For much of the AI market, the conversation about cost has centered on a relatively simple question: What does a million tokens cost?

It is an understandable place to start. Tokens provide a useful way to measure model usage. They do not, on their own, tell us how capable the model is, how useful its output will be or what that output is worth.  

But focusing on token price alone leaves much of the economic picture out of view.

Consider two AI workflows that consume roughly the same number of tokens over a business day. One processes documents overnight. The other supports customers in real time. Their tolerances for delay differ, as do the consequences of failing to meet their service requirements. Token consumption alone cannot tell us what either workflow costs to operate or determine whether its outcome is worth that cost.

The token count may be the same, but the economics can be very different. 

Which raises a bigger question: What exactly should we be pricing? 

The price of intelligence depends on what we ask of it

It is useful to separate three questions that can easily get mixed together.

  • What are we paying for AI? 
  • What does it cost to put that AI to work? 
  • What is the completed work worth?

Token pricing helps us answer the first question. The second requires a broader view of how the AI is being delivered and operated. The third depends on the business outcome. They are clearly connected, but they are not the same thing.

We are already seeing this play out in the commercial models of leading AI firms. Pricing can change depending on whether processing happens immediately or can wait, whether previously processed context can be reused, what model tier is selected, what service level is required and how much volume is consumed.

AI pricing is already multidimensional. Enterprises need to manage it that way.

Instead of asking only, How many tokens did we consume?,  We need to ask a broader set of questions:

  • How quickly do we need it? 
  • How capable does it need to be? 
  • How predictable must its performance be? 
  • How much can we reuse? 
  • What happens when the system is wrong, and who bears the consequences?
  • What is the intelligence actually worth to the business? 

The answers will be different depending on the work being performed. 

Not every problem requires the same intelligence

This becomes much more important as AI moves beyond experimentation and into day-to-day enterprise operations.

A routine classification task may work perfectly well with a smaller, more economical model. A complex engineering, financial or strategic problem may need considerably greater reasoning capacity. 

Similarly, an AI agent that repeatedly works with the same instructions and enterprise context may be able to reuse information the system has already processed, rather than incurring the same processing cost again and again. 

Yet it is easy to treat every AI request as though it requires the same level of intelligence. 

We shouldn’t use the most powerful model for every task, or simply choose the cheapest model. The goal should be to match the cost and sophistication of intelligence required to the economic importance of the problem at hand.

Cheaper tokens do not necessarily mean cheaper AI

There is another wrinkle in all of this.

When token prices fall, it's tempting to assume that enterprise AI costs will follow suit. But that is not necessarily what happens.

Lower costs make more use cases economically viable. More use cases drive more consumption. And as enterprises move from individual prompts and copilots toward agents that continuously reason, retrieve information, call tools and perform work, that consumption can grow quickly. In many agentic workflows, each step carries forward more accumulated context, so the amount of information processed can grow faster than the number of steps.

We described this in our earlier work as the “token paradox.” Even as the unit cost of AI consumption declines, overall spending can increase as AI becomes embedded in more complex, continuously running workflows. 

But there is an important point here. A growing AI bill is not necessarily evidence of poor economics.

If spending is growing because AI is completing more work, and that work is creating meaningful business value, then the additional investment may be entirely justified. The concern is not higher consumption by itself. The concern is consumption that grows without a corresponding increase in useful, reliable outcomes.

This is why I don't think enterprises can optimize AI economics simply by negotiating a lower price per million tokens.

Cheaper intelligence is useful. Economically efficient intelligence is much more important. 

Who bears the consequences?

Every AI task and workflow has a cost that never appears on the token bill: the cost of being wrong.

When a model misreads a contract or misroutes a customer request, someone absorbs the consequences. That may be the employee who fixes it, the customer who gets the wrong answer or the enterprise, through rework, regulatory exposure or lost trust. The token price stays the same, but the economic impact does not.

When an enterprise pays the model vendor per token, it pays the same amount whether the AI gets the answer right or wrong. If the output is wrong, the enterprise also pays to find and fix the mistake.

A services partner that runs the full AI-enabled process can take on some of that risk. If the partner is paid for each invoice processed correctly, rather than for the tokens used, the partner pays for the errors. It has a direct reason to make the work accurate, whether that means choosing the right model, designing better checks or keeping people involved where it matters.

Neither approach is right for everything. For routine, low-risk tasks, paying per token and handling the occasional error in-house may be the most economical choice. For work where mistakes are costly, it can make more sense to pay for results and have a partner share the risk.

How risk is allocated shapes how intelligence should be priced.

From buying AI to managing a portfolio of intelligence

This leads me to what I believe is a more useful way to think about enterprise AI.

Today, organizations often ask which AI model they should use. Over time, I think the better question will be: What mix of intelligence does our business require?

Some work needs more capability. Some needs greater speed or risk assurance. Some can wait. And some may eventually make more sense under outcome-based commercial models rather than being tied to the amount of AI consumed.

The answers to those questions start to look less like choosing a single technology and more like managing a portfolio of intelligence.

For enterprise leaders, the starting point is to define what successful completion looks like for each important AI workload. Those requirements should guide the choice of models and commercial arrangements, not the other way around.

This is where I see the greater opportunity: allocating intelligence strategically rather than just buying it cheaply.

If intelligence becomes abundant, competitive advantage will depend on more than securing the lowest unit price. It may come from who gets better at deciding where premium intelligence matters, where more economical intelligence is sufficient and how to allocate each across the enterprise.

Moving beyond the token

This does not mean token pricing disappears. Tokens remain a useful way to measure model usage. But the enterprise objective is not simply to buy intelligence cheaply. It is to apply the right level of capability and assurance to the work that matters, at a cost justified by the outcome. That is the shift from purchasing AI to managing the economics of intelligence.

But we also don't have to invent these principles from scratch. Other industries have spent decades learning how to price scarce capacity, urgency, flexibility, risk, access and outcomes. Those lessons may offer some useful clues about what the economics of intelligence will look like next.

The implications extend beyond the price of AI itself. In our viewpoint Pricing when you can't predict, we explore how AI is reshaping the broader economics of managed services and why commercial models built for predictability will need to become more adaptive as a result. 

Read Pricing when you can't predict to get the full perspective.