AI spending is climbing faster than the returns it produces, and finance chiefs are noticing. Deloitte's survey of executives in Europe and the Middle East found that while average tech investments pay off in 7 to 12 months, a typical AI use case takes 2 to 4 years. That gap puts many projects at risk of cancellation.
Token costs themselves are falling fast - adjusted for performance, LLM inference costs drop about 10x a year - but usage is rising faster. Agents consume far more tokens than chat interfaces: single agents about 4x more, and multi-agent systems roughly 15x, according to Anthropic. Experts expected average monthly AI spend to rise 36% in 2025, with no sign of slowing in 2026. A DoiT survey of 500 finance leaders at large US and UK enterprises found only 15% of companies said they could calculate AI ROI easily, while 83% expected clear, quantifiable results within a year.
The dbt Labs team points to Model Context Protocol (MCP) as a major cost driver. Loading every tool definition upfront is expensive: Anthropic measured a five-server setup with 58 tools at about 55,000 tokens before a conversation even starts - at 10,000 agent calls a month, that's 550 million input tokens of overhead, or $2,200 a month on Claude Opus 5.5. Raw payloads can also pass through the model twice, and the same context gets pulled again and again with no reuse.
Their fix was to build a permanent context layer in the data warehouse, so retrieval and data shaping happen once and get reused across invocations. Agents query one governed warehouse layer instead of 20+ MCP servers, cutting upfront token usage by as much as 85% by Anthropic's data. Using dbt to chunk, embed and classify transcripts means only relevant passages go to the model - Anthropic cut one agent workflow from 150,000 tokens to 2,000. Storing key elements of customer calls from Gong also avoids paying about $3,000 each time AI reads the full dataset. The team says the approach shaved more than 50% off its token costs.
For CFOs, the advice is to establish baseline spend, find MCP sources that overlap with warehouse data, start with the most expensive context-rich sources, process unstructured data into embedded chunks with dbt, and measure before-and-after costs.
Source: dbt Labs Blog