Case Study: Sales Team Replaces $18k/yr Enrichment with an AI Pipeline

Background

PipelineHQ sells workflow software to mid-market ops teams. Their 6-person SDR team ran 2,000 outbound touches/month. Apollo enriched the list — titles, industries, headcount — but personalization hooks were boilerplate: "I saw you're growing fast."

Reply rates plateaued at 4.1%. Leadership asked whether a custom pipeline could beat generic enrichment without hiring a data team.

The problem

The AI Lead Enricher workflow was the starting point.

Architecture

CSV in (name, email or company)
  → Per row (parallel, capped at 10):
      Tavily: "{name} {company}" + "{company} news last 30 days"
      Claude: extract title, role, industry, headcount, hook + source URL
  → Rows without source URL → NEEDS_REVIEW flag
  → CSV out → Google Sheets for reps

Implementation phases

Phase 1 — Parity check (1 week)

They ran 200 leads through both Apollo and the custom pipeline. Title match rate: 88% (Apollo) vs 82% (custom). Hook usability (blind rep review): 43% (Apollo) vs 67% (custom). The hooks won.

Phase 2 — Production pipeline (2 weeks)

Python script with:

• Structured output schema (Pydantic + Instructor)
• Company-level news cache (7-day TTL in SQLite)
• NEEDS_REVIEW gate for missing sources
• Nightly cron via GitHub Actions

Cost per lead: ~$0.07 (search + extraction). 2,000 leads/month ≈ $140.

Phase 3 — Rep workflow integration (week 4)

Output landed in Google Sheets with columns: hook, hook_source, recent_news, confidence. Reps filtered confidence >= 0.8 for same-day sends; lower confidence went to manual research queue.

Results

The +2.4pp reply rate translated to ~12 extra conversations/month — worth far more than the $1,360/mo saved.

Quality controls

1. Source URL required — no URL, no hook shipped.
2. 10% weekly audