跳到正文
原文
Accounting Today(美国会计行业媒体)· Chris Gaetano·· 3 天前精选站内评分58

技术导向型小型会计师事务所AI成本上升,支出封顶与模型路由成为控本手段

Tokenomics: Tech-forward small firms see higher AI bills but controls keep costs down

导读

技术导向型小型会计师事务所正面临更高的AI账单,主要表现为常规软件订阅涨价,支出封顶、超额审批、模型路由和培训可帮助控本。Financial Cents对486名主要来自30人以下事务所的会计与簿记专业人士的调查显示,95%已以某种形式采用AI,但仅11%将其嵌入大部分日常工作;CPA Trendlines调查中,63%至少使用AI,仅3%称AI已覆盖大部分日常工作。

推荐理由

文章以小型会计师事务所的AI使用数据为线索,归纳支出封顶、ROI审批、模型路由和培训等可迁移的控本方法。

正文
AI blocks
Wooden blocks showing human mind and larger AI symbol, representing changing relationship, growing influence, support or reliance between human thinking and artificial intelligence

WEERAPAT W./patpitchaya - stock.adobe.com

  • Key Insight: Find out how small accounting firms are regulating employee access to expensive AI tools.
  • Expert Quote: "smaller firms will experience increased monthly subscription costs on their regular software that now has AI features" — Rebecca Driscoll, Founder, AI Lab for Accountants

AI costs have spiked throughout the economy, and accounting firms are no exception. While this issue has been especially acute for large firms with significant AI investments, leading to new governance and oversight features to control costs, some tech-forward small firms are also feeling the pinch. 

Processing Content

While the vast majority of firms have used AI in the past, not as many would consider themselves fully fluent with the technology. A report from Financial Cents, which polled 486 accounting and bookkeeping professionals, mostly from firms under 30 people, found that while 95% of such firms have adopted AI in some form or another, only 11% would say they have embedded it in daily workflows across much of the firm. The majority, 67%, have either parts of the team using AI regularly for defined tasks or a few people who use AI informally for occasional tasks. Another survey from CPA Trendlines found similar data: While most firms use AI in some form (63%), the largest group (37%) uses it for only a few tasks and just 3% say it is built into most of their daily work. 

Rebecca Driscoll, founder of the AI Lab for Accountants, a practitioner community for sharing insights into the technology's use, said smaller firms are still very early in their AI journey, and while vendors have been adding many AI features to their products, most of their experience with it has been with the big public models versus complex specialist platforms. 

"The really AI-heavy firms are more focused on the cost of Claude, ChatGPT, etc., for their firms, than traditional accounting firm software AI usage. I don't even think [many] firms are getting savvy about best practices to control costs quite yet. It is all still new and a lot of them are still hung up on the security concerns," she said in an email. 

Still, for those firms that have already invested heavily in AI, as well as those firms who aspire towards doing so, costs have become much more of a factor than before. Part of this can be attributed to the shift this year from subscription-based to consumption-based token pricing by many AI model providers. 

But the cost of an individual token is only part of the story. While the cost of tokens has fallen dramatically, the number of tokens an average person consumes has increased by many orders of magnitude more, especially when taking agentic capacities into account. Compared to asking an AI model what to make for breakfast, performing a full end-to-end workflow with agentic AI requires much more computation, which, in turn, leads to higher token costs. Mike Levy, managing principal of Cherry Hill Advisory in New Jersey, which makes extensive use of AI agents, has needed to contend with this at his firm, so he prioritizes things that deliver actual concrete ROI. 

"It's more workflows," he said in an email. "We've enabled a lot more around our AI-enabled team and the services we deliver to clients. We're very ROI-driven. If a workflow doesn't drive value for our customers or our team, we don't do it."

Beyond the cost of individual tasks themselves, there's also more people using AI than before, which creates more costs regardless of pricing model. This is one of the reasons Arizona's High Rock Accounting has seen higher costs compared to last year, according to CEO Liz Mason, who speaks frequently on the intersection of AI and accounting. 

"It's more people using more tools in general, and that makes sense," she said in an interview. "AI is a commodity, like a utility, it is not software the way we traditionally think of software and the cost of it is incredibly expensive. It's not a cheap thing to do in general." 

Driscoll noted that as software providers add more AI features, which themselves require ongoing costs to maintain, they will also raise prices on subscriptions. While smaller tech-forward firms aren't necessarily suffering from token costs, she said they are still facing higher AI costs in general. 

"I think larger firms will face direct token costs on the enterprise level AI tools and smaller firms will experience increased monthly subscription costs on their regular software that now has AI features," she said. 

