会计师事务所如何通过预算、监控与培训控制AI令牌支出
Tokenomics: How to control your firm's AI spending
会计师事务所正通过令牌预算、模型选择、软件监控和员工培训控制AI成本,并将考核重点从使用次数转向可衡量产出与投资回报。文中介绍按岗位和技能分配额度、用软件跟踪工具使用及提示词,并对高使用低产出人员开展复盘;LBMC AI业务负责人称,其部门设有每人每月$25软上限和$500硬上限。相关管理者认为,额度限制需与培训和监控结合,并将节省的时间重新投入客户互动、关系建设和项目执行,以验证AI投入价值。
文中多家事务所的做法覆盖令牌额度、模型选择、使用监控和人员培训,为兼顾AI采用与成本控制提供了可迁移的治理路径。

Andrii Yalanskyi - stock.adobe.com
- Key Insight: Find out how accounting firms are tracking and restricting employee AI token usage.
- What's at Stake: Firms risk losing employee AI adoption if new cost-control measures are too restrictive.
Over the past year, AI spending at accounting firms has soared by many orders of magnitude, which in turn has led leaders to quickly develop new governance models, new training methods, new budgeting considerations and new technical guardrails to ensure cost-effectiveness and optimal use.
Processing Content
A big part of this increase can be attributed to the shift from subscription to consumption-based pricing by many model providers, but other major factors include rising numbers of people using AI for increasingly complex tasks, as well as the overall cost of software increasing with vendors raising prices as they add new AI features.
However, despite these price shocks, firm leaders are not at all pulling back from their AI ambitions, they're actually spending more than ever because they remain confident that the benefits of the technology outweigh even these increased costs. The last thing they want is for people to stop using AI. They just want to make sure they're using it effectively. It's not so much what they're paying, but what it is they're paying for, according to Marc Staut, shareholder and chief innovation and technology officer for accounting firm consultancy Boomer.
"I'm not seeing a meaningful slowdown. The refinement is in sequence and in what firms expect to pay for. … So the long-term vision hasn't shrunk. Firms are asking how to get value from what they already pay for before they buy the next thing, and they want a measured result before the next renewal," he said in an email.
This means that leaders must walk a delicate balance between controlling costs and encouraging use. Firms want people to use AI efficiently, and so have implemented controls, but they also don't want to be so restrictive that people stop using AI at all.
"We put in place the guardrails to limit the excess binge component but in general we still think there's tremendous value to using AI in your daily workflow so we don't actually want to discourage usage even though we're getting charged for that usage," said Philip Comerford, CFO of top 25 firm PKF O'Connor Davies, in an interview.
This is in marked contrast to the rhetoric up until now, where people were encouraged to use AI for just about anything and everything they could think of, whether that be writing emails or composing memos or even planning a dinner. Firms ideally wanted as many people as possible using AI, and so using it for just about any reason was viewed favorably. Some companies even evaluated employees based on simple sheer AI use, a few even had leaderboards to that effect. Charlie Apigan, AI practice leader for top 50 firm LBMC, said those days are over now.
"For example, if one of my team members right now puts into Claude he wants to make dinner tonight and he needs to know a recipe, guess what? I just got charged," he said in an interview.
Firms are now heavily focusing on not using fewer tokens but making sure the tokens that are used are put towards clear goals with real business purpose. While there are technical measures in place to prevent overspending, Hrishikesh Pippadipally, chief technology officer for top 100 firm Wiss and Co., noted that most cost control measures are behavioral and governance-related at their root. Measurable outcomes are now the order of the day.
"[It used to be] more usage is good, more people using it is good. That was a general statement back then because we wanted people to use it, adopt it, experiment with it. Now it's more like 'what is the measurable outcome?' Yeah, you're using it now, but how is it helping you in what task? Is it helping you automate 50% of the task and saving X number of hours? Are you able to do more engagements now? All those are conversations. No more 'oh, you're using it a lot, that's great.' It's 'How effectively are you using it? How is it changing your work?'" he said in an interview.
Comerford raised a similar point in that they want to make sure AI is used strategically. They want people not only to save time using it, but to find other things to do with the time they saved to make the investment worthwhile. It is of less concern what an accountant spent using AI than what they did with the results.
