As generative AI moves from pilots into everyday work, the question of what it actually costs has become harder to answer. When companies first started using AI tools, they were mostly buying flat-rate licenses, which made budgets easy to predict. That era is ending. In 2026, AI providers moved to consumption-based pricing, meaning companies now pay for every request a model handles. With more—and more complex—agents than ever, requests can quickly stack up. The costs of running AI also show up in cloud bills, software subscriptions, and the hours people spend checking outputs.
In this video Explainer, McKinsey’s Lari Hämäläinen, Wasim Lala, and Kaavini Takkar break down why AI bills keep climbing and what it takes to run an AI agent. They draw on examples from their client work, including one bank where projected AI spend jumped from $3 million a year to more than $25 million once employees had access to every tool and model with no guardrails in place.
The three agree that AI costs can be measured and managed and that control starts with visibility. They differ on some of the details, including whether the falling price of a token tells leaders much at all. Each offers a practical way to judge whether an AI tool is worth what it costs.
This article has been edited for length and clarity.
Why is it so hard to tell what AI really costs?
Wasim Lala: AI is a very fast-growing technology and is truly consumption based. I think we’re at a level of maturity right now where our estimates of what something will cost when we put AI and generative AI into workflows amount to what I’d call napkin math.
Kaavini Takkar: Our research says that, on average, when you move from pilot to production, your AI spend nearly quadruples. You’ve got AI spend rolling through your AI provider. Then there are the indirect spend buckets. Every time there’s a model call on the cloud, there’s an inherent compute cost associated with it. And then there’s all the AI embedded in SaaS [software-as-a-service] tools. So you first need to identify all the sources of AI spend in an organization. Then think about going from pilot to production. A lot of organizations are experimenting and running many pilots.
Lari Hämäläinen: Predictably, most of the cost today is on licenses. What is not predictable is the cost of giving open-ended access to AI tools. Here’s an example: Many McKinsey colleagues use Cursor. Most Cursor users, if given a choice, choose the latest frontier models. And that then often results in quite unpredictable costs.
Wasim Lala: We’ve seen this pattern across organizations. Between months two and four, their usage doubled and their spend doubled. In the first couple of months, [Claude’s] Sonnet was the model being used. Months three, four, five? All of that usage had switched to [the more advanced model] Opus. I’m pretty sure all those users didn’t need a much more advanced model. They saw the ability to access a more powerful model, and they went there.
Kaavini Takkar: AI cost is perhaps not hard to decipher, but it is very complicated. However, with the right visibility, it’s solvable.
What is a token?
Wasim Lala: A token is the smallest unit of information a model sees. These are the pieces that carry data into the model. The model then looks at those pieces of information and uses a weighted average to determine what the next word could be.
Lari Hämäläinen: If you look at how language models, and the agents built on them, work today, at the end sits a token. It’s not quite a syllable. But it is something comparable to a syllable of a word, or it could be a set of pixels in a picture, or it could be a set of pictures in a video frame.
Looking for direct answers to other complex questions?
Kaavini Takkar: A token is a chunk of text. At its core, it’s a unit of measure that AI providers use today to bill their clients. Typically, billing happens per million tokens. At a very basic level, if I were to say, “Hi ChatGPT, how are you?” that could be about six tokens. But if I included a lot of images or references, I would essentially expand that context window.
If it were as easy as counting words, everybody would be able to do it. But there are input tokens, or the prompt you’re giving the system, and there are output tokens, the output the system is generating. Output tokens are way more expensive than input tokens. There are also context windows. People are providing as much reference as possible to get the best outcomes. Then you have reasoning models. There are intermediate thinking tokens that are generated that we, as users, don’t get to see. But we’re essentially paying for it.
Why is AI pricing moving from per seat to per token?
Lari Hämäläinen: It is simple economics. If you look at the labs—Anthropic and OpenAI—mostly two things have happened. Number one, they either increased or capped pricing. Number two, they’re increasingly shifting to consumption models, in which consumers pay per token. Both models coexist, even within the same tooling, but we are seeing a transition toward consumption-based models for pretty much all AI tools.
