Here's the lowdown
Accounting firms are moving away from AI usage leaderboards and asking employees to show what each tool changed in their work.
One firm uses a $25 monthly soft limit and a $500 hard limit, with managers approving higher spending as the user proves value.
Model choice has become a staff skill because one model can cost 50 times more than another for nearly the same task.
The first phase of workplace AI rewarded motion. More prompts were good. More users were better. Some companies even built leaderboards.
That phase is ending inside accounting firms.
Leaders still want people to use AI. They just want an answer to a harder question: What did the spending produce?
A cap is not the strategy
The easiest cost control is a hard limit. It is also the easiest one to misuse.
Set the limit too high and a careless workflow can burn money all day. Set it too low and an employee with a valuable use case gives up before the work is finished.
LBMC's AI team uses a ladder instead. A user has a $25 monthly soft limit and a $500 hard limit. As spending approaches $25, the user can approve an increase to $50. A manager can then allow another step to $75.
The point is not to punish usage. The pauses create a moment to ask what the model is doing and whether it should continue.
That matters because an employee can start a long prompt, walk away, and return to a large bill with nothing useful to show for it. Another firm leader reported seeing one person spend $600 in a single day by repeatedly revising the same deliverable.
Model choice is now a money skill
When every model came with the same flat subscription, choosing the most advanced option felt harmless.
Consumption pricing changed that math. Firms now pay for the computing work behind each request, and more advanced models often use more of it.
One technology leader tested the same email task on two models. The answers were almost identical, but one cost 50 times more.
That does not mean firms should ban the expensive model. It means staff need to match the tool to the job.
A simple request, such as finding last month's client revenue, may work on a cheaper model. A complicated analysis with several documents and judgment calls may justify a stronger one.
Some firms are testing software routers that choose the model automatically. Those systems are not always reliable yet, so training remains part of the control.
Usage data needs context
Firms can now see who uses which AI tools, how much they spend, and sometimes what types of prompts they run.
That dashboard can identify two very different people who look the same on a bill.
One may be a power user who automates half a task and creates capacity for more client work. Another may be repeating prompts without using project folders, context windows, or other features that make the tool more effective.
Wiss & Co. reviews heavy usage with weak results and asks the employee to explain the process. The firm may follow up with training instead of simply cutting access.
This is the governance shift in one sentence: measure effectiveness, not enthusiasm.
The return is what happens next
Time saved is not automatically value created.
If AI gives an accountant one extra hour, the firm still needs a plan for that hour. The employee could spend more time with a client, prepare another pitch, work on a second engagement, or study a client's unusual problem.
Or the hour could vanish into the day.
That is why finance leaders are looking beyond the AI invoice. A $500 model bill can make sense if it saves more than $500 of staff time and that time becomes useful work.
The same logic applies to firm software. Leaders are reviewing point solutions, pricing escalators, agent-level charges, and added fees for access to the firm's own data. A small AI feature can make a familiar contract much more expensive.
Training keeps the controls from backfiring
Firms do not want a cost program that teaches employees to avoid AI.
The better approach pairs limits with education. Staff learn which model fits which task, how to write efficient prompts, when to stop a bad run, and how to show the result.
Different users can also receive different budgets. A development team building complex tools may need more capacity than someone using AI for routine drafting. A proven power user can earn a larger allowance than a new user who is still learning the basics.
The goal is not fewer tokens at any price. It is fewer wasted tokens and more useful work.
Accounting firms have moved from asking, "Are our people using AI?" to asking, "What changed because they used it?" The firms that can answer the second question will have a much easier time defending the bill.
For more plain-English accounting news, follow Ledger Lowdown.
