Revana, the lead intelligence agent at estaie, reads thousands of signals to find a handful of real buyers. Sending every signal to a frontier model would cost more than the leads are worth. Instead it runs a cascade, and the cost per qualified lead dropped ten times with better precision.
The three tiers
| Tier | What runs | Share of traffic | Cost |
|---|---|---|---|
| 1 | Embedding similarity filter | 100% in, ~28% out | Near zero |
| 2 | Small classifier or small model | 28% in, ~9% out | Cents |
| 3 | Frontier model with tools | 9% in | Real money, worth it |
Rules for a good cascade
- Each tier must be allowed to say "not sure" and pass the item up, never down.
- Measure precision and recall at every tier with the eval suite from lesson 4. A cheap filter that drops good leads is not cheap.
- Log what each tier dropped. Sample the drops weekly. That is where your false negatives hide.
Where cascades apply
Support triage, lead qualification, content moderation, document routing, search re-ranking. Any flow where most inputs are obviously irrelevant and a few deserve real thought.
Do this today: look at your agent's last 1,000 calls. What fraction needed a frontier model to decide? If it is under half, you have a cascade to build.