Major accounting firms have been placing huge bets on artificial intelligence, having invested billions upon billions of dollars in the past few years alone. This is done with the understanding that AI will ultimately reduce expenses and drive profits. Yet, as always, it takes money to make money: fully realizing the potential of artificial intelligence can come with a hefty price tag, encompassing both short and long term expenses for not just the AI systems themselves but everything else that enables their effective use.
The AI models themselves, of course, represent a significant R&D expense. Whether for internal efficiency, client engagements or both, building and training these models is no casual affair, requiring skilled specialists operating sophisticated software to create, something with which Doug Schrock, managing AI principal for top 25 firm Crowe, is well familiar. His own firm has spent a great deal of money developing custom AI solutions for things like tax and audit that are now used by staff every day, as well as Crow Mind, a gateway portal for all of the firm’s AI solutions. It has also devoted significant resources towards building bespoke AI solutions for clients, particularly in cases where they need something that simply does not exist in the market today. He compared it to making a custom Excel spreadsheet but far more complex.
“It’s like you buy Excel. Here’s Excel. But you’ve got to configure it to your business case, so there’s a whole lot of customization to make the actual spreadsheet do what you need it to do. We see that a lot: you buy the suite, but you need a bespoke solution… Configuring the hardware, chaining together multiple agents to do the tasks, automating it, that takes work,” he said.
Studio217 – stock.adobe.com
Chris Kouzios, chief information officer for top 50 firm Schellman, added that developing an AI system may appear to be a one-time spend at first, but considering things like maintenance, integrations and upgrades, each model can also represent an ongoing expense.
“If you think of the initial build, you could call the initial build one time, although like any piece of software it will be continually approved over time, so I look at it from both perspectives,” he said.
Big data, big costs
But the development costs of AI models are only one part of the overall expense. Just as significant, perhaps even more so, are the fees that come with hosting and accessing these models in the cloud. Running AI, especially generative AI, is very data intensive, which has served to accelerate cloud costs that have
“Your compute will go up at least exponentially over time and one of the things I think we’re going to see, and this is just future forecasting a little bit, I think clients will in general, not just in my space, be more comfortable when they feel they’ve got a little control over what they’re doing and what is done. In the cloud at the beginning people were terrified of putting their stuff there, we’ll see the same stuff with AI, we’ll probably have additional costs for spinning up instances for clients nervous about what goes where,” said Kouzios.
Crowe’s Schrock reported similar things, noting that the major cloud hosting companies saw the opportunity for revenue generation via AI hosting and are already capitalizing on the situation, as evidenced in the fees they charge. The reality is that generative AI uses a lot of data, which means higher data costs from cloud providers who run the infrastructure it rests on. He talked about a recent meeting he had with Microsoft, a strategic partner with Crowe.
“They’ve got 4 million servers across the US. They’re super interested in AI, not just because of Copilot but because we’ll be using Azure, using their server computing power to run the LLMs we write. They want to drive more Azure service dollars. So… we’ll be having more computing power costs for us through Azure,” he said.
Accounting solutions vendors have noticed this too. Brian Diffin, chief technology officer for business solutions provider Wolters Kluwer, also noted that generative AI has indeed led to higher cloud costs, which has challenged the company to find ways to release AI-functional products in an economically sustainable way.
“Gen AI is very CPU intensive, so one of the challenges we face—we’re doing a lot of experiments with this— is there’s so many approaches on how you would implement a gen AI based piece of functionality in software. We’re evaluating not just the LLMs in terms of what those capabilities would produce but what is going to be the cost of that feature when we go to production,” he said.
Data shows that this is happening not just in the accounting space but across the economy as a whole.
“GenAI is creating a cloud boom that will take IT expenditures to new heights,” said Chris Ortbals, chief product officer at Tangoe. “With year-over-year cloud spending up 30%, we’re seeing the financial fallout of AI demands. Left unmanaged, GenAI has the potential to make innovation financially unsustainable.”
The report noted that cloud software now costs businesses an average of $2,559 per employee annually. Large organizations spend an average of $40 million on cloud fees annually, with very large organizations worth more than $10 billion spending $132 million annually.
However, while cloud costs are rising due to AI, leaders are also confident that they can be managed. Schrock said his own firm has controls in place specifically to monitor data usage to avoid outsized costs. For instance, recently they tried a new LLM tool from Microsoft that caused a short 3,000% spike in usage, but firm leaders received an alert and quickly stepped in.
“It’s not like when you get surprised by the electric bill. You put controls in place to do things smart,” he said.
Further, while the costs have increased, he said they have still gained more than they lost in terms of increased efficiency and productivity. The extra fees are still lower than the cost of hiring an entirely new human, and the quality of work is better than what humans would accomplish alone. So while their Microsoft Azure bill is higher, they’re also able to deliver more for less cost overall, so it has been a net positive.
