Automatically sending each AI request to the cheapest model that can handle it well, instead of running everything through the most expensive option.
Also known as LLM routing, intelligent routing
Model routing is a layer that sits between a request and the AI models available to handle it, deciding in real time which model actually gets the job. A simple task, cleaning up a headline, formatting a table, might go to a small, cheap model. A complex one, reasoning through a media plan with conflicting constraints, gets routed to a frontier model that costs far more per query.
The economics are the whole point. Premium models can run 15 to 30 times the cost per token of lightweight ones, and most of what any team asks an AI tool to do does not actually need frontier-level reasoning. Teams that tune a routing layer well have reported cutting AI spend by 40 to 85 percent without a noticeable drop in quality, because most requests were overpaying for capability they did not use.
AI tool costs scale with usage in a way media budgets do not always anticipate. An agency running AI across dozens of accounts, for copy variants, image generation, reporting summaries, can see costs climb fast if every request defaults to the most expensive model available.
Not every task deserves the same model. Drafting ten headline variations does not need the same horsepower as untangling a client’s attribution data across five platforms. Routing formalizes a distinction most teams already make informally when they choose which tool to open for which job.
Badly tuned routing creates its own problem. Routing too aggressively toward cheap models to save money can degrade output in ways that are not obvious until a client notices, so the savings need to be checked against actual quality, not just the invoice.
An agency’s ops lead notices the AI tooling bill has tripled in two quarters as more teams adopt AI for daily work, from quick social captions to full campaign strategy documents. Instead of restricting access, she sets up a routing layer: routine formatting, tagging, and first-draft copy tasks go to a lightweight model by default, while anything flagged as strategy, client-facing, or final-draft gets routed to the premium model automatically. Within a month the bill drops by more than half, and when a copywriter complains that first-draft quality feels slightly flatter than before, the team adjusts the routing threshold for creative tasks specifically rather than rolling the whole system back.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.