My default AI routing policy is still simple:
Cheap models draft. Expensive models judge.
Not because you are stingy for sport. Because most of the tokens in a workflow are rough clay, and clay does not need marble prices.
What counts as a draft job
- first passes of copy
- summarizing a long thread into bullets
- extracting fields into a table
- proposing three options you will discard two of
- turning a checklist into a template
- code suggestions you will still read line by line
These jobs benefit from speed and cost control. If the output is wrong, I can tell quickly. That is the key: cheap drafts must be easy to evaluate.
What counts as judgment
- final customer-facing wording in sensitive contexts
- architecture choices with long blast radius
- tricky debugging when the failure mode is expensive
- anything involving policy, compliance nuance, or ambiguous tradeoffs
- synthesizing conflicting sources into a decision you will own
Judgment jobs are where I pay for a stronger model, a slower think, or a second pass with a cleaner brief.
Why this beats “always use the smartest”
Always-smart is lazy routing. It feels premium. It trains you not to notice which tasks were easy.
It also blows budgets in quiet ways. A team can burn serious money on summarization that a smaller model could do fine, then get weirdly cheap at the exact moment a hard decision shows up.
Aim for the opposite: frugal on volume, generous on consequence.
A practical router
Use labels in prompts and in code:
- `tier=draft`
- `tier=judge`
- `tier=extract`
Extract is sometimes even cheaper than draft. Structured extraction with a schema often does not need poetry.
The router is allowed to escalate. Draft first. If confidence is low or validation fails, escalate once. Do not escalate in a loop forever. That is how you recreate always-smart with extra steps.
Quality control still sits with me
Cheap drafts fail differently than expensive ones. They miss nuance. They get generic. They invent plausible filler.
So keep validators:
- schema checks
- forbidden phrase lists
- “cite the source sentence” requirements
- human review for external sends
The model tier does not replace QA. It just changes the cost of getting to QA.
Story from the trenches
I once ran a content pipeline entirely on a frontier model because it “sounded better.” It did sound better. It also cost enough that Start skipping runs.
I moved outlines and first drafts to a smaller model, reserved the big model for final polish on pieces that mattered, and suddenly the pipeline ran daily again. More iterations beat occasional perfection.
Ops truth: a system you avoid because of cost is not a system.
Checklist for your own routing
- List your top ten AI jobs
- Mark each as draft, extract, or judge
- Put cost caps on draft volume
- Require an eval snippet for judge-tier changes
- Review the split monthly as models and prices move
Closing
Cheap drafts, expensive judgment is not a slogan about models being dumb. It is a slogan about matching spend to irreversible decisions.
Most tokens are reversible. Spend like it. When the decision is not reversible, pay for the better brain and give it a better brief.
That default has survived a lot of model launches. It will probably survive the next ones too.
Teach the team the split
If only you route this way, spend leaks through other people’s defaults.
Put the routing policy in the internal doc. Show examples of draft-tier versus judge-tier prompts. Celebrate cost saves that did not hurt quality. People copy what gets praised.
Measure both quality and spend
Track cost per successful task alongside quality notes. If draft-tier quality collapses, escalate the model. If judge-tier spend balloons on easy chores, push work back down.
Bottom line
Without both meters, you will optimize the wrong one and call it taste.