"What it really comes down to is, with these time savings, what is that individual doing? Is that individual who now has another hour in the day going to spend additional time with clients? Are they talking to referral sources? Are they building their own network? Are they prepping for another pitch? Are they spending additional time on the actual engagement? Are they understanding the individual nuances of a client that historically they might not have had the time to do? So it's all about the redeployment of time saved. In our mind, the value of that time saved is greater than discouraging usage on mundane tasks," he said.
Apigan noted that what they're really after is return on investment. He noted that the firm understood that building out a strong AI capacity costs money, but so long as it provides real results that grow the bottom line in other places, they're perfectly fine with spending more. If someone spent $500 on AI, but used it to save more than $500 worth of time, he said, the usage paid for itself.
"That's what we're trying to do today. Nobody's complaining about that cost. We accepted it. We knew it was going to be a cost. We knew it was going to be a lot. We didn't know how quickly we would get to a certain threshold. But we're spending," he said.
Richard Kopelman, CEO of top 25 firm Aprio, said in an email that this sort of efficiency is not just about the immediate impact but also the long term skill and relationship development that serves the firm in the long term.
"When technology can effectively handle more routine work, our people gain capacity to solve problems, spend more time with clients, deepen relationships, and bring more of their experience and judgment to their work. They also gain opportunities to develop new skills and grow their careers as the profession evolves," he said in an email.
Cheryl M. Burke, the project management officer for PKF O'Connor Davies and a specialist in operations management, noted that her own firm has worked these principles into their training. People now are taught to be efficient with their prompts and judicious with what model they decide to use in the first place, as more advanced ones consume more tokens.
"If you're going to ask a simple prompt you can use a simple instance of ChatGPT. You don't need so high [a model] for everything, you need to think about whether this is a task it is good for. We're accountants, so we know the costs and the money, so they're starting to use some of that discretion and saying 'I'm going to use Instant for this, but when I have something more complex, [I'll use a more advanced model.]" she said in an interview.
The specifics of making these evaluations depend on diligent monitoring of AI use by staff. Leaders said they use software to track who is using which AI tools in the firm, how much they're using them, and in certain cases even what specific prompts they're inputting. Pippadipally noted that his own firm takes a data driven analytical approach to overseeing AI spend, which helps see who are the power users and who is still struggling.
"We have lots of smart analytics where I can get to know who the users who are just spinning their wheels, not getting the effectiveness out of their tools, and not using the projects or context windows or skill tools appropriately or are just going randomly. So there is an effectiveness piece we can measure now, so it's getting better," he said.
People who are heavy users without much impact, he said, get an email asking them to explain their processes and how they use the models, with the possibility of a follow-up meeting after. They're also provided with a prompt to input into their own AI that will analyze and summarize their AI use for their own information.
Another common practice among firm leaders was user segregation. Users are given a certain token budget that can vary based upon their needs and skill. Extremely prolific AI users who make full use of their tools' capacities to great effect will get more tokens in their rations versus those who are not as skilled in using it, or don't use it as much. Apigan said people in his own department have a $25 per month soft limit and a $500 hard limit.
"What that means is if you get close to $25, you personally approve yourself to go to $50, and when you get close to $50, I can let you go to $75. And the reason for this is that you could do a prompt and walk away from your desk: what if that costs $100? And it [produced] nothing. So we want to stop it at 25 bucks," he said.
Pippadipally also stressed the importance of allocations, whether that be on an individual level or tiers based on whether they're an advanced programmer or a more mundane day-to-day user, and to constantly monitor how those allocations are consumed. Leaders, he said, need to watch how people are using AI tools, where they're seeing value, and how managers can support them in becoming more effective users. Overall he described it as a governance model based on understanding your staff.
"Because otherwise I have had developers come to me saying they used all their capacity over two days or the weekend. So you've got to be careful. You need to segregate user groups and educate them because there's a broad context for people there… Whether staff or partner level, everyone has different ways of doing things, plus the development team is completely separate," he said.