“What we’ve been talking about are the costs to run AI. I’ve got the cost to run a car but it also gets me places more easily. The cost will be a thing but used appropriately it will be great,” he said, adding that it’s important to use the right tool for the right situation; maybe you don’t need to access the high-data AI model to solve a problem, maybe Copilot would work fine.
Diffin raised a similar point. While he conceded overall costs have gone up, the money has been well-spent in terms of product development.
“Certainly gen AI capabilities are increasing in cost, and overall costs have gone up because we’re using more and more of what [Microsoft] offers, and so what translates into for us is developing and releasing products faster than if we were to develop everything ourselves,” said Diffin.
On top of cloud fees, subscriptions and licenses were also mentioned as a significant ongoing expense. This includes subscriptions not only for the tools used to create and maintain AI systems but also for AI solutions that the firm chooses to buy rather than build. While the individual subscriptions may not be much, when considering the size of certain firms, like Crowe, they can quickly add up, especially considering there are multiple products the firm subscribes to.
“Everything is a subscription. So you have all the different types of subscriptions. Crowe is making significant investments in ongoing software licensing for the leading enterprise AI solutions, things like Microsoft Copilot for example. We expect everyone in the firm to be using that in 2025. It’s over half right now … We’re also buying specialty AI based applications to fit particular needs and things like copy AI for marketing and search, and there’s a whole suite of specialty apps that we sign up for with specialty use cases, so that becomes the ongoing expense,” he said.
Labor costs, training costs
And then there are the people who create and maintain these models, often software engineers and data specialists. While often touted as a labor saving device, AI can come with surprisingly large labor costs, according to Schellman’s Kouzios.
“I would say in general, probably as close to 15-20% of my IT budget will be spent on AI, closer to 25% for the first year [of deployment]. Of that, if you take that number and break it out, 85-90% is labor,” he said.
The firm, which already hosts a large number of technical specialists, recently hired more to support the firm’s AI ambitions, seeking to shore up its machine learning, data analytics and product management expertise, which allows its staff to focus on “building what it is we want to do.” While this does represent a spending increase, he is confident that the efficiencies they uncover will increase firm-wide capacities over time.
“I think we’ll get to a point where, [though] we know the costs will go up, ROI on this should be deferral of cost or deterrence of cost, not having to spend money in the future we’d otherwise have to spend. For example, peak season comes up and you need to either hire employees or temp employees,maybe we can avoid that in the future,” he said.
Another component of labor costs is training the non-technical staff in using the AI systems the technical staff develops and maintains. Schrock, from Crowe, said that, in addition to hiring more experts, the firm has dropped cash on in-depth training and development in things like how to use Microsoft Copilot and other generative AI tools and incorporate them into a workflow. With this training has also come changes in business processes and job descriptions that needed time to properly digest. While there is some learning curve involved, he felt education like this was essential to fully implement the firm’s AI vision.
“These tools don’t inherently have value, they derive it only through their application to solve problems. So there is one time cost of upskilling and process redesign to incorporate that into the business,” he said.
And it is not just the humans who need training. Kouzios said one idea he has been exploring lately is assigning those trainers who’ve been educating the human staff to the AI models themselves, which often begin in an almost child-like state and require data input to be effective.
“I’ve been exploring talking to them about training the models because, this is my experience in IT, nerds are very good at the tech, but here are some things we lack and teaching—when I brought it up to them, I meant teaching the models—the tech people hated the idea, so I might tap into some of [the trainers’] time too,” he said.
Heat vs light
Yet, while big money is being spent on AI at accounting firms, they should not necessarily take too much stock in the marquee headlines of this firm spending that many billions on AI or that firm spending many more billions still.
“The billions of dollars here, is more bragging about an investment level. Well, investment level can be measured in a number of different ways. It can be measured by some ginned up cost where you reallocate peoples time and come up with some marketing number on costs, but I don’t put a lot of confidence in those as an expert in the field,” said Crowe’s Schrock.
Kouzios, from Schellman, raised a similar point, noting that there are a lot of people making big dramatic announcements that, upon closer inspection, are not that significant.
“You’ve seen those press releases, saying we bought chatGPT for our 85,000 employees, we’re AI enabled. Yippee, well done. For 20 bucks a month I could do that too,” he said.
When looking at what firms are spending on AI, Schrock said to look not at the jaw-dropping number they announce but in actual deliverables they produce.
“What I wanna understand is how many people are utilizing it, what unique IP they have created, how aggressively is it being incorporated into service lines, how aggressively do they take this into market—that is a measure of your investment level in AI more so than some number,” he said.
But what about smaller firms? Turns out, their experiences with AI costs are much different than large scale firms with international footprints. We intend to explore this issue more deeply in another story soon.
Credit: Source link
