.
Suppose a company spends $4,000 a month on an AI workflow that saves 100 hours. The invoice alone does not settle the question. The company still needs to know whose hours were saved, what those people did next, and whether the result was accurate enough to use.
If the saved time disappears into the week, the return may be weak. If the team closes the books faster, serves more customers, or avoids new hiring, the same bill can look reasonable.
The example is hypothetical, but the test is real: Connect cost to an outcome a finance team can measure.
A practical control loop
Clients do not need a giant AI committee before they can control spending. They need a repeatable loop.
First, inventory the tools and contracts. Record seat fees, usage charges, model prices, data-storage costs, and automatic renewal terms.
Second, assign owners. Someone should approve new tools, review access, and investigate unusual usage. Finance, technology, security, and the operating team all have part of the answer.
Third, set different limits for different jobs. A flat cap can block a valuable heavy user while giving an inefficient workflow too much room.
Fourth, review cost beside output. A dashboard should show more than dollars and tokens. It should connect spending to completed work, time saved, error rates, revenue, or another useful result.
Finally, make stopping normal. A pilot that fails its value test should end. A workflow that needs a costly model should be able to explain why a cheaper one did not work.
The advisor still has to make the call
AI can collect usage data and flag a spike. It cannot fully decide whether the spending helped this particular business.
That decision needs context. A client may accept a higher bill during tax season, a system conversion, or a product launch. The same increase during routine work may deserve a closer look.
This is where an accountant's judgment becomes valuable. CPAs already connect messy operating data to budgets, controls, and decisions. Consumption-based AI adds a new meter, but the advisory skill is familiar.
The winning question is no longer, "How many people used AI?" It is, "What did the company get for the bill?"
For more plain-English accounting news, follow Ledger Lowdown.
