Model routing rule template
There are two kinds of AI model — fast/cheap and reasoning/thinking — and the most expensive mistake is using the wrong one for the job, in either direction. Use this one-page guide to map your tasks to the right tier before anyone routes by instinct.
The dependency test — use it for every task
Imagine handing the task to a sharp employee. Would they answer off the top of their head, or would they say "give me a few minutes" and go work it out?
- Answers off the top of their head → Fast model. The task is a lookup, a classification, a reformatting, a summary. No dependency chain. No deliberation needed.
- Has to go away and work it out → Reasoning model. The task requires holding several things in mind at once and reasoning through how they connect. Getting step one wrong throws off everything after.
Routing reference: task type → model tier
| Task type | Model tier | Why |
|---|---|---|
| Classify an email or ticket (complaint / question / other) | Fast / cheap | Single-step lookup; no dependency chain; a wrong answer is cheap to catch |
| Extract structured fields from a document | Fast / cheap | Pattern recognition; no multi-step reasoning required |
| Draft a routine reply, summary, or formatted output | Fast / cheap | Generation from given inputs; quality bar is adequate, not premium |
| Route or triage incoming work to the right queue | Fast / cheap | Classification; the whole point is speed and cost efficiency at high volume |
| Complex contract or regulatory document review | Reasoning / thinking | Must reason about how multiple clauses or rules interact; step-by-step dependency |
| Multi-step planning where step 3 depends on step 2's conclusion | Reasoning / thinking | Errors in early steps compound; deliberation earns its cost |
| Vendor or options analysis with multiple weighted factors | Reasoning / thinking | Requires holding trade-offs in mind simultaneously; fast model skips contradictions |
| Regulatory interpretation — how several rules combine for our situation | Reasoning / thinking | High stakes; answer depends on reasoning across connected constraints |
| Deep research — synthesize sources into a considered answer | Reasoning / thinking | Multi-source reasoning; cost trivial against value of a correct answer |
Your task routing map (fill in for your workflows)
| Our task (be specific) | Model tier assigned | Why (dependency test result) | Estimated daily volume |
|---|---|---|---|
Routing health check
- ☐ We are not routing all tasks to the reasoning model "to be safe" — capability we do not need is cost we cannot justify.
- ☐ We are not routing everything to the fast model to save money — on the tasks that genuinely require deliberation, a shallow answer is not a saving; it is a defect.
- ☐ The tasks we assigned to the reasoning tier genuinely require multi-step dependency. We can name what the dependencies are.
- ☐ Our fast-model tasks are classified, extracted, or drafted — not analyzed, compared, or interpreted across connected rules.
- ☐ We have estimated the per-task cost at production volume, not demo volume.
The rule of thumb
In practice, the genuinely hard tasks — the ones that earn reasoning-model prices — are roughly 10 to 20 percent of total volume. The other 80 to 90 percent are fast-model work. If your routing sends more than 20 percent to the premium tier, review each one with the dependency test before committing the budget.
Owner of this routing rule: ______
Last reviewed: ______
Next review: ______
Want a second set of eyes on this in your firm? The no-sell promise applies — if it isn't a fit, I'll tell you in the first ten minutes.
Book a 30-Minute Call →