Beyond tokens, leaders also limit model usage based upon need. There was a clear consensus among leaders that most of the time, for most tasks, you do not need to use a highly advanced frontier model. If all someone wants to know is, for example, what the client's revenue was last month, a much simpler version is usually just fine. Apigan noted that people can certainly request to use more advanced models, "but you're probably going to have to show you're doing some pretty complex tasks to need that."
This reflects a point Staut made about how model selection needs to become a staff skill. When costs were the same all around, it didn't much matter what model people used. Now that more providers are charging for compute, however, knowing the right model for the right job has become important to controlling AI costs.
"Using the right model for the task matters when the same task can cost 50x more on one model than another. More advanced firms are looking at routers that pick the model automatically, but the routers aren't always good enough yet, so for now it comes down to training. Caps alone don't change behavior either. One CIO in our community caught a single user running up $600 in AI costs in one day by iterating on the same deliverable. They had no idea what each round cost. The firms getting this right pair caps with training and monitoring," he said.
Past AI tools themselves, firms are also controlling costs through reevaluating their current tech stack. Apigan noted that there's a lot of point solutions that aren't necessary anymore because general AI can do the same things already or, failing that, can be used to create a custom application that does. For example, rather than buy a point solution to monitor AI spend, they vibe coded their own application which they now use to oversee model usage across the firm.
"You know what AI can do right now. So do I need to spend that huge subscription for that one small niche thing that our accounting firm does, or do I create a Claude skill that does that for you? And so I think that's where the reevaluation in the early in the early times is going to need to be done because most companies, they don't have a big, I'm assuming, tech department. So for them to go build their own, have their own servers or do other things, they're not going to do that," he said, adding they're not at the point yet where they're going to be switching around solutions or redistributing costs, but said they will in about a year.
It has also meant being more judicious with vendors, with leaders now thinking more carefully about whether they want to buy that latest solution.
"Where firms are going back is the vendor contracts: reading the price escalation clauses, asking for per-unit or agent-level pricing so they can tie cost to value, pushing back on pricing tied to the firm's revenue when the value isn't clear, and noticing which vendors now charge more for full API or MCP access to the firm's own data. A few firms have also told me the total cost of ownership on newer platforms came in higher than they expected once everything was counted," said Staut.
All of these cost-cutting measures are especially important in light of the fact that leaders, despite rising prices, are generally not passing on the expense to their clients. If an engagement requires a lot of tokens burnt, perhaps for complex financial modeling, the firm will choose to eat the cost before upping their fees. While firms are increasingly adopting technology fees, these are separate from AI spending.
Pippadipally said that Wiss does not even charge a tech fee. There had been heavy debate within the firm as to whether to do so, but ultimately they decided the costs were too difficult to disaggregate to meaningfully charge one.
"We had debated a lot on adding a tech fee … But the clients are not ready for it. We haven't figured out a clean way where tech costs are directly there because it's everywhere. Like, it's there in our ERP systems, our CRM systems, our service systems, our billing and AR systems, which is more horizontal cost versus workflow AI cost," he said.
Comerford similarly said they do not pass on their AI costs, feeling there is little need to do so as they're able to actually be more productive with the technology anyway and therefore derive more value from what they spend.
"In our minds, with AI, we're able to deliver additional value to the customer through better insights, more robust testing, and overall a better output to our end clients. And so we haven't seen an impact to date around pricing," he said.
Overall, it would seem that while firms are seeing higher AI bills this year, leaders appear to view this as simply the price of developing an AI-first organization that will, ultimately, bring far more value to the firm than whatever they spent getting there.
"Looking ahead, I believe leadership in AI will be defined by how thoughtfully organizations apply it. The greatest value will come from combining the capabilities of technology with the experience, judgment, curiosity, responsibility, and trusted relationships that people bring. That is where we see the real promise of AI," said Kopelman.
But how is this affecting the client relationship? How have price spikes affected what advisory firms tell their clients? We'll find out more in Part 3.
Introductory bullet points created by AI with editorial review.
Technology Editor
Chris Gaetano is the technology editor for Accounting Today. He brings with him more than a decade of experience covering the accounting profession... Read full bio
来源:Accounting Today(美国会计行业媒体) · accountingtoday.com