Businesses are alarmed by the costs of AI - the CFO's return to the loop

In plain terms

After two years of FOMO, large companies are taking stock: their AI bills (OpenAI, Anthropic, Azure OpenAI, AWS Bedrock) exceed initial projections by several times. Result: the CFO takes back control, PoCs that don't pay off are killed, and models are put in competition line by line.

The affair

According to The Economist (June 2026), corporate AI spending continued to grow strongly in 2025-2026, but a majority of use cases do not pass the positive ROI threshold according to the CIOs surveyed. Three causes recur:

1. The token cost is not the real expense. The real cost is the human-in-the-loop: review, correction, orchestration, governance. Observed ratio in the field: typically 3 to 5× the raw cost of the model.
2. The context kills the margins. Moving from a short window to a long window (with RAG) multiplies the bill by an order of magnitude for the same use case.
3. The multi-model explodes the orchestration. A good practice - routing simple tasks to Haiku/Nano - is paid for in infrastructure complexity that few CIOs truly master.

The transition takes two forms: (a) partial repatriation to open-weight models hosted (Llama, Qwen, DeepSeek) - the self-hosted share is rapidly increasing in corporate AI workloads; (b) threshold contracts with providers - Microsoft now negotiates floor rates for its large accounts.

Under the hood

• Key metric: the "cost per successful task" - not the tokens billed, but the average cost of a deliverable that passes the review.
• Intelligent router: significant cost reduction (often 30-50%) when an upstream classification layer directs to a small model by default (LiteLLM, Portkey, Anthropic's own tiering).
• Cache prompt: underutilized - up to -75% on the repeated system prompt (Anthropic prompt cache).

So what

Three consequences. One: the phase "let the devs choose" is over, finance is back. Two: vendors who can demonstrate measurable ROI (not slides "productivity + 30%") take the prize. Three: it's the time for open-weight - not ideologically, economically. To watch: the inbound attrition rate of OpenAI on its 500 largest accounts at the end of 2026.