LOCAL SERVICE SPOTLIGHT · FIELD NOTES
Model Judgment: picking the right AI for each part of the job
Route each task to the cheapest AI that still clears the quality bar. Do the grunt work on a cheap, fast model. Keep the judgment calls on the best one. Let the system decide, every time.
The one idea: Stop asking “which AI is best?” Ask, for each piece of work, what is the cheapest tier that still clears the quality bar — and route it there. Cheap for the mechanical work, the best model for the judgment. That is how the Content Factory keeps your brand work both thorough and affordable.
Where this comes from
Two ideas a month apart that turn out to be mirror images.
Delegate down. The team behind Claude — and the developer Simon Willison, who wrote it up on July 3, 2026 — found it is better to let a top AI use its own judgment about how to work, including which cheaper model to hand a sub-task to, than to script it. His one-line instruction: “For all tasks, use your judgment to decide an appropriate lower-power model and run that in a subagent.”
Escalate up. Anthropic’s “Advisor Strategy” (April 9, 2026) is the inverse: a cheap model drives the whole task and calls in the top model only when it gets stuck. Their own numbers show a cheap model’s accuracy on the hard parts can more than double — from 19.7% to 41.2% — with a top-model advisor riding along, at a fraction of the cost of running the expensive one throughout.
The ladder — the whole thing, not just “the expensive one”
| Tier | The work | What runs it |
|---|---|---|
| 0 · mechanical | pulling data, publishing, formatting, file wrangling | plain scripts — no AI at all |
| 1 · bulk / first draft | gathering mentions, first-draft articles, routine pages | a fast, cheap model |
| 2 · judgment | the honest score, the real strategy, and writing in your own voice | the top-tier model |
| 3 · hands-on | logged-in publishing and anything behind a password | a person, or an agent on their computer |
Most of the work — and most of the cost — lives in tiers 0 and 1, and none of it needs the most expensive model. Only the thin judgment slice earns the ceiling.
The decision, in one breath
Does getting this wrong take taste, honesty, or your real voice?
No → push it down to a cheaper model.
Yes → keep it on the best model, and let the cheap tiers tap it on the shoulder when they get stuck.
Why it matters for your brand
A brand build is not one job — it is a stack of them. Pulling the data is mechanical. Drafting a first pass is bulk. But the honest read on how the world sees you, the strategy, and copy that sounds like a real person — that is judgment, and it is a sliver of the work. Run all of it on the priciest model and you burn the budget doing grunt work with a brain surgeon. Route by tier and you get roughly three times the output at a fraction of the cost, with no drop in quality where quality counts.
Worked examples
- Gathering every good mention across the web → tier 1. Cheap models sweep the sources; the top model only ranks and interprets the shortlist.
- A wave of articles for one topic → tier-1 drafting, tier-2 editing. Drafts on a cheap model; the one canonical page and the voice pass on the best model.
- Scoring a brand honestly → tier 2 only — and the honesty rules live in code, not a prompt, so no model can hand out a flattering number.
- Writing in your voice → tier 2, the top model. This is the one thing that reads as “AI” the moment you cut the corner. So we don’t.
The guardrails
- Check before it ships. A cheap model left to loop unwatched just automates its own mistakes.
- Honesty rules go in code, not prompts. The score bands run the same on every tier.
- Never cut corners on voice or the final call. Those are the crown jewels.
- One shared set of files underneath. Every tier reads and writes the same place, so models stay swappable and nothing is locked to one vendor.
Cheap where it is mechanical. The best model where it is judgment. Complete, always. That is the discipline behind the Content Factory.
See how we keep it fast
See the Content Factory
Method by Dennis Yu · Local Service Spotlight, as part of the Content Factory. Background reading: Simon Willison, “Fable’s judgement” (Jul 3, 2026); Anthropic, the Advisor Strategy (Apr 9, 2026).