Wasim Lala: When generative AI became widespread, a lot of us were used to models where we paid monthly subscription fees and got effectively unlimited tokens. In the first quarter of 2026, a lot of the frontier companies switched from per-seat licensing—that is, per user—to fully consumption-based pricing, meaning the more you eat, the more you have to pay. And so, as these contract structures evolve, you could be using the same amount of AI but seeing two to three times the cost instantaneously.
Kaavini Takkar: Per-seat pricing shaped how many of us learned to use these tools. I could prompt a million, two million, three million tokens, and I’m still covered under a per-seat licensing umbrella because my price is fixed. But now I’ve gotten used to not prompting efficiently, and vendors have switched to a consumption-based model. So with every extra word I’m typing, every extra file I’m putting in, I’m forgetting how to cache and forgetting how to reduce my context window.
If the price per token keeps falling, why are the bills going up?
Kaavini Takkar: This is the standard Jevons paradox. As the cost comes down, demand and usage go up. In the mobile world, you may recall that as prices per gigabyte fell, people suddenly shifted away from email to video calls. The same thing is happening in the AI world. The per-token price has fallen by almost 90 percent, a figure commonly cited in the industry. But the key to focus on here is the volume, the consumption that has drastically gone up. Everyone’s experimenting. They’re building their own agents.
Wasim Lala: Initially, we were mostly using web-based windows to correspond through a chat interface. We have now entered the agentic paradigm, where we are using agents to think through tasks and connect with the large language model [LLM] directly. What’s coming next is networks of agents: not just one agent connecting with the LLM, but one agent orchestrating four subagents to go into the LLM. So you can see the exponential growth in the number of times we send tokens to the LLM and receive outputs.
Lari Hämäläinen: Up until the end of the first quarter of 2026, everybody was worried about, “Are we getting enough usage?” Every enterprise executive was asking, “How do we drive adoption in our organization? How do we get people to use more, and more, and more?” And then the coding models, which are driving the current discussion around token economics, started to take off at the end of last year. So by April and May 2026, executives were seeing both high adoption and uncontrolled usage. And people suddenly woke up to the fact that a single user could rack up thousands of dollars in a single day.
Wasim Lala: It’s not just about cost per token. It’s about cost per task. That’s really what we are paying for at the end of the day.
What does it cost to run an agent?
Wasim Lala: It’s hard to say how much an agent costs, right? It’s almost like asking what color water is. If designed poorly, the same agent could consume up to 30 times as many tokens [as if it were designed well] because it could be really inefficient at figuring out how to do a task. In infrastructure operations, not all parts of the process need agentic AI. Many other parts of the workflow are typically very clear: “If it’s this kind of device, go check this repository.” That does not need agentic AI. You could put generative AI in there, but it would be a waste of the token cost.
Lari Hämäläinen: The models are stateless, which means that if you want a model to remember something, you need to feed everything back into it. So, it keeps repeating; it keeps compounding. And that’s the core essence of the models and agentic uses today: The longer you discuss with ChatGPT or Copilot, the more the costs compound. The more context you provide, the higher the cost. And then there’s interpreting the output, refining the output, and all the other layers of the agent’s existence, which consume even more tokens.
Kaavini Takkar: Every single time an agent receives a prompt, it is going to make model calls. You multiply your model calls by the per-token price. That will give you one cost element. How much does it cost for a human to validate or review the outcome? There’s observability. There’s vector storage. There’s orchestration cost in the mix. Think about an airline that has built an agent for rebooking. By the time users make up their minds, the agent may have to inform them, “This flight is actually already gone. Let’s go through this process again.” That failure rate is also costing a lot of money because you’re having to repeat the same process.
How do you know if an agent is worth it?
