Pick the model per step: fast to sort, strong to judge
Why I use different Claude models for different steps of a support workflow: a fast model to classify, a stronger one for hard escalations, and spend limits.
The first version of any AI workflow tends to use one model for everything. It works, then the volume grows and the bill or the latency starts to hurt. The fix is to treat each step as its own job.
One workflow, three jobs
01Classifyevery message, needs speed
02Summarisefor the team channel, needs clarity
03Judgeonly hard escalations, needs depth
Matching the model to the job
The bars show roughly how much of the traffic reaches each step. Most messages only need the first, so that is where a fast, cheap model pays off. The expensive model only sees the small slice that needs it.
Guardrails that come first
Spend limits per workspacea runaway loop cannot surprise anyone
A timeout on every calla slow answer falls back to the review queue
Retries with a capthen the message goes to a person
Logged inputs and outputsso the daily QA sample has something to read
This is the same layered idea as the quality gate: the cheap, automatic path handles the clear cases, and the expensive attention goes where it matters.