CORE IDEA: Modern AI agents think before they act, and their thinking is metered. Most of that spending is decided by the machine, not the user. Managing AI now means managing how much thinking we ask it to do, and whether the work is worth the price.

In 2024 the question about artificial intelligence was one of capability. Which model was smartest? Who was winning? By 2026 the question has become financial. Can we actually afford to run this at scale?

The paradox is that AI has never been cheaper. McKinsey reports that model costs fell by more than 99% during 2024, from around $20 per million tokens to roughly $0.07. Yet enterprise spending tripled in a single year, budgets are routinely exceeded, and some companies are now capping their own usage to control costs. It is cheaper by the unit, but dearer by the task. Something fundamental has changed in what we ask these systems to do.

From chatbots to colleagues

The AI that most organisations met in 2024 was simple. You typed a prompt, the model produced a reply, and the exchange ended. Costs were modest and, importantly, predictable. Deloitte, surveying enterprises through that year, captured a mood of cautious experimentation.

By late 2025 the centre of gravity had shifted from chat to agents. Deloitte found agentic AI to be the single most sought-after development, with a quarter of organisations already building autonomous agents in earnest. An agent does not simply answer. It plans, searches, calls tools, checks its own work and decides its next step, often across hundreds of rounds with little human involvement. That step change in capability is precisely what rewrote the economics.

What you are actually paying for

AI is billed by the token, a small piece of text roughly the size of a short word. To read a modern AI bill, it helps to separate three kinds of token, because their costs behave very differently.

The three kinds of token

Type

What it is

What drives it

Input

Everything you send the model: your instruction plus the background it must read, such as documents, data and earlier steps.

Grows with how much context you supply.

Reasoning

The model’s private thinking as it works towards a result: planning, weighing options, checking itself. You rarely see it, but you pay for it.

The agent decides how much to do. This is the hidden cost.

Output

The visible answer: the reply, the report, the code, the drafted email.

Grows with answer length.

In a chatbot, input and output dominate, and you control both. In an agent, reasoning is usually the largest cost, and the agent controls it.

The essential point is where control sits. In a chatbot the user holds the wheel, choosing what to ask and how much to send. In an agent, reasoning dominates, and the machine steps on the accelerator.

Why the bill keeps rising

McKinsey identifies six drivers that explain why agentic workflows cancel out falling token prices. The first two matter most, because they sit largely outside the user’s control.

The six drivers of agentic cost

Driver

Why it inflates the bill

Long-lived context

Models have no memory between steps, so an agent resends earlier context again and again. Agentic tasks can consume roughly 1,000 times more tokens than a single question.

Refinement is the sink

The costly part is not the first answer but the checking, repairing and re-verifying that follows. About 60% of an agentic task’s cost goes on refinement.

Autonomy creates variance

The same task can take different paths and retries, producing very different costs. In programming, the same job has shown a thirtyfold variation between runs.

Premium models for simple work

Sending every task to the most powerful model is like taking a Formula One car to buy groceries. Deep reasoning is wasted on easy work.

The ecosystem charges too

Security, orchestration and monitoring tools increasingly bill by the token, and the same tokens may be metered several times as they move through a system.

Structure and language

How information is presented affects cost. Dense prompts waste tokens, and non-English text is often split into more tokens for the same meaning.

Source: McKinsey, “Is that AI agent worth it? Agentic economics and the modern operating model”, July 2026.

The psychology of more

These drivers explain how the bill grows. They do not explain why we allow it. The deeper cause is not technical but human. In 2024 we typed a prompt and read a reply, because our expectations were modest. As we have grown comfortable, those expectations have expanded. We now ask machines to plan, research, verify and work unattended for minutes at a time. We have begun to demand that they think, and think deeply, on our behalf.

Thinking, for a machine, is computation, and computation is billed. The model applies no judgement to the value of a request. It does not pause to ask whether a question deserved so much effort, nor offer a discount for a trivial task handled lavishly. It simply thinks as hard as we have designed it to, and meters every token regardless. The technology is perfectly indifferent, reflecting our demands back to us, priced to the decimal.

