Nearly nine in ten respondents from a McKinsey survey report regular AI use in at least one business function. Similarly, 54% of larger companies report enterprise-wide scaling of AI.
This is good news, right?
For years, one of the biggest challenges with AI adoption has been getting employees and customers to use it instead of finding ways to work around it. Now, AI is becoming a regular part of how organizations operate, whether everyone is enthusiastic about it or not. However, as companies move from deterministic AI to scaling AI agents, they're running into a different challenge: cost.
One in five organizations in McKinsey's survey say AI operating costs are already constraining AI use. Amid all the excitement around AI and the pressure on companies to take the next step into the future, it's easy to overlook a basic reality: more capable technology comes with a price tag.
Still, major organizations aren't backing away from AI. They're trying to figure out how to afford and manage more of it.
Take Uber's CTO as an example; he made headlines earlier this year after revealing that the company had already blown past its AI budget in the first four months of the year. Yet Uber isn't the only organization increasing their usage and spending. About 60% of organizations expect to increase their AI investment over the next year, even though many continue to struggle to demonstrate clear ROI from those investments.
So, what does this tell us about where AI is headed? And as spending continues to increase, how can small, midsize, and large companies prepare for what their AI bills might actually look like?
Let's explore this together.
Have you ever walked into a store because everything was on sale, only to walk out having spent more than you planned? Everything costs less, so you put a little more in the cart.
AI costs can work in a surprisingly similar way.
Sam Altman, CEO of OpenAI, has predicted that the cost of using AI will fall by roughly ten times every 12 months. Why? Because improvements in hardware, more efficient models and algorithms, and massive investments in AI infrastructure have all helped drive costs down.
But lower underlying costs don't necessarily translate into lower overall AI spending. As AI becomes cheaper and more capable, organizations are finding more places where it makes sense to use it, expanding adoption across employees, departments, workflows, and customer interactions.
Agentic AI adds another layer to the costs because one seemingly simple task can involve multiple model calls, tool calls, reasoning steps, and actions. An employee might ask an AI agent to prepare a summary of last quarter's sales performance. What seems like a simple task to them can involve much more happening behind the scenes. The agent might need to retrieve data from multiple systems, analyze the results, identify trends, compare them with previous quarters, and generate the final summary. What looks like one request may actually involve multiple model and tool calls to complete it.
That's why falling unit costs don't necessarily mean lower AI bills. As your company finds more ways to use AI, those seemingly small interactions can turn AI into a much bigger expense than expected.
Uber blowing through its budget wasn't the end of the conversation. It forced the company to start thinking differently about how AI consumption fits into its overall operating costs. Once AI becomes a meaningful operating expense, companies have to start comparing token consumption and its associated costs with other resources, including headcount. Suddenly, an organization isn't only asking whether AI makes employees more productive. It also must consider whether the additional AI consumption is worth the investment and at what point another solution might make more economic sense.
Axios reported that an AI consultant said one of its clients spent $500 million in a single month on Claude after failing to establish usage limits for employees. At this point, you might be thinking the solution is to put limits on your AI usage. But it's not that simple.
If we spent years trying to make employees and consumers embrace AI, are we going to now discourage people from doing so? No. More AI usage isn't necessarily a bad thing, especially when that usage is creating measurable value. The challenge is understanding when it is and when it isn't.
It does, however, introduce an interesting topic called: tokenmaxxing.
If you haven't heard of this word before, tokenmaxxing describes the tendency to maximize AI usage when generous or unlimited access is available, sometimes without a clear connection between that consumption and the business value it creates.
As Dave Treadwell, an Amazon senior VP, suggested, don't use AI for the sake of using AI. Use it to solve customer problems, solve business problems, and innovate. That's where the distinction between consumption and value becomes important. Adopting AI shouldn't be about consuming as much AI as possible. It should be about creating traceable, valuable outcomes.
But companies also need to be careful about swinging too much in the opposite direction. Restricting access too aggressively could push employees who have grown dependent on these tools toward unapproved alternatives. This behavior contributes to Shadow AI, the use of AI tools without organizational approval or oversight, creating risks for adoption, governance, security, and visibility.
A better approach is to look at AI consumption through the lens of outcomes. If increased usage leads to greater productivity, better customer experiences, more revenue, or other meaningful business results, then a growing AI bill may be perfectly justified.
The reality is that controlling AI costs isn't going to be only about managing usage. It's about understanding what that usage produces and whether the value outpaces the investment. And as AI adoption continues to grow, how organizations pay for it becomes an important part of that equation.
If companies don't know exactly how much AI they'll consume, and consumption itself doesn't necessarily equal value, how are they supposed to budget and pay for it?
There's no single answer yet. CX Today identifies at least six pricing approaches emerging across the AI market, including seat or bundled pricing, channel-based consumption, component-level usage, credits or tokens, action-based pricing, and outcome or resolution-based pricing.
The variety alone tells us something. The market is still figuring out what companies should actually pay for when it comes to AI. That decision gets even more complicated because the advertised price of AI isn't always the full price of running it. Model usage may be only one part of the bill. Depending on the use case, companies may also need to account for integrations, telephony, orchestration, governance, monitoring, and even the costs associated with interactions that don't go as planned.
All of this makes forecasting difficult. A company may be able to estimate how many employees will have access to an AI tool, but predicting how much they'll actually use it is another story. Customer-facing AI brings similar uncertainty. Demand can fluctuate, new use cases can emerge, and agents may consume different amounts of resources depending on the tasks they're asked to accomplish.
Ultimately, the right pricing model depends on how your organization plans to use and scale AI. A retail company, for example, may experience significant seasonal peaks around the holidays. A usage-based model could allow spending to rise alongside higher demand and come back down when that demand subsides.
The point isn't to find one pricing model that works for everyone. It's to find a commercial model that reflects your organization's demand, use cases, industry, and plans for growth.
And you may not know exactly what that looks like from day one. Scaling AI sustainably isn't just about finding the lowest price. It's about finding a model that gives your organization room to grow while keeping spending connected to actual demand and, most importantly, the value that AI creates.