多家会计师事务所AI成本较去年激增,按量计费与复杂工作流推高支出
Tokenomics: Firms see surging AI costs compared to last year
会计师事务所的AI成本较去年大幅上升,模型供应商转向按量计费及企业采用更复杂工作流共同推高账单。LBMC称,单个席位由每月20美元的订阅费叠加使用量费用,重度用户可能超过500美元;PKF O'Connor Davies、Wiss and Co等事务所讨论了模型选择、用量控制与预算治理。
通过多家事务所的计费变化、用量差异和预算数据,呈现按量计费下模型选择、用量控制与预算治理的实务权衡。

tadamichi/tadamichi - stock.adobe.com
Key Insight: Discover how accounting firms are experiencing unexpected spikes in their monthly AI bills.
Expert Quote: "It would be like if Netflix charged you for binging based on how many minutes you watched," — Philip Comerford, CFO of PKF O'Connor Davies
Processing Content
AI spending at accounting firms has risen dramatically compared to last year, a shift that is primarily–but not completely–explained by the move from subscription-based to consumption-based pricing by many model providers.
These issues reflect changes within the broader economy as rising AI costs affect enterprises of all sorts, including accounting firms. As early as the beginning of this year companies were observing costs spike as a direct result of AI use, with a Deloitte report from January noting that cloud bills are rising nearly 20% year over year, driven by AI workloads, and spending has only increased since then. More recently, Gartner said that worldwide spending on AI is forecast to total $2.7 trillion in 2026, a 49.5% increase year-over-year.
We see similar growth when looking specifically at CPA firms. Inside Public Accounting, for example, found that 70% of firms report increasing their AI budgets over the past year, and a majority are actively implementing or refining AI initiatives across their organizations. Meanwhile, Wolters Kluwer found that firms are planning to spend even more in the future, as 87% said they plan to increase their investments in AI-enabled tech. Top 50 firm LBMC is just one of many firms that have been feeling the squeeze this year, with AI practice leader Charlie Apigian saying cost increases since the beginning of this year have been "substantial," with much of the spike attributed to, this year, going from the standard to enterprise version of Anthropic's Claude, which opened the firm up to consumption-based pricing on top of the regular subscription cost.
He noted that, prior to this year, they had only been spying $20 a month per seat for access to the bot; converting to the enterprise version increases the price from the same $20 a month plus additional token costs. So already he is facing the same costs he had last year, plus usage fees, which can vary dramatically based on the user. While some people might only spend within the previous $20 range, heavy users might incur much more than that, possibly $500 — or more.
"And I bet you I'm probably going to have even more than that. So you know, that's more than 5x. Gosh, it's way more. We're talking 25x at that point, going from $20 to $500. So yeah, that's pretty big," he said in an interview.
Philip Comerford, CFO of top 25 firm PKF O'Connor Davies – which has seen its own share of rising AI costs – noted that this variably has been one of the biggest changes in how to think about AI spending. When things were mainly subscription-based, costs were a lot easier to predict and control because the bill would be roughly the same every month. Now, however, it can go up or down based on how many tokens they used in a given period.
"It would be like if Netflix charged you for binging based on how many minutes you watched," he said in an interview.
Hrishikesh Pippadipally, chief technology officer for top 100 firm Wiss and Co, which has also seen AI costs climb over the past year, said another confounding factor is the subjectivity of the current market, while model providers are now charging for tokens it appears they have yet to agree on how much a token should cost, making it even more difficult to predict one's AI bill each month. This means that there is only so much they can control from their end versus the vendor, who may not necessarily understand the firm's processes.
"One of our vendors in the past year, they went from seat-based to consumption models. They gave us 3 million tasks per month on our subscription, it sounds great. And we hit our annual limit within two and a half months. So that's the maturity I'm talking about, it's still not there in the AI space for even the vendors to price it. … We did some optimizations, but it's still not enough," he said in an interview.
While he wouldn't mind a consumption-based model in and of itself, he said right now things are much too chaotic for him to feel comfortable, especially considering he cannot know the degree to which the vendor has architected their product to optimize token spend, or if they even did so at all.
"If they don't use it effectively, they might burn tokens because they're not optimizing their cost models, so there's a carrying degree of cost. That's why I'm not ready. But in the future, when it becomes very close and margin for error is small on the vendor side, then yeah,it makes sense to go to a consumption-based model," he said.
The why behind the rise
While the shift to token-based pricing from many providers has led to increased costs, the absolute price of tokens has actually gone down over the past year. In fact, enterprise data solutions provider Modern noted in its recent report on AI costs that token prices have fallen roughly 3000 times over the past three years. This past June alone, OpenAI and Anthropic have cut their prices a further 80%.
But the cost of an individual token is only part of the story. Because while the cost of tokens has fallen dramatically, the number of tokens your average person consumes has increased by many orders of magnitude more, especially when taking agentic capacities into account. A report from Big Four firm E&Y noted that not only are people using AI more, they're using it for more complex use cases which take even more compute power.
"Consider a customer service AI assistant built two years ago to answer product questions. A chat that once cost $0.04 may now involve tool retrieval, planning and subagents that turn it into a $1.20 orchestration. Additional costs — such as knowledge-base updates, agent evaluation and human collaboration design — may not appear on the model vendor's invoice but still surface to the enterprise," said the report.
