Tokens to usage: Why cheaper AI still leaves enterprises with bigger bills

Token prices have fallen sharply, but agentic workflows can consume far more of them, while data, infrastructure and governance add to costs, shifting the focus from usage to business returns

Artificial intelligence (AI) model prices are falling sharply, but that is not necessarily making the technology cheaper for businesses. The price of a token, a unit used by large language models (LLMs) to process and generate text, has fallen drastically since early 2024. Yet enterprise AI bills continue to rise as companies use the technology more widely and move from basic chatbots to AI agents.

 

The discussion is therefore shifting from the price of individual tokens to how many tokens companies consume, what systems they build around AI and whether that spending produces measurable business results.

Why cheaper AI tokens are not lowering enterprise bills

The economics of AI are changing as companies move from experimenting with the technology to using it in everyday business operations. A chatbot may require only a few interactions to answer a question, while an AI agent can break the same request into several steps, including planning, information retrieval, tool calls and verification.

 

 

Stanford University's 2026 AI Index found that organisational AI adoption reached 88 per cent in 2025, while generative AI was being used in at least one business function at 70 per cent of organisations.

 

AI agent deployment, however, remained in the single digits across nearly all business functions, leaving substantial scope for usage to increase further.

 

Agentic AI drives up token consumption

 

The shift from chatbots to AI agents is increasing token consumption because agents do more than generate a single response. They can retrieve information, use tools, check results and repeat steps until a task is completed.

 

Moving from a chatbot to an agent can increase token consumption for a task by one or two orders of magnitude. In agentic workloads, the final output can account for only 5-15 per cent of the tokens consumed, with the rest going towards processing context.

 

This context can include information retrieved from company databases, previous steps in a workflow and documents supplied to the model.

 

PDFs are one example. They are designed primarily to display information rather than make it easy for AI systems to interpret. An AI model may therefore need to process a document's structure and layout before extracting the information it requires.

 

As companies use AI across more business functions and models perform deeper reasoning, token consumption can increase further.

 

Research into agentic coding has found a similar pattern. A 2026 study by McKinsey stated that multi-turn agentic workflows consume up to 1,000 times more tokens than standard chat or single-turn code reasoning.

 

Model costs are only part of the enterprise AI bill

 

For enterprises, the cost of using an AI model is only one part of the overall bill. Companies also need to prepare data, connect AI systems with existing software, monitor usage, manage security and put governance mechanisms in place.

 

The API or token bill therefore does not reflect the full cost of enterprise AI.

 

Businesses also need to account for infrastructure, storage, integration, security, monitoring and governance.

 

Data preparation can add further costs when information scattered across systems needs to be cleaned and transformed before an AI model can use it.

 

This engineering work may not appear on the model provider's invoice, but it remains part of the cost of deploying AI.

 

How should companies calculate AI costs

 

McKinsey's July 2026 analysis argues that per-token pricing has "stopped being a useful measure" of what enterprises actually pay for generative AI. Its research identifies long-lived context, repeated refinement and the broader architecture around AI agents as important drivers of costs.

 

For businesses, therefore, the cost of deploying AI should be calculated beyond the headline price per million tokens.

 

The first step is to measure the total number of model calls involved in completing a task rather than looking only at the price of one interaction.

 

Companies also need to measure how much context is being sent to the model, how frequently information is retrieved and how many times an agent repeats or verifies a task.

 

The second step is to include costs surrounding the model. These can include data preparation, storage, infrastructure, integration with existing systems, security, monitoring and governance.

 

A May 2026 McKinsey Enterprise AI FinOps Survey found that 93 per cent of respondents had exceeded their AI budgets. One-fifth of respondents to its 2026 global State of AI survey also said their organisations had limited AI use because of AI-related operating costs.

 

Is higher AI spending producing better results

 

The bigger question for companies is whether rising AI expenditure is producing measurable gains in productivity, revenue or other business outcomes.

 

Stanford's 2026 AI Index shows that AI adoption is widespread, but agent deployment remains relatively early. This suggests that many companies are still working out how to translate adoption into large-scale business results.

 

PwC's 2026 AI Performance Study provides a similar picture. Based on a survey of 1,217 senior executives across 25 sectors, the study found that the top 20 per cent of companies captured 74 per cent of AI's economic value.

 

These companies were more likely to use AI to pursue growth opportunities and redesign workflows rather than simply add AI tools to existing processes.

 

PwC's 2026 Global CEO Survey found that only 12 per cent of chief executive officers said AI had delivered both cost and revenue benefits. Another 33 per cent reported gains in either cost or revenue, while 56 per cent said they had not seen significant financial benefits from AI so far.

 

The findings suggest that simply increasing AI usage does not guarantee better financial results. 

 

From cheaper AI to better AI economics

 

The sharp decline in token prices remains important because it makes AI models cheaper to access. But for enterprises, the falling price of an individual token is only one part of the equation.

 

As AI moves from chatbots to agents, systems can consume far more tokens through reasoning, retrieval, tool calls and repeated processing. At the same time, businesses have to pay for the data, infrastructure, integration, governance and monitoring needed to make those systems work.

 

The calculation for enterprise AI is therefore shifting: not how cheap a token is, but how much a complete AI workflow costs and what the business receives in return.