Unexpected AI costs force enterprise project reviews, Mavvrik survey finds
A survey of 396 enterprises found surprise AI spending altered business decisions for 62% of respondents, exposing gaps in cost forecasting.
By Wei-Lin Zhao · AI Correspondent
· 3 min read
Unexpected AI costs enterprises encounter are changing project decisions, according to a survey commissioned by AI cost-management vendor Mavvrik. Of 396 enterprise organizations surveyed in April and May, 62% said a cost surprise materially changed a business decision, while one in four said it cancelled an AI initiative outright.
The results are not an industry census, and Mavvrik did not provide respondent names, detailed sampling methodology or audited spending data in the material reviewed. Still, the findings point to a problem for companies putting AI features and internal tools into production: a budget line for infrastructure does not necessarily produce a reliable view of what an AI workload costs.
Mavvrik said 95% of respondents assign formal AI budgets and 98% track AI infrastructure costs. Yet only 11% forecast AI spending within plus or minus 10%, while 89% miss that range. The company sells products for AI cost visibility and governance, giving it a commercial interest in the issue, so its results should be read as vendor-sponsored survey findings rather than a market-wide measure.
Why are unexpected AI costs hard for enterprises to forecast?
The reported gaps extend beyond large-language-model token charges. Mavvrik said 47% of respondents identified data-platform overages as their leading source of unforeseen AI costs, compared with 43% who named LLM token costs. Its reporting coverage figure was 67% for public cloud but 36% for agentic workflows and GPU infrastructure.
Developer tooling is another blind spot. Mavvrik reported that 98% of organizations use AI coding tools, but only 42% include those tools in AI cost reporting. The implication is operational rather than theoretical: finance and technology leaders can track part of the bill while missing costs accumulating in data systems, compute infrastructure and development workflows.
Agentic AI adds uncertainty as companies use systems that pursue tasks with greater autonomy. Gartner research chief Rita Sallam told CIO Dive that raw model-unit pricing may decline while the cost of completing a task rises as agentic workflows become more complex and require more advanced reasoning. Usage-based pricing can add to that expense, she said.
That makes a token-only budget inadequate for judging return on investment. Cost attribution means connecting spending to the product, customer, team, workflow, model or environment producing it. Without that connection, reported savings or revenue benefits cannot be tested against the full operating cost.
What should enterprise teams track?
The practical issue is coverage across the stack, including models, data platforms, cloud and GPU infrastructure, developer tools and agentic workloads. Mavvrik’s survey does not prove that cost surprises alone caused each delayed or cancelled initiative. It does show that respondents associate those surprises with budget actions: CIO Dive reported that about one-third imposed emergency spending freezes and nearly half escalated the issue to their boards.
Separate reporting points to the same broader pressure without offering directly comparable figures. AI Business reported that enterprises are wrestling with ROI, inference expenses and implementation choices as they adopt generative AI and consider agents. IBM has also reported that every executive in one of its surveys had cancelled or postponed at least one generative-AI initiative because of cost concerns. Different samples and methods mean those surveys should not be combined.
This story draws on original reporting from CIO Dive.
