Among UK and Irish businesses that have moved beyond pilots and integrated generative AI into regular operations, 41% cite prohibitive LLM costs as an active barrier, according to new research from SAS.
The study, commissioned by SAS and carried out by Coleman Parkes across 100 senior enterprise technology decision-makers in May 2026, finds that 22% of UKI enterprises have now fully integrated GenAI into regular processes, up from 9% in 2024. Within that more advanced group, 45% say GenAI has delivered below-expected ROI — the highest dissatisfaction rate recorded in the research.
The picture is not uniformly negative. Nearly half of UKI enterprises actively using GenAI report significant improvements in operational costs and time savings, and more than two in five say customer satisfaction has significantly improved. But the financial pressure at scale is real: enterprise GenAI spending continues to rise even as per-query costs fall, because usage grows faster than unit prices decline. Unlike traditional enterprise software, where costs are relatively predictable, GenAI charges scale directly with consumption.
Dr Iain Brown, Global Head of AI and Data Science at SAS, points to examples already in public view. Uber consumed its entire 2026 AI coding tools budget within four months of rolling out tools including Claude Code to its 5,000-strong engineering team, with monthly API costs per engineer running between $500 and $2,000. Amazon shut down an internal AI adoption leaderboard after engineers began generating unnecessary AI workloads to improve their rankings, inflating costs without productivity gains. The same pattern, referred to as "tokenmaxxing", has been reported at Microsoft and Meta.
The problem is set to intensify as agentic AI moves out of the lab. The SAS research finds 30% of all UKI enterprises are already investigating or piloting agentic systems — tools that handle multi-step tasks without human prompting at each stage. Gartner analysis indicates agentic models require 5 to 30 times more tokens per task than standard generative AI. Where a standard chatbot query generates a single round of token consumption, an agentic workflow may make 10 to 20 separate AI calls to complete one task.
Dr Brown argues that governance and cost controls need to be built into GenAI strategies before organisations scale, not as an afterthought once budgets are already strained.
To stay across the latest in cloud, AI and enterprise tech analysis from Compare the Cloud, subscribe to our weekly newsletter at https://www.comparethecloud.net/newsletter