Mason has seen this happening as well, though she herself hasn't experienced it, as she maintains very good relationships with her vendors (as well as participates in a lot of free beta programs.) But she knows she won't be able to avoid the issue forever as AI costs mount for vendors. Still, she feels fortunate that she can see this coming down the pipe and prepare accordingly. 

"As the tools start realizing the cost of throwing this inside their software, they're starting to charge for some of that. Luckily, we have pretty good relationships with our vendors, so most of it we know is coming, but it hasn't quite hit yet," she said. 

Controls and guardrails

Part of keeping AI costs down for Mason is monitoring and capping use, saying there are limits baked into everything that her team does. This helps keep costs predictable as people cannot spend more than they've been allotted without approval. What gets someone that approval is a clear business case and return on investment. If someone pitches a use case that will save 10 hours a week, even if it costs $200 or $500 more than normal, management will usually approve it. While they have seen a spike in costs, it is not unexpected "because of the way we've locked everything down," said Mason. 

"What we're doing is saying, Ok, our house can only run this much power a month, and once you max out that power, then you're in the dark. No TV, no computers. You're off. There is no more power for you. You have to come get approval for more, and that's how we've regulated costs," she said. 

She noted that this method does sometimes interrupt workflows, meaning people need to "figure out how to do it the old fashioned way again or they have to get approved for more," but this is more an occasional stumbling block than a persistent issue. 

Levy, from Cherry Hill Advisory, also monitors and uses spending caps so they don't get surprise bills. Similar to Mason, if someone wants to go past this typical allotment, they need approval and authorization, and that approval is contingent on ROI conversion about how that spend delivers value to their clients. Beyond that, he pointed to education and training as another component, saying that people learn to use the technology judiciously, with as little waste as possible. 

"We've gotten a lot more efficient in how we use AI. We're on a mix of consumption-based and fixed pricing. So much of using AI thoughtfully and cost-effectively comes down to education and training, and that's a big focus for our team," he said. 

His firm also takes a variety of technical measures focused on the architecture of its AI software. This includes pre-processing, which means preparing information before it reaches an AI model, so the model receives only the data it actually needs to accomplish a particular task; optimized workflow routes, which means designing a system that automatically determines which steps are necessary, in what order they should occur, and which tools should perform them; and mode routing, the practice of automatically selecting the appropriate model for each task, so someone would use a smaller cheaper model for a lower-stakes task and reserve big frontier model use for complex tasks that actually demand their power. 

"That lets us automate much of the optimization so it's seamless to our team and our clients," he said. 

Mike DeKock, CEO of MJD Advisors in Iowa, is an enthusiastic vibe coder who has also discovered a number of measures to use AI in a cost-effective way to the point where, when asked about the degree to which his own AI spend has dramatically spiked, he said it hasn't.

"It's not a problem I'm having. I would pay four times what I pay for AI because of the amount of value I get. I looked at the numbers. We paid $4,000 last year. We paid $5,000 this year," he said. 

When asked what his secret was, he admitted he wasn't sure whether there was one. Part of it might be the size of his firm. There are only nine people, so absolute usage might be lower than in other practices. But another factor might be that, as much as he uses AI and codes agents, he is intentional when it comes to where and how he uses them. His overall strategy, he said, is to do very simple workflows with very simple agents that do very simple tasks. 

"One specific paragraph in a report has an agent, and one specific meeting reviewer that's looking and building out tasks, that's another agent," he said in an interview. "I found that both increase the accuracy, the biggest thing that I'm after. But I think it also impacts the token cost because I don't have to have these super long context filling threads that take up a bunch of tokens."

He noted that he's learned that AI does best when it doesn't need to sort through too many things. For example, if there is a folder with 100 documents, he introduces a filter that pulls out 40 that are not relevant. so the model only looks at the 60 they actually need. 

"I'm not just dumping things into AI," he said. 

Mason noted that efficient and cost-effective AI is not only good from a budget perspective but an ethical and environmental one as well. While many firms think of this only in terms of cost, she said the question of cost cannot be decoupled from data rights and environmental conservation, as they pertain to how people develop and use AI models in the first place. Considering the energy consumption alone, someone using a frontier model to make an alien face gif to react to a message is not just a matter of cost but a matter of ethics. 

"I live in Arizona," she said. "In the middle of the summer, would I set my AC on 50 degrees and open all the windows and the doors? Absolutely not. That's a terrible decision from a financial perspective, but it's also a terrible decision from an environmental perspective. … I think that that's kind of the same mentality that we're at right now — people are starting to develop frameworks for judging. Is that abusive? Is that ethical? Is that environmentally sound? Is that in line with my core belief systems?"