A person, by contrast, often gives more than they are paid for, rounding the effort up out of pride or goodwill. A machine rounds nothing up for free. The same logic applies to the businesses behind it. Having moved from subscriptions towards consumption pricing, most now earn more when we consume more. This is not a criticism; it is the plain economics of expensive computing. The point is simply that the incentive to escalate sits on both sides of the exchange, and few forces naturally push the other way.

Treating AI as a utility

The most useful mindset is therefore the humblest one. AI is not a colleague who volunteers extra effort. It behaves like electricity or water, billing strictly for what it consumes. You would not leave every tap running because water is cheap by the litre, nor feel wronged by the utility for metering it. You manage your own consumption. The same discipline now applies to intelligence.

This shifts the executive question from how much AI costs to whether the cost of a machine’s thinking is justified by the value of the outcome. The answer is a balancing act rather than a one-off cut, and it rests on a simple principle: spend thinking where thinking pays.

Matching the cost of thinking to the value of the work

Not all thinking is worth the same. The value of a token depends on the gap it bridges between the task and the person doing it. A lawyer cannot hold every statute in memory, nor keep pace with each amendment. Here deep reasoning is not indulgence but competence, and a generous budget is easily justified, because the alternative is not a slower professional but an unreliable one.

A junior analyst producing the same report each morning is a different case. The work is repetitive and largely knowable, and paying premium rates to reason it out afresh each day is the equivalent of sending a Formula One car to fetch the shopping. Yet even here the answer is not simply to spend less, but to spend differently. The tokens are better invested once, in teaching an agent to produce the report, than paid out daily to generate it by hand. The cost belongs to the transition of the role, not the repetition of the task. In the process the analyst is not diminished but promoted, from performing the work to designing the system that performs it.

None of this is mechanical. Roles are neither uniform nor linear, and a routine job has demanding moments just as a demanding job has routine ones. The aim is not to price every role to the decimal, but to hold a rough sense of the intellectual demand each one carries and to budget in that direction.

Measuring to notice, not to police

This is where measurement earns its place. Once each role carries an expected level of demand, the token report becomes a diagnostic rather than a bill. The interesting number is the one that does not fit. Why is a junior analyst consuming so much? Perhaps the role is harder than anyone assumed, or perhaps it is ripe for automation. Why is a senior lawyer consuming so little? Perhaps that is discipline, or perhaps a useful tool is going unused.

In each case the gap between expectation and reality points to something worth knowing: a role to redesign, a task to automate, a habit to correct or an adoption barrier to remove. An outlier is not a verdict but an invitation to look closer. Sometimes the heaviest user is the best return on intelligence, and the lightest user is the real problem. Managed this way, cost data stops being a bill to dread and becomes a map of how work is actually done.

Five habits that keep spending tied to value

Habit

Why it helps

Measure per user and task

You cannot balance what you cannot see. Showing people the cost of a request often changes behaviour on its own.

Match the model to the work

Reserve frontier reasoning for high-value, demanding work. Smaller models and plain automation handle routine execution at a fraction of the cost.

Set hard stopping rules

No autonomous agent should run without explicit limits on how far it may go and how much it may spend.

Keep prompts and context lean

The more you resend, the more you pay. Concise prompts can cut consumption by 30 to 40% without compromising quality.

Judge by outcome, not tokens

Track cost per completed outcome rather than raw throughput. Tokens are the bill; the finished work is the value.

The real prize

None of this is an argument against agents. Deployed well, they perform work that was previously impossible, and that work can be worth far more than it costs. The danger is not spending on AI. It is spending without knowing what we are buying, or why.

Falling token prices created an illusion of ever-cheaper intelligence. The agentic era has replaced it with a subtler truth. The unit is cheap, but the thinking is not, and the machine now decides how much thinking to do while charging for all of it. The organisations that prosper in this phase will not be those using the most AI. They will be those that treat machine work with the discipline they already apply to capital and labour: allocating it deliberately, measuring it honestly, and ensuring every unit of thinking earns its place.

References

WANT TO SPEND LESS AND GET MORE FROM AI?

Interested in deploying AI that is both effective and affordable? Get in touch. We help teams match the cost of AI to the value of the work through practical coaching, sensible model choices and a clear adoption plan that keeps spending tied to outcomes.