AI tools have become commonplace at advisory firms, shaving time off tasks from meeting prep to research. But many firms still have only a hazy understanding of how the technology is priced — and why falling token prices can still produce rapidly rising AI bills.
A couple of months ago, Kevin Hughes, president of financial planning at Advyzon Investment Management, tested out a planning agent. During testing, "one of our developers came back and said: 'did you know that just doing one of these things is 50,000 tokens?'" he said.
Hughes isn't alone. Token costs fell from $20 per million tokens to as low as 7 cents through 2024, according to a McKinsey report citing Stanford Institute for Human-Centered AI (HAI) data, yet enterprise LLM spend tripled over the following 12 months.
In light of this, 93% of survey respondents reported blowing through their token budgets and one-fifth said they have had to reel in their AI usage due to rising costs.
Firms in search of a way to curb costs can start with a few concrete moves:
- Match the model to the task
Audit the tech stack before layering in AI- Define if the firm is on a fixed or variable budget before scaling usage
- Measure against outcomes
READ MORE:
Tokens: An explainer
Tokens are the units used to meter and price AI usage. Each AI prompt or response is broken into numbered fragments before a model can process it, and those fragments are called tokens.
AI models break text and other information into small units called tokens before processing it. Input tokens include the information sent to the model — such as a prompt, document or conversation history — while output tokens are generated in the model's response. Providers typically charge separately for each, with output tokens often costing considerably more.
Using Google's Gemini, Sentisight added, about 1,000 tokens equate to 750 words, and for Gemini, four characters cost one token. The word "fantastic," Sentisight wrote, "might become three tokens: 'fan,' 'tas' and 'tic.' Each token receives a numerical ID."
As models become more intuitive and efficient, that reduces the amount of required token processing, "which directly lowers computing costs," wrote Sentisight.
Convenience, falling costs and rising spend
The disconnect between cost and consumption requires thought before addressing. Many firms and advisors are used to the more commonplace subscription pricing models, such as CRM software.
Users should approach tokens more like they would approach an old-school arcade video game coin slot.
"You put your quarters in it, and you keep playing until you're officially done," said Hughes, "or you stop and come back at a later time to finish the project when the clock resets."
At present, token model pricing reveals a wide spectrum, Ray Wu, founding managing partner of Alumni Ventures, said in an email. "OpenAI currently offers models from roughly 20 cents per million input tokens for GPT 5.6 Luna to substantially more for frontier models, while Anthropic prices Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens."
The decline in costs, however, does not translate to actual money saved or token-related bills falling.
"Companies are seeing their AI budgets rise rapidly because they are consuming tokens at mass scale," said Wu. "It is very similar to what happened with cloud computing. Storage and compute became cheaper, but once companies moved hundreds of workloads to the cloud, cost management became a discipline of its own."
But what about ROI?
Another layer underneath token spending and consumption is the inability to effectively measure the ROI. One issue, according to Todd Ahlsten, chief investment officer of Parnassus Investments, is that, "Everybody's using tokens for different things.
"You might use a token to help research something. I might use a token to pick a stock. Someone else might use a token to do HR and payroll," he said. "The value of that token is radically different."
This inability to measure ROI, according to Ahlsten, means, "there's going to have to start being a rationalization of it. It's starting to get expensive…and it's really hard to measure the ROI," he said. "CFOs around the country are starting to go 'whoa, this expense is becoming a really big expense,' and it's going up at a vertical level."
READ MORE:
Addressing the dilemma, for now
To combat some of the issues, firms and advisors should remember that every task doesn't demand the costliest tools.
"A high-end reasoning model may make sense for analyzing an unusual tax situation or complicated portfolio," Wu said, "but it is unnecessary for summarizing a routine meeting or classifying a document."
Firms can keep spending grounded by designating cheaper models for simpler work and deploying costlier models when extra intelligence is applicable. A foundation for that discipline is knowing what's actually running underneath the AI layer.
Because AI is only as useful as the data it can see, firms limit its capability if they are running a patchwork of disconnected tools.
"I'd look at my tech stack, and I would ask myself: how strong are all these integrations today," said Hughes.
Staying grounded also means knowing the firm's budget and sticking to it before rollout. A firm that doesn't know whether it's on a fixed or variable model risks getting surprised mid-project: The token-clock equivalent of running out of quarters.
Those firms that are used to budgeting line items, in particular, need to be prepared, according to Ahlsten.
"Let's say you own a business, and you're like, 'hey, listen, I'm going to spend $50,000 a year on IT or something,'" he said. "If it's tokens and you don't have control of that, there's things all of a sudden getting real now."
Ultimately, keep in mind that it is the return that matters more than the receipt.
"If an AI workflow costs $2 but saves an advisor 30 minutes, that may be an excellent economic trade," Wu said.
Even so, that cheap workflow triggered millions of times a year, without adding real value, can cost more in the long run than it looks like it will on paper.
"I don't think the winning financial institutions will necessarily be the ones that spend the least on AI," said Wu. "They will be the ones that understand the economics of intelligence."
