Companies expanding their use of artificial intelligence are facing a growing cost-management problem as falling prices for individual AI interactions are offset by rapidly increasing consumption.
The challenge is becoming more pronounced as businesses move from standalone generative AI tools to agentic systems that use multiple AI agents to perform tasks, make decisions and interact with other software.
Each interaction consumes tokens, the units used by large language models to process prompts and generate responses. As AI systems handle more complex tasks, token consumption can increase in ways that are difficult for businesses to predict.
Simon Gooch of identity management company Saviynt said the changing economics make it difficult to offer customers fixed-cost commitments over periods of one, two or three years.
At the same time, the cost of individual tokens has fallen sharply. Goldman Sachs forecasts that monthly token consumption will rise 24-fold between 2026 and 2030, reaching 120 quadrillion tokens as businesses expand their use of AI agents.
Rising AI use makes costs harder to predict
Large language models break user prompts into tokens before processing them. They also generate responses as tokens before converting them into text, code or commands.
Token consumption can vary depending on the prompt, model and complexity of the task. Agentic systems can make the calculation even harder because several AI systems may work together to complete a single process.
Will Venters, associate professor of Digital Innovation and Information Systems at the London School of Economics, said companies can struggle to control spending when employees experiment with AI or deploy it across business operations.
The problem can extend beyond the main AI task. Companies may also use AI for software testing, security checks and safeguards designed to control AI systems.
Agentic technology adds another layer of uncertainty. Businesses can deploy additional AI agents more easily than they can expand a human workforce, potentially increasing usage without an equivalent physical expansion of operations.
Businesses look for ways to control spending
Some companies are already examining ways to make AI costs more predictable.
Oliver King-Smith, founder of engineering software company smartR AI, said smaller businesses may sometimes manage AI spending through flat-fee personal accounts. He suggested that model could become harder to sustain if major AI providers face greater pressure to improve profitability.
Choosing the right AI model for each task can also reduce unnecessary spending. Companies do not always need their most capable or expensive model for routine work.
Rob Steele, chief financial officer at UK accounting software company iplicit, said clearer instructions can also help control consumption. More precise prompts can reduce unnecessary processing and make AI systems more focused on the task.
However, reducing token use is not always the best financial strategy.
Higher consumption can produce better or more useful results. Companies therefore need to compare the additional cost of AI processing with the value generated by the resulting work.
AI pricing models remain unsettled
The uncertainty extends to companies that sell AI-powered products and services.
Bill Peterson, senior director of product marketing at Sumo Logic, said the cybersecurity software company was developing new services using agentic AI while discussing pricing options with corporate customers.
Companies could respond by raising prices across their products, charging customers according to results or creating bundled services based on specific outcomes such as security incidents.
Those models could still change as AI providers adjust their own pricing.
Variable pricing creates difficulties for corporate customers that need to prepare budgets months or years in advance. Frequent changes in AI infrastructure costs can make it harder for businesses to determine how much an AI-powered service will cost to operate over the life of a customer contract.
Falling prices do not guarantee lower AI spending
The economics of AI are therefore becoming more complicated as adoption increases.
Lower prices per token can make individual AI interactions cheaper. Yet businesses may use vastly more tokens as AI becomes embedded in software, customer services and autonomous workflows.
For companies building AI into their products, the central challenge is increasingly about managing variable consumption rather than simply reducing the price of individual AI interactions.
Businesses will need to measure token usage, select models according to the value of each task and develop pricing structures that can absorb changing infrastructure costs.
As agentic AI expands, the question is no longer simply how much an AI model costs to use. It is how companies can forecast that cost, control consumption and determine who ultimately pays for the additional computing required.
Reporting credit: Based on reporting and analysis from the sources cited in the supplied material, including comments from executives and researchers quoted in the report.
Editorial note: This Journos News rewrite preserves the supplied reporting and attribution. No additional claims have been introduced beyond the source material.