Levy added that cost efficiency is also valuable from a client service perspective. The price increases may be unfortunate, but he said they have helped clarify what truly is important to the bottom line, tying everything back to ROI. 

"We take an outcome-driven approach rather than thinking about how many hours an engagement will take. I appreciate how that has sharpened our focus on client service and value. If we focus on value and outcomes, we'll do the right thing," he said. 

The client aspect

The lessons learned in their own practices are also brought to their clients. Much as Mason has learned to be judicious with AI, she teaches her clients the same thing. And in this age of rising costs, it also encompasses questions of whether AI is even the right fit for their particular need. For example, one client she worked with, a public company, wanted to know which of their workflows they could automate with agentic AI, which involved documenting the workflows themselves to analyze their suitability. In the course of doing this, she said, they found about half the time it was actually more expensive to go through the whole process of fitting the workflow to AI and having agents run it than it was to just keep paying one person. 

"We did a bit of an ROI analysis on that — and those are shifting because we're able to utilize better, more efficient models for bigger types of data analysis — but the way that it trained my brain to think about this is: does it make sense to implement the AI right here if we can continue to employ a human?" Mason said. 

Instead of redoing the entire workflow, they focused more on upgrading the tech stack with more AI-native solutions with MCP support that she said are more future-proof than closed off manual general ledgers. 

Levy, similarly, advises his clients on not just using AI but doing so without breaking the bank. A lot of conversations, he said, is how they can operationalize the technology in a way that is both effective and efficient. 

"This is a big part of where we drive value. We make sure enablement is done thoughtfully, with a training and education component, and that people focus on strategy and architecture so they measure twice and cut once. We also stay agile enough that change can happen effectively," he said. 

The two diverge on the matter of passing on tech costs. Mason said on the tech stack side of things clients are usually paying for their own general ledger, payroll and AP accounts, which go on their bills. Similarly, she said that costs of certain AI-intensive workflows may also be passed on to clients, though she said they are usually OK with this due to the increased performance. 

"The AI structure, particularly around some of the AP workflows and AR workflows, we tell them if we turn this on it's going to cost you more, [but] they are all for it from the timeframe perspective, the processing perspective, [because of] the information they get out of it. And so it hasn't changed our bills per se, it's changed theirs. But we are managing it, so we do have to be upfront about that," she said. 

Levy, on the other hand, said that since everything they do is AI-enabled, they choose not to charge fees or pass on costs. 

"We never charge a tech fee. Everything we do is an AI-enabled service. We absorb the costs ourselves, so clients know what to expect and there are never surprises," he said. 

Up or down?

The AI landscape is often in flux, so it can be difficult to know what to expect next week, let alone next year. Still, when looking at long term trends, Levy thinks that while AI will always be a major expense for firms, as the technology improves, costs will start going down. 

"I do think AI will be a major cost, and the efficiency it creates is very significant. The heaviest spend is in development and coding, but there are efficiency moves there that can materially drive down the cost," he said. 

Mason was less sure. She said AI acts more like a utility than traditional software, so she believes its price will fluctuate in a similar manner. She agreed with Levy that innovation over time will lead to more efficiency, saying she has seen a similar progression with power grids, but wasn't certain this will mean lower prices. While AI costs are rising, it remains heavily subsidized by providers, just less so than before. Even if they're charging more than ever before, they're still running at a loss. While she said prices could go down, it's not inevitable. 

"It will become more efficient by the function of innovation. This has always been the trend and will continue to be the trend. What I fear is that we are sitting in a late-stage capitalistic society where we can take advantage of the naivete of companies across the board, and there will be increased abuse on pricing models and funding in a time when maybe it is more efficient than they're letting on," she said. "But I do think it'll come to a moment where we'll have public regulation." 

This was the fourth part in a five-part series on AI tokenomics. You can find Part 1, Part 2 and Part 3 at these links. Find out next week how the vendors who sell software solutions to accountants are managing their own AI costs, and how this affects their firms, in the final part of our series.

Introductory bullet points created by AI with editorial review.

Gaetano-Chris

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

审计热点 是聚合摘要和阅读索引,原文版权归各来源所有,引用请以官方原文为准。如果你是来源方,希望更正、下架或调整展示方式,可以通过反馈页联系我们。

来自微信公众号的内容,本站仅呈现摘要与原文链接,不转载文章全文;文章权利归属原作者及原公众号

关于使用条款隐私信源