Our last monthly bill from a hosted model provider arrived at $4,200, and almost half of it was spend we couldn't trace to a customer outcome. That single line item pushed us to rebuild how our internal agents think, route, and pay for intelligence. Three months in, we're running leaner, faster, and cheaper — and the playbook is simpler than the conference circuit makes it sound.
The first shift was admitting that not every task deserves a frontier model. Routing simple classification, formatting, and schema checks to a small local model handled the bulk of our volume without losing accuracy where it counted. We reserved the expensive hosted calls for the genuinely hard reasoning work — strategy framing, nuanced synthesis, and edge cases. The principle is unglamorous but powerful: match the cost of the call to the value of the answer.
The second shift was treating agent skills as code, not chat. Each skill is a small, testable function with explicit inputs and outputs. When a skill is well-defined, a tiny local model can execute it reliably, and we can cache the result. Over a week, those cached and locally-served calls are where the savings compound. The hosted model becomes a fallback, not the default.
Third, we measured tokens like a budget, not a vanity metric. We tracked cost per resolved task across every workflow and refused to ship anything that didn't beat a baseline. This forced honest conversations about which prompts were actually pulling their weight and which were paying premium prices for filler. A surprising amount of "intelligence" in our pipeline was repetitive boilerplate that a 7B parameter model handles comfortably.
Fourth, we gave every agent the ability to escalate deliberately. Local-first with a clear, documented path to a stronger model when confidence is low. The team trusts the system because the escape hatch is real, and the budget trusts the system because most calls never reach it.
Finally, we stopped optimizing for the demo and started optimizing for the Tuesday at 2pm. Real workflows, real latency budgets, real cost ceilings. That's where local models and disciplined skill design earn their keep.
If you're staring at a usage dashboard that doesn't reconcile with the value your team is shipping, the path forward isn't another contract negotiation. Start by mapping your highest-volume calls, route the simple ones locally, formalize the rest as reusable skills, and put a real cost-per-task number on every workflow. The 90% reduction isn't a marketing claim — it's what falls out when you stop paying frontier prices for chores.
Want a starter audit checklist we use to find the easy wins before committing to infrastructure? Subscribe and we'll send the one-page worksheet on Friday.