Here's the lowdown
A survey of 486 accounting and bookkeeping professionals found 95% had adopted AI in some form, but only 11% used it across the firm.
Small practices are using spending caps and approval gates so a useful experiment cannot quietly become an open-ended monthly bill.
The cheapest workflow is often the narrowest one: filter the data first, give each agent one job and reserve the strongest models for work that needs them.
One small firm's bill rose by $1,000
A nine-person accounting firm spent about $4,000 on AI last year. This year, it spent about $5,000.
That is hardly a budget disaster. It is also a useful clue.
MJD Advisors uses AI heavily, including small agents built for individual tasks. Its annual cost rose 25%, but CEO Mike DeKock said the value still made the bill easy to defend.
The more important question is not whether AI costs more than it did last year. It is whether each extra dollar buys useful work.
Adoption is wide, but deep use is rare
The Financial Cents 2026 survey covered 486 accounting and bookkeeping professionals across North America.
It found that 95% were somewhere on the AI adoption curve. Only 11% said AI was running across the entire firm.
The report also found that only one in five respondents could point to measurable return on investment. Half said the technology looked promising but was hard to measure.
That gap explains why many small firms have not yet felt the kind of token bill that arrives at a large enterprise. A few people using a chatbot for a few tasks create a very different cost base from a team running agents through full workflows all day.
But the bill can change quickly once usage spreads.
AI has a meter running behind it
Traditional software often charges one predictable fee per user. AI can work more like a utility.
Every prompt, document and answer uses tokens, which are small pieces of text processed by a model. A short question uses relatively little. An agent that reads files, plans steps, calls tools, checks its work and rewrites an answer can use far more.
The price of one token may fall while the total bill rises. That happens when more employees use more tools and when each workflow asks the model to do more.
Small firms can also pay indirectly. A familiar accounting product may add AI features and raise its subscription price to cover the computing cost.
This makes AI spending harder to see. The cost can appear in direct model bills, software subscriptions, development work and client-specific tools at the same time.
Caps turn surprise bills into decisions
High Rock Accounting uses limits inside its AI systems. Employees cannot spend beyond their allotment without approval.
The approval test is practical. If a use case can save 10 hours a week, management may accept an extra $200 or $500. If the case is vague, more capacity is harder to justify.
Cherry Hill Advisory uses a similar approach. The firm watches usage, sets spending caps and asks employees to connect any request for more capacity to client or team value.
The control can interrupt a workflow. Someone may have to wait for approval or finish a task the old way. That friction is intentional because it forces the cost decision into the open.
The alternative is letting every useful experiment become a permanent expense.
Smaller agents can be cheaper and more accurate
MJD Advisors does not ask one giant agent to handle an entire process. It gives narrow jobs to narrow agents.
One agent may draft a specific paragraph in a report. Another may review a meeting and build a task list.
The same idea applies to data. If a folder contains 100 documents and 40 are irrelevant, the firm filters those out before the model begins. The model reads 60 files instead of 100.
That can improve accuracy because the model has less noise to sort through. It also cuts the amount of text the firm pays the model to process.
Simple design is doing two jobs at once.
Route the work before buying more power
Cherry Hill Advisory also controls the architecture around the model.
First, it prepares information so the model receives only what it needs. Second, it maps the steps and tools required for the job. Third, it routes each task to the right model.
A lower-cost model may be enough to classify documents or extract standard fields. A harder judgment call may deserve a stronger model.
Using the most expensive model for every task is like assigning a partner to every data-entry job. It may work, but the economics will not.
Training matters too. Employees who understand prompt size, context, model choice and workflow design are less likely to burn capacity on repeated or unfocused requests.
Sometimes the human process still wins
High Rock reviewed a public company's workflows to find candidates for agentic automation.
About half the time, fitting a process to AI and paying agents to run it was more expensive than keeping one person on the job.
That does not make the analysis a failure. It is the answer the analysis was supposed to find.
The comparison should include setup, testing, monitoring, subscriptions and ongoing usage. A task that looks repetitive may still be cheaper with a person if the volume is low or the exceptions are messy.
Firms also need a policy for client costs. High Rock tells clients when AI-heavy accounts-payable or accounts-receivable workflows will increase their software bill. Cherry Hill absorbs its technology costs inside a fixed client price.
Either model can work. Surprise fees do not.
Every AI budget needs an owner
AI prices may fall as models become more efficient. Total spending can still rise as firms build more workflows and put more people on them.
That leaves small practices with a familiar management job: set a budget, name an owner, measure the result and stop work that does not earn its keep.
The best cost control starts before the next token is used: What valuable outcome is this workflow buying?
