The Hidden Cost of AI Agents: Why Companies Can't Predict Their Future AI Bills
The cost of using artificial intelligence models has fallen sharply in recent years, but businesses are facing a new challenge: figuring out how much they will actually spend on AI.
As companies move beyond simple chatbots and begin deploying AI agents that can write code, analyse data, handle security tasks and automate business processes, the number of tokens consumed by these systems is rising rapidly. That is making AI pricing increasingly difficult to predict—and potentially creating a major headache for companies trying to budget for the technology.
AI may be cheap for consumers, but businesses face a different reality
For individual users, free versions of services such as ChatGPT, Claude and Gemini can appear to offer extraordinary value. Behind those services, however, companies including Microsoft, Google and Anthropic have invested enormous sums in developing large language models (LLMs).
AI companies therefore offer paid plans with additional capabilities, particularly for areas such as coding, business automation and enterprise applications. At the same time, a growing ecosystem of third-party companies is developing specialised AI-agent services on top of these models.
The problem is that pricing AI services is far more complicated than pricing traditional software.
Simon Gooch of identity management company Saviynt says businesses cannot easily commit customers to a fixed AI cost over several years because the economics of AI are changing so quickly.
Why tokens are at the centre of the AI pricing problem
Large language models process information through units known as tokens. A user's prompt is converted into tokens before being processed by the model, while the model's response is also generated as tokens before being converted back into text, code or instructions.
The amount of computing required can vary considerably.
A small change in a prompt can generate a different response. Different AI models can consume different numbers of tokens for similar tasks, while complex AI-agent systems can involve several models working together.
That makes predicting future AI expenditure particularly difficult.
And although the price of individual tokens has fallen dramatically, the volume of tokens being consumed is rising at an extraordinary rate.
AI agents could drive token consumption 24 times higher
According to Goldman Sachs research, monthly token consumption could increase 24 times between 2026 and 2030, reaching around 120 quadrillion tokens per month as businesses increasingly adopt AI agents.
That creates an unusual economic situation: the cost of each individual unit is becoming cheaper, but companies are using vastly more of those units.
For businesses, the result could be an AI bill that is difficult to forecast.
Many organisations also have limited visibility into how many tokens their employees or automated systems are consuming until they reach usage limits or receive their monthly invoice.
Companies are already struggling with AI spending
Some businesses have already encountered unexpected AI costs.
Microsoft has reportedly reduced engineers' use of certain third-party AI coding tools because of their expense. Uber, meanwhile, reportedly consumed an annual AI coding-token budget within just a few months earlier this year.
Will Venters, associate professor of Digital Innovation and Information Systems at the London School of Economics, says businesses can struggle to manage AI spending because the output—and therefore the amount of computing required—is not always predictable.
The challenge becomes even greater when AI is introduced across an entire organisation.
AI agents make the problem even harder
Traditional software typically has relatively predictable infrastructure costs. AI agents work differently.
A company might initially deploy one AI agent for a specific task. It can then add more agents almost instantly, potentially increasing token consumption dramatically.
Businesses may also discover that they need AI not only for their primary task but for testing, cybersecurity, monitoring, quality control and safety guardrails.
That can turn a seemingly affordable AI deployment into a much larger expense.
Venters notes that increasing the number of AI agents can be as simple as clicking a button, whereas expanding a human workforce normally requires hiring plans, approvals and additional management.
Companies may need to become smarter about AI models
Businesses are increasingly looking at ways to control their AI expenditure.
Oliver King-Smith, founder of engineering software company smartR AI, says smaller companies can sometimes use flat-fee personal AI accounts, although he believes that approach is unlikely to remain viable indefinitely as major AI providers focus more heavily on profitability.
Companies could also reduce costs by choosing AI models more carefully and using cheaper models for simpler tasks.
Another important factor is prompt quality.
Rob Steele, CFO of UK accounting software company iplicit, argues that companies need to give AI systems precise instructions. Poorly designed prompts can lead to unnecessary processing and higher token consumption.
The bigger question: Who will ultimately pay for AI?
For companies building AI-powered products, controlling internal costs is only half of the problem.
They also need to decide how those costs will be passed on to customers.
Bill Peterson, senior director of product marketing at Sumo Logic, says there is still no clear industry standard for pricing agentic AI services.
Possible approaches include:
Increasing the overall subscription price
Charging customers according to usage
Charging based on results
Selling bundles of AI-powered services or incidents
Introducing separate pricing tiers for AI features
But each approach has drawbacks.
A usage-based model could expose customers to unpredictable bills, while a flat subscription could leave the software company carrying the risk if AI usage suddenly increases.
AI pricing could become a major business issue
The biggest uncertainty is that AI providers themselves can change their pricing.
If the cost of accessing an underlying large language model changes every few months, companies building products around that model may find it difficult to maintain predictable pricing for their own customers.
That creates a difficult balancing act.
AI is becoming cheaper per token, but businesses are using exponentially more tokens.
For companies adopting AI agents at scale, the key question may therefore no longer be simply “How much does AI cost?” but rather “How much AI will we actually use?”
As AI agents become more capable and autonomous, that distinction could become one of the most important factors shaping the economics of the AI industry through the rest of the decade.