The rising cost of using large language models (LLMs) is now giving enterprises pause for thought. As artificial intelligence (AI) models have become more sophisticated, queries are costing businesses significantly more in “tokens” and, in some cases, ratcheting up disastrously large bills.
And their inherent variability means it is no longer possible to predict how much each task will cost. The same prompt one day could produce an instant response, but an almost identical prompt on another day could take five minutes and burn through 10% of your monthly token budget, says Don Schuerman, chief technology officer (CTO) at Pegasystems.
Enterprises have incentivised employees to maximise their use of AI without fully considering the benefits that it produces for the organisation. Such “tokenmaxxing” has left companies with unexpected bills.
Last month, for example, Axios reported that a single unnamed enterprise spent over £500m on AI tokens for Anthropic’s Claud AI platform in one month after failing to put a cap on its employees’ IT use.
And in April, Uber’s CTO disclosed that the minicab and delivery service had burned through its entire budget for Claude Code for 2026 in the first three months of the year.
AI token cost rising
Commercial pressures to fund AI datacentres and increasing energy costs have led AI suppliers to raise their prices in recent months, leading organisations to question where the value of spending on AI lies.
In some cases, Schuerman says companies have replaced people with AI only to realise that AI is costing them more than the people they replaced.
Last week, Mary Daly, president and CEO of the Federal Reserve Bank of San Francisco, put it succinctly in an interview with Bloomberg. “Productivity growth is everywhere except in the data,” she said.
“What’s happened is the models have gotten more sophisticated,” Schuerman tells Computer Weekly. “The model reasons with itself, sometimes it dispatches other agents to do other things, and as it does that, it’s continuously running the token meter.”
Enterprises waking up to AI costs
He argues that enterprises are waking up to the fact that the cost of AI does not increase linearly with the number of calculations the model makes. “Every step is adding quadratically to the cost of the process,” he says.
If you use AI to work out a business process, the first step might take 500 tokens, but the context of the first step will need to be carried over, so the second step will require 1,000 tokens. The third step will use 1,500 tokens, and so on, says Schuerman.
As calculations become more complex and require more context, not only does the cost increase, but the risk of AI hallucinating or behaving unpredictably also increases.
“The best possible use of AI is to help me get that repeatable process right – help me define it, help me design it, help me ensure it follows best practices. And it turns out I don’t need much AI for that”
Don Schuerman, Pegasystems
Pegasystems’ answer to keeping the costs of AI under control is to use the technology in a more strategic way.
Pega supplies Fortune 500 companies with low-code platforms to automate their business processes and manage relationships with their customers.
This week, the company announced that it would charge its customers for business outcomes, rather than charging them for how many AI tokens they use. Schuerman argues that at least 60% to 70% of the high-volume mission-critical processes enterprises need can be automated using rules-based approaches.
“The best possible use of AI is to help me get that repeatable process right – help me define it, help me design it, help me ensure it follows best practices,” he says. “And it turns out I don’t need much AI for that.”
Designing workflows
Pega’s Blueprint software, for example, uses AI to help people design automated workflows for their organisations. Because AI does the bulk of its work at the point of design, AI agents don’t have to rethink the process from scratch each time the process runs.
The workflows can, however, call AI agents to execute specific tasks – summarising a document or seeking input from a human, for example – giving enterprises the ability to use AI reasoning in a controlled way.
Pega has now made its automated workflows compatible with the open source Model Context Protocol (MCP). This means companies can give AI agents built on other platforms – such as Anthropic Claude, Google Gemini, OpenAI, AWS AgentCore, and others – access to Pega’s processes.
Schuerman says that for Pega, adding MCP was a “relatively light lift”. Pega is agnostic about how organisations access its workflows. Organisations can already access Pega workflows through other software, such as Salesforce, for example, without having to use Pega as a front end.
But for businesses, the change is huge, he says. They can take their existing AI agents and use them to follow Pega’s workflows.
“Once I’ve connected in the Pega MCP, that agent is going to follow the rules, it’s going to do it without excessive reasoning [by making] that relatively simple call to find the right workflow and finding the skill,” says Schuerman.
From banks to pizza
Companies, including the Dutch bank Rabobank, are using MCP calls to convert chatbots into intelligent agents that complete tasks for customers, such as checking an account balance or potentially making a payment.
Blueprint has opened the door for more companies to become Pega customers, more quickly. In the past, it might have taken several years for Pega to develop prototype workflows and for companies to decide whether to use Pega software.
“It often took a long set of conversations and really skilled salespeople to help the customer understand how their business problem fit into our platform,” he says.
With Blueprint, the software can take a description of what they do as a business, scan material from the company website, and within minutes design workflows that make sense to the business.
Schuerman won’t put any numbers on it, but says Pega has made a conscious decision to sell its platform to a broader range of companies, partly directly and partly working through business partners.
Previously, Pega specialised in supplying its platforms to highly regulated industries such as banks, insurance companies and healthcare. But now it is seeing opportunities from companies like Papa John’s, the pizza chain, which was seen among the attendees of Pega’s annual conference.
AI can’t create the extraordinary
Schuerman is adamant that AI is not going to replace human creativity as enterprises become more automated.
Unlike people, AI does not create things that are extraordinary or show genuine creativity. “It just averages everything else that it has read in the past,” he says. “What we want to do is continually compress the time from an idea to that idea making a meaningful change in how my employees work and how customers engage with us.”
That leaves people free to focus more on strategy, ideas and creativity. There is too much focus on the next version of ChatGPT or Claude, when last year’s version was just fine, according to Schuerman.
“The hard work is in turning all this AI potential into realities for people, and that’s where I think the interesting work is,” he says.
