Here's the lowdown
A $20 monthly AI seat can still cost a heavy user $500 or more after a firm moves to enterprise and consumption pricing.
One accounting firm used what looked like a year's allowance of 3 million tasks in only two and a half months.
Cheaper tokens do not guarantee a smaller bill because firms are adding users, tools, data infrastructure, and longer multi-step workflows.
The old software budget was easy to understand. Count the seats, multiply by the monthly price, and move on.
AI is turning that tidy little spreadsheet into a guessing game.
At LBMC, a standard Claude seat had cost about $20 a month. The firm's move to an enterprise version kept that base price and added usage charges. A light user might stay near $20. A heavy user could reach $500 or more.
That is not a small price increase. It is a completely different way to buy software.
The seat is only the cover charge
Traditional subscriptions give a firm a mostly fixed bill. Consumption pricing adds a meter.
Every prompt is split into tokens, the small pieces of text an AI model reads and writes. Longer inputs, bigger documents, repeated revisions, and more powerful models can all use more tokens.
The cost gets harder to see when an AI product performs several steps behind one button. It may search a knowledge base, plan the job, call another tool, check the answer, and try again before the user sees a result.
EY's research on agentic AI gives a useful example. A customer-service chat that once cost four cents can become a $1.20 job after adding retrieval, planning, and subagents. That is 30 times the cost for an interaction that may look almost identical to the employee.
The bill can also hide work outside the model. Firms may pay for data preparation, evaluations, human review, cloud storage, and the connections between the AI tool and tax or audit systems.
Cheaper tokens can still produce a bigger bill
Here is the strange part: the price of individual tokens has fallen sharply.
The problem is volume. Firms are giving AI to more people, testing more products, and asking the software to complete whole workflows instead of one narrow task.
Extracting a number from a document is a small job. Building a workpaper, passing it to a reviewer, sending the result into another system, and preparing a client-ready file is a chain of jobs. Each link can call the model again.
The market is growing just as fast. Gartner forecasts that worldwide AI spending will reach $2.7 trillion in 2026, up 49.5% from 2025.
That spending is not proof that every AI project creates value. It shows how quickly experiments, infrastructure, and embedded features are becoming normal business expenses.
A generous allowance may not be generous
Wiss & Co. received 3 million tasks per month from one vendor. On paper, that sounded large.
The firm hit its annual limit in two and a half months.
That gap shows why buyers need more than a number. What counts as one task? Can a single workflow create dozens of tasks? Does the allowance reset monthly, or does heavy early use consume an annual pool? Can the firm see usage by person, client, model, and workflow?
If a vendor cannot explain the meter, the firm cannot build a reliable budget around it.
Model choice can move the bill 50 times
The fanciest model is not always the right model.
One technology leader tested the same email prompt on two models. The answers were nearly identical, but one cost 50 times more.
That is where governance becomes a money issue. Staff should not need to understand every model price, but the firm's system can route routine jobs to a cheaper model and reserve expensive models for work that needs them.
Think of it like staffing. A partner should not do every data-entry task simply because the partner can. The same rule applies to AI models.
Budget the outcome, not the token
A low token price can be comforting and useless at the same time. Firms care about the total cost of finishing a return, building a workpaper, drafting a memo, or reviewing a client file.
Start with those outcomes. For each important workflow, track the model cost, connected software, cloud expense, review time, failure rate, and hours saved. Then compare that total with the value of the completed work.
The questions become practical fast:
Which users or workflows produce the largest bills? Which model creates an acceptable answer at the lowest cost? Which tool duplicates something the firm already owns? Where does human review erase the promised time savings?
AI vendors may eventually make consumption pricing calmer and easier to predict. Firms cannot wait for that day to learn what they are buying.
The $500 user is not automatically a problem. A $500 bill that produces $5,000 of useful capacity may be a bargain. The real problem is paying the bill without knowing what work created it or what the firm received in return.
For more plain-English accounting news, follow Ledger Lowdown.
