Cheaper tokens are creating a bigger AI bill

Falling model prices are making AI cheaper by the token, but not necessarily cheaper to run. As agentic systems drive consumption, enterprises need a new discipline for measuring whether usage is translating into business value.

For much of the generative AI boom, falling model prices have encouraged a reassuring assumption about the economics of the technology. As models improve, hardware becomes more efficient and competition intensifies, the cost of using artificial intelligence should keep falling. For individual requests, that is largely happening. For the businesses deploying AI at scale, however, the overall bill may be moving in the opposite direction.

Boston Consulting Group’s September 2026 report Token Cost Management: Three Pillars to Take Control of Token Cost argues that token consumption is increasing faster than token prices are declining. What looks like a steadily cheaper technology at the unit level can therefore become an increasingly significant corporate expense as more employees use AI, more workflows incorporate it and autonomous agents begin consuming computing resources without waiting for a person to initiate every interaction.

The implication extends beyond another round of IT cost optimisation. Enterprise AI is beginning to behave less like conventional software purchased through predictable licences and more like a metered resource whose cost depends on how intensively it is used. That changes the financial discipline required around deployment. The important question is no longer simply how much access costs, but what the organisation is paying for each useful business outcome.

When lower prices do not mean lower costs

Tokens are the units AI providers use to measure the information processed by a model. Every prompt consumes them, as does the context supplied to the model and the answer it generates. Improvements in models, infrastructure and computing have pushed the price of those tokens down, creating an apparently favourable cost curve for businesses expanding their use of AI.

BCG identifies growing employee adoption as one source of rising consumption, but the more significant shift comes as AI moves from personal assistants into operational workflows, where usage can expand far more quickly. A member of staff interacting with a chatbot creates a natural constraint on consumption because the system generally waits for another request before doing more work. An agent can behave very differently, decomposing an objective into individual tasks, invoking tools and models, evaluating results and repeating steps until it believes the work is complete. Each stage can create another round of token consumption.

Longer-running systems can also repeatedly process material that has already appeared earlier in the workflow. Instructions, retrieved documents, conversation history and tool outputs may all be carried forward as context. A business may therefore move from paying for an employee to make an occasional request to paying for a machine to conduct multiple interactions in pursuit of a single outcome.

This creates a different relationship between scale and cost. Traditional SaaS economics encouraged companies to think predominantly in terms of seats and subscriptions. Once software had been deployed, greater usage did not necessarily create a proportionate increase in the provider’s charge. AI inference introduces a more direct connection between activity and cost. For AI embedded in a customer-facing product, each additional interaction can carry its own processing expense, potentially making inference a factor in gross margin rather than merely another item in the technology budget.

The problem becomes harder when organisations lack visibility into what is driving consumption. A monthly AI bill might incorporate employees experimenting with models, production applications serving customers and agents completing thousands of intermediate steps. Treating all of that as a single infrastructure expense conceals whether the organisation is paying for productive work, experimentation or computational activity that could have been avoided.

The agent changes the economics

The rapid development of agentic systems makes that visibility increasingly important. An organisation controlling conventional software can normally make reasonable assumptions about the resources required to complete a transaction. Agent behaviour is less predictable because the amount of work undertaken may vary according to the task, the model selected, the context supplied and the number of attempts required before an acceptable result is produced.

Using the largest available model for every task can increase costs unnecessarily, but selecting the cheapest model does not automatically solve the problem. A less capable model that needs several attempts and greater human correction can ultimately cost more to achieve the required result. The economic unit that matters is therefore not the price of a token or even the cost of a request. It is the cost of producing an acceptable outcome.

BCG has developed this argument further in its work on what it calls “return on AI”. Rather than judging deployment principally by activity, organisations need to connect the economic return produced by a workflow with both its token consumption and the human effort required to operate, review and correct it. A coding agent that produces thousands of lines of software has generated considerable activity, for example, but that output has limited economic meaning if much of the code fails review or requires substantial reworking before it reaches production.

Similar questions will emerge wherever businesses introduce agents. A customer service system might process more enquiries while simultaneously creating more escalations for human staff. A marketing system might generate dramatically more material without increasing the amount that is used. Measuring the volume of work performed by AI can make deployment appear productive while obscuring the cost of turning that output into something the organisation values.

For finance leaders, this creates an unfamiliar governance problem. Restricting AI consumption too aggressively could suppress applications that generate genuine value, while allowing unrestricted access risks creating an expense whose relationship with business performance becomes increasingly difficult to explain. BCG’s argument is that controlling AI expenditure should therefore be about improving its economics rather than simply reducing usage.

Cost management moves into the architecture

The latest BCG analysis identifies potential savings of up to 50 per cent from three areas: smarter usage, smarter infrastructure and smarter contracting. It estimates that changes to usage could account for around 15 per cent, infrastructure improvements around 20 per cent and procurement measures another 15 per cent, although actual savings will inevitably depend on the organisation and its existing AI estate.

Some of those changes challenge the assumption that every intelligent workflow needs the most sophisticated model available. Routine lookups, deterministic decisions and conventional calculations may sometimes be handled more cheaply and reliably by existing software. Other workloads can be routed dynamically between models according to the complexity of the task rather than sending every request to a frontier system.

Infrastructure can also determine how efficiently tokens are consumed. Reusing previously processed context through caching, controlling the amount of information passed into models and setting limits around agent loops can reduce unnecessary processing. Cost management consequently becomes partly an architectural discipline. The way an AI application is designed may have as much influence over its economics as the headline price charged by the model provider.

Buying AI capacity on the expectation that prices will continue falling offers limited protection when consumption is simultaneously accelerating, which means procurement will need to adapt alongside the technology itself. Organisations need sufficient visibility into demand to understand what they are committing to purchase, while maintaining enough technical flexibility to shift workloads where appropriate. BCG argues that competition between providers and the ability to move workloads can strengthen the buyer’s negotiating position.

None of these measures matters greatly without attribution. Businesses first need to understand which users, applications and models are generating consumption and connect that activity with an owner and intended outcome. The same discipline increasingly needs to extend into the P&L. Tokens used to build reusable capability resemble investment; consumption supporting day-to-day internal work behaves more like operating expenditure; inference embedded inside a commercial product can become part of the cost of providing that product.

The emerging challenge is therefore not that AI has suddenly become prohibitively expensive. The greater risk is that organisations become successful at increasing adoption without becoming equally sophisticated at understanding its economics. Falling token prices can disguise that problem during the early stages of deployment because each individual interaction appears inexpensive. Scale changes the calculation.

As agents take on more work, businesses may find that they are no longer simply buying software for employees to use. They are purchasing machine intelligence continuously, in variable quantities and increasingly on behalf of other machines. Managing that consumption without constraining useful innovation will require a much clearer connection between architecture, finance and business performance.

The companies that gain the most from falling AI prices may therefore not be those that consume the greatest number of cheaper tokens. They will be the ones that can identify when additional consumption produces additional value and when the meter is simply running.