WaterfallReach: An AI Agent for Cost-Aware Contact Enrichment via a Five-Tier Cascade

1.Introduction

The standard approach to contact enrichment couples a vendor API to a CRM pipeline and bills per row regardless of resolution path. We observe that a non-trivial fraction of enrichments are recoverable from public-records signals at zero marginal cost — yet vendors do not expose this option. We propose a waterfall cascade that performs cost-aware tier selection, with an AI agent acting as both signal-fuser and exit-criterion judge.

2.Method

The cascade comprises five tiers T₀…T₄, ordered by ascending marginal cost. Each tier produces a candidate set and a confidence distribution; the AI judge fuses these and decides whether to exit the cascade or descend to the next tier.

  ┌──────────────┐    seed     ┌────────────────────────┐
  │  Input row   │ ──────────→ │  T₀ Free signals (20+)  │
  └──────────────┘             └────────────┬────────────┘
                                            │
                              candidate · confidence c
                                            │
                       ┌── exit if c ≥ τ ───┴───────────────┐
                       ▼                                    │
                  ┌─────────┐         else descend           │
                  │ verify  │                                ▼
                  └─────────┘                       ┌─────────────────┐
                                                    │  T₁ … T₄        │
                                                    └─────────────────┘

2.1 The AI Judge

We use an LLM as a judge over heterogeneous signals (permutation candidates, IRS Form 990 disclosures, CommonCrawl mentions, MX fingerprints, GitHub commits, NPPES taxonomy, certificate transparency logs, archive.org snapshots, and others). The judge is explicitly not tasked with generating contacts; it ranks and arbitrates only. Prompt caching keeps cache-read ratio above 90%, materially reducing inference cost.

2.2 Domain Memory

On successful verification, the agent persists the resolved pattern (e.g., first.last@domain) to a per-domain memory layer. On future invocations matching the same domain, the cascade short-circuits past Tier 2 onward. This monotonically reduces blended cost as workload size increases.

2.3 Adversary-Aware SMTP

Tier 1 (SMTP verification) is sensitive to security gateways (Proofpoint, Mimecast, Barracuda, Cisco) that produce false-positive 250 responses for unknown mailboxes. We mitigate via MX fingerprinting and a catch-all probe that walks the MX priority chain.

3.Compliance Posture

Each emitted email carries a four-class source-risk label {T0, T1, T2, T3} indicating the legal-disclosure category of its underlying source. Customers cap the maximum allowable class per workspace, enabling configuration-level rather than code-level compliance posture.

4.Conclusion

An AI agent that judges weak signals, learns domains, and exits cheaply provides a credible alternative to single-vendor enrichment. We release the system as a public API with a free tier for evaluation.