Pippadipally, from Wiss, raised a similar point. Not only are they using AI more than before, they're using it for increasingly complex things that require more than a simple one-shot prompt. Before they may have used AI to automate one part of the workflow; now they're using it to automate entire workflows from start to finish, something a lot more complicated than, say, extracting data. This is part of Wiss's overall AI vision.
"Previously, it used to be, 'hey, extract these documents.' This is a mini-task. Now we think about end-to-end workflows, not just form extraction but building a work paper for me that my accountants can see and edit. How does that flow into my reviewer work? How does that flow into the tax systems? How can I review the end output of the PDF documents? How can I automate sending these PDF documents to my clients? It's from point solution to end to end workflows, that is the vision.
Beyond the cost of individual tasks themselves, there's also more people using AI than before. Comerford, from PKF O'Connor Davies, noted that costs have increased not only because the individual applications have grown more complex but simply because there's more people than ever using AI for these reasons, and they're using even more tools than before.
"Our costs within our existing tools have obviously gone up. We've added individuals, we've added people on the platform, and we've got higher rates of adoption so that is driving some amount of increase in our existing cost base for tools we had last year relative to the same tools this year and that's just driven by greater usage. Number two, we're also adding new tools in order to position ourselves to experiment and figure out what works for our PKFOD way of doing things. We're experimenting and piloting a number of different solutions and that is driving a portion of the cost increase as well," he said.
Cheryl M. Burke, the project management officer for PKF O'Connor Davies and a specialist in operations management, added that another issue completely separate from tokens is that the subscriptions are getting more expensive too as software providers continue to add new AI features into existing products that then drive price increases.
"Adding AI functionality into existing software platforms, that drives [cost increases.] Not all of them are consumption-based, some are still subscriptions, but there's these new features, new tools within the tools that are essentially add-ons, so that's another piece where software cost are increasing per person," she said in an interview.
Apigian, from LBMC, raised a similar point, saying "so many things now have 'it's the same thing but with AI' and now that's 10 bucks per person per month, so that's a huge cost to companies, probably more so than the tokens." Adding further expense is the cost of the entire data infrastructure that makes AI work in the first place, like Microsoft Azure or AWS. While traditionally these have been relatively fixed costs, the addition of AI features in both means that they, too, have gone up in price.
"We were using all these tools behind the scenes that were in Microsoft Azure or AWS and all of the sudden now they're charging per token. And so [before] all you had to do was pay for the right to use that and all of a sudden it was a new cost," he said.
Adding to all this is the fact that as models grow more powerful, they can do more things but also use more tokens. Marc Staut, shareholder and chief innovation and technology officer for accounting firm consultancy Boomer, noted that choice of model can make a huge difference in terms of both the predictability of costs as well as the overall expense.
"Model choice alone [can move] the bill. One member of our CIO community ran the same prompt for the same email through two models, got nearly identical results, and paid 50x more for one of them. If staff default to the most expensive model for everything, the firm pays for it," he said in an email.
This is why model control is so important to Pippadipally: without it, there's no control over how much compute the user chooses to burn on a given prompt, and so therefore no control over how much it will cost. This is why, he said, "if I'm not in control, I'm not going to pay consumption-based."
"I [want to] pick the model I want to use. I might use Fable only 1% of the time. I might use Sonnet 5 as my default model, which is ten times cheaper. So I have the control. … Think of your home. You switch off the lights when you don't need it. You don't run the water when you don't need it. So it's in our control. And I can teach those habits," he said.
Growing pains?
Despite higher costs, no one was planning to scale back their AI ambitions. In fact, all firms were doubling down on the technology, saying that while the cost might be high, the long term returns will far outweigh them.
"Our fundamental view is that AI is going to have a transformative effect on the productivity of our staff, and we want to make sure that we are positioning ourselves to take advantage of some of these advancements. To do that you need a mindset of being nimble and experimental because we're trying to build the business for the long term, and so we take a longer term view of some of our spending. So we're not optimizing for earnings today but we are optimizing to create the most valuable platform over the next 20, 30 years," said Comerford from PKF O'Connor Davies.
And while the current price spikes are certainly unpleasant, no one likes spending more money least of all accountants, there is a general impression that this is temporary. As the technology improves, there will be less reliance on pricey frontier models, with Apigian from LBMC saying there will be more powerful and reliable open-weight models as well as more effective on-premise AI that can be accessed on one's own device.
"If we can get a better way of doing that we'll become more efficient I think the costs are going to level, first of all prices are going to go down and then we're going to have new models that do so much more and we now have what are called open weight models that I could have on this machine that I'm sitting with. So you're gonna have more on-device models, which means on your phone, on your laptop, and then you're also gonna have cheaper models. So yes, it's a hurt as we figure this out, and it's a shock to the system. But I think it is going to be a new line, and it's going to be one that we accept," he said.
Burke, from PKFOD, felt confident about this as well since she has seen it before with cloud computing. Much like AI, its costs steadily dropped until, now, it is almost taken for granted. AI, she said, might see a similar curve.
"If you think about it from years ago when cloud computing started and everyone moved their applications to the cloud, it was 'oh we're going to add this to the cloud and there is an extra cost' and then you were always saying 'well something else may [get cheaper] because you don't need to host your own storage now or different elements.' I think that that's all still evolving in that kind of way," she said.
But while prices are indeed higher for various reasons, CPA firms are not taking this lying down. The sudden spike in AI costs has spurred new governance models, new budgeting practices and new technical guardrails meant specifically to curb costs and encourage efficient use of AI. We'll find out more on what that all involves in Part 2.
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