Kaavini Takkar: One good metric to look at would be the cost of the agent per successful output. What are all the costs associated with inference, human validation, and vector storage? And how many times has the agent failed to deliver the expected output? Then compare that cost with the alternative you had: the fully loaded cost of a human or of another tool that had been doing the job. And how frequently is that agent being used? There are thousands of rebooking requests coming to an airline every day. But an agent that just audits tax filings for employees? Perhaps once a year.
Lari Hämäläinen: Today, we’re using AI very much as a blanket term. If you simplify it, there are three different types of AI. There’s personal tooling that affects your work as a person. There are automation- and efficiency-driven AI applications: We take what exists today and rebuild it for higher productivity. And then third, there’s rebuilding for AI and running the business differently. All those have very different value-versus-cost equations. For personal tooling, the value is very indirect. It’s more the cost of doing business. And then on the other side, for automation, it’s very much a question of, “How much can I save? How much does it cost to build this? What is the maintenance cost?” And you have to understand that equation.
Wasim Lala: There was one small team that I worked with. It had five engineers, and they were completely overloaded. We helped them redefine their workflow so that the inputs coming to them said, “There’s a problem. We have done the diagnosis. Here’s the proposed resolution. Do you agree?” This created about 20 to 30 percent capacity in their days. Was there any dollar value in that? Not really. I still need those five engineers. But it is not enough to just create an agent and do something faster. Otherwise, it’s like building roads. You can build a lot more roads, but you’ll just fill them with more cars, and you’ll still have the same traffic problems.
How can companies get AI costs under control?
Lari Hämäläinen: Companies need to be in control of their own AI. Companies should own their own workflow design and process designs for AI. They need to own the harnesses, which is the software that sits around the language model. They need to control and own their data. And advocating for open-source models, where you can control what the model does and where it runs, will be preferable for many of our clients.
Kaavini Takkar: One of my clients is from the banking industry. When the client started, their annual spend was in the $2 million or $3 million range. Eventually, they began signing new contracts that were consumption based. Everyone in the organization then had access to every single AI tool, every single model. Without any governance guardrails in place, their spend projections went way above $25 million a year. And that’s when they said, “I need one AI control plane, one place where I can bring all these different sources of my AI spend together, so I can track it back to an individual and understand, why didn’t you route this particular request to an open-weight model?”
Wasim Lala: Am I using the right model for the right task? Am I teaching my employees how to write the prompts so I’m not overloading the context window? I have seen folks entering large documents into Cursor to get one or two facts. Now imagine at a 100,000-employee scale, everyone doing this 20 times a day. There’s certainly going to be value leakage. We have been on the cloud journey for 13 years now, and I routinely find savings opportunities of 15 to 20 percent of cloud spend. You can always find the savings. Executing on them is the hard part. So automating these controls is going to be really important.
Is AI worth it?
Wasim Lala: I think it is absolutely worth it, with a big caveat: We use it for good, for the kind of outcomes we want to drive, and we do it carefully, with a commitment to also upskilling our workforce at the same time. AI will be worth it as long as you can measure it. Without measurement, there can be no success.
Lari Hämäläinen: In some ways, yes. In many ways, I personally feel it’s not worth it. It’s always worth it for an individual company. It’s always worth it for an individual to use it. But it still might not be worth it for society. Let’s say you’re competing in a 100-meter sprint or in weightlifting. If you know that everybody else in the world is adopting a new advantage, you probably would adopt it as well, right? As a society, we tend to optimize for profits. So, of course, you apply what you can apply. Then there’s a different question: Is that good?
Kaavini Takkar: The answer is yes. But there are layers behind that yes. The first step is to ask what workflows do I have in my organization? How do I reimagine them? Then, how do I automate? Then I should think about the AI aspect of it all. If those are the steps I am pursuing, it will always be worth it. I’ll share an example. Two clients brought the exact same use case from pilot all the way to production, with drastically different results. One had massive efficiency gains and higher productivity [than the other] because they started by reimagining the workflow. The other just threw a bunch of AI tools at the process.
