UK businesses warned over rising generative AI costs

SAS has warned that businesses in the UK and Ireland face rising generative AI costs as they expand use of the technology. Research commissioned by the company found cost pressures are emerging most sharply among organisations that have moved beyond pilot projects.

The survey of 100 senior enterprise technology decision-makers found that 22% of organisations have fully integrated generative AI into regular processes, up from 9% in 2024. Within that more advanced group, 41% said prohibitive large language model costs were an active barrier to implementation, compared with 32% across the wider market.

Returns also appear uneven. Among organisations that have fully integrated generative AI, 45% said it had delivered a below-expected return on investment, the highest dissatisfaction rate recorded in the research.

At the same time, the findings suggest some businesses are seeing benefits. Nearly half of enterprises in the UK and Ireland actively using generative AI reported significant improvements in operating costs and time savings, while more than two in five said customer satisfaction had improved significantly.

Dr Iain Brown, Global Head of AI & Data Science at SAS, said the main issue for many businesses has shifted from proving the technology works to containing the cost of using it at scale.

"GenAI is genuinely working for organisations that have deployed it thoughtfully - we see that in the data and hear it from customers. But we're also seeing the first signs of a cost challenge that most businesses in the UK and Ireland aren't prepared for because they haven't moved out of pilots. Nearly half of the organisations that have moved beyond pilots are telling us clearly that costs at their current scale are becoming prohibitive. I'm hearing the same thing directly from enterprise customers we work with. The cost of AI tokens has fallen in recent times, but consumption at scale is more than wiping out any savings. They moved first, they scaled first, and they're feeling the cost challenges first because they didn't have the luxury of market proof. That's why governance and cost controls need to become part of GenAI strategies much earlier than many organisations anticipate," Brown said.

Cost pressures

The research points to a familiar software budgeting problem in a new form. While the price of individual AI interactions has fallen, spending can still rise if employees and systems use the tools more frequently.

That consumption-based model differs from many traditional software contracts, where costs are more fixed and easier to forecast. In generative AI, usage tends to increase with every new workflow, user group, or automated task added to a deployment.

Brown said that shift is changing the questions large organisations are asking suppliers and advisers.

"The organisations we work with that have moved GenAI into production aren't asking us whether it works anymore. They're asking how to stop costs from running away. That's a fundamental shift in the conversation, and it's happening faster than most anticipated. Organisations need to focus first on good governance of their GenAI and AI systems, then on controlling costs and treating GenAI as an enterprise capability before scaling," Brown said.

The survey also found that 38% of enterprises in the UK and Ireland already see difficulty proving return on investment as a barrier, even though most respondents have not yet reached full production use.

Agentic AI

Another pressure point is emerging in agentic AI, where systems handle multi-step tasks with less human intervention. SAS said 30% of organisations surveyed are already investigating or piloting agentic AI for tasks such as case handling, onboarding, and investigations.

These systems typically generate more AI calls than standard chatbot-style interactions. SAS cited Gartner analysis saying agentic AI models can require five to 30 times more tokens per task than standard generative AI, with a single workflow making 10 to 20 separate AI calls to complete one task.

That means businesses that based cost assumptions on earlier generative AI pilots may find real-world spending rises much faster once more autonomous systems are introduced. The warning comes as many large employers look for ways to embed AI more deeply in customer service, software development, and internal operations.

SAS also pointed to examples from large technology groups to illustrate the risk of uncontrolled usage. According to the company, Uber exhausted its 2026 AI coding tools budget in four months after rolling out tools including Claude Code across 5,000 engineers, with monthly API costs per engineer of between USD $500 and USD $2,000.

Amazon also ran an internal AI adoption leaderboard that encouraged engineers to use AI tools, SAS said. Employees then generated unnecessary AI workloads to improve their ranking, inflating costs without corresponding productivity gains, before the programme was shut down.

Similar behaviour, sometimes referred to as tokenmaxxing, has also been reported at Microsoft and Meta, according to SAS.

Governance focus

Brown said businesses should avoid using staff adoption alone as a measure of success and should put spending controls in place before broad deployment.

"Stop measuring GenAI success by how much your people are using it. Governance, budgets, and usage controls need to be in place before you scale. And critically, before any organisation moves from pilot to production, especially into agentic AI, they need to stress-test what costs might look like at full deployment. Good foundations and planning are key to generating strong returns from GenAI," Brown said.