The wrong tool for the job costs more than the right one

The default is maximum

Most AI coding setups share a common configuration: one model, one context window, every task. It is simple. It requires no judgment at the routing layer. It also means that renaming a variable and redesigning an authentication system go through the same pipeline.

The cost difference between those two tasks is significant. The routing overhead is zero because there is no routing. Everything runs at maximum capacity, regardless of whether the task needs it.

At low volume this is invisible. At the task frequency most active developers run, it becomes a meaningful line item with no corresponding output quality improvement for the low-risk tasks.

What routing actually reads

Avorelo routes tasks based on observable signals from the instruction, the file boundary, and the prior session receipt. It does not ask the developer to classify the task manually. It reads intent from the task structure.

Task type	Risk signals	Weight
Rename, reformat, extract constant	Narrow scope, no logic change, no cross-boundary writes	Low
Add test, update docs, fix typo in message	Additive, isolated, reversible	Low
Refactor function, change interface	Cross-file writes possible, logic impact moderate	Medium
New feature, auth change, schema migration	Cross-boundary, high logic complexity, proof required	High

The wrong turn problem

Routing errors are not always about cost. They are also about quality in the opposite direction: sending a high-risk task to a route that lacks the context depth to handle it correctly.

Over-routing and under-routing both generate rework. The over-routed low-risk task completes expensively and correctly. The under-routed high-risk task may produce a plausible result that does not hold up under review.

What changes when routing is automatic

When developers do not have to classify tasks, two things happen. First, the classification actually occurs on every task instead of only when someone remembers to do it. Second, the classification is based on the task structure, not the developer's current mental model of what the task involves.

That second point matters more than it appears. Developers routinely underestimate scope at the time they issue a task. Avorelo reads what the task is going to touch, not what the developer thinks it will touch.

The outcome

Teams that route by intent rather than by default see AI spend become more predictable. Low-risk tasks run efficiently. High-risk tasks get the depth they require. The ratio between spend and output quality improves across the full distribution of work, not just the high-complexity tail.