The problem
Corporate extended-stay demand hides in thousands of weak signals: relocations, project announcements, hiring surges, event calendars. Reading all of them with a frontier model would cost more than the leads are worth. Reading none of them means waiting for inbound.
What we built
Revana is a lead intelligence agent that detects buying intent, enriches and qualifies prospects, and carries outreach through to close. The economics work because of a cascade: cheap models do the sorting, and the expensive model only sees what survives.
| Tier | What runs | Passes through |
|---|---|---|
| 1 | Embedding similarity filter against known buyer patterns | About 28% |
| 2 | Lightweight classifier on intent and fit | About 9% of the original |
| 3 | Frontier model with research tools, writes the outreach | Only what matters |
The trust layer
- No outreach to existing customers or active conversations, checked in code against the CRM.
- Claims and tone policy enforced on every message before it is queued.
- Daily send caps per segment and per agent.
- Two human moments: one to approve a new segment, one before a deal closes. Everything else is autonomous and logged.
Results
Cost per qualified lead fell about ten times compared with running a single frontier model over every signal, with better precision because each tier is measured on its own false negatives. A live qualified pipeline existed within 90 days of launch. Revana is now being productised for other businesses.
What this means for you
If your AI bill is growing faster than your results, you probably have a cascade to build. Most inputs are obviously irrelevant; a few deserve real thought. Pay frontier prices only for the few.