Corporate finance teams are rewriting the AI spending playbook. After two years of routing every query through top-tier models regardless of task complexity, CFOs and boards are pushing back against ballooning AI bills. The emerging solution? Model routing—a system design pattern that dispatches simple tasks to cheap, fast models and reserves expensive frontier models for complex work.

Scott Wu, CEO of Cognition, reports routine coding tasks see 5–10x cost improvements with sufficiently capable smaller models. Glean’s Arvind Jain estimates that 95% of enterprise AI usage still runs on premium models for work that cheaper alternatives handle easily. The math is stark: Cisco’s Jeetu Patel calculates $200 per employee per week in tokens—$10,000 annually per person. For a 90,000-employee company, that’s $900 million a year. Cisco, with 30,000 engineers using AI for product development, blew past its budget and now reallocates resources from other line items to token spend.

Routing introduces a critical architectural decision: where to place the classifier that determines query complexity. It can run as a lightweight model at the gateway, a rules engine, or an ensemble. The trade-off is latency versus cost—wrong routing wastes money or degrades user experience. Cognition backs its agent with a productivity guarantee, refunding up to $10 million if Devin’s saved engineering hours fall short. The metric shift is from token volume to output.

For OpenAI and Anthropic, the trend threatens their pricing model. As enterprises route high-volume, simple work to cheaper open-source or alternative models, frontier labs lose revenue on easy tasks. Patel argues this doesn’t kill demand for cutting-edge models but forces a pricing shift—vendors must optimize usage efficiency rather than simply charging more. The market’s center of gravity moves from premium model sellers to cost-conscious buyers. How much enterprise workload qualifies as “hard” will determine the valuation of AI’s leading labs.

Source: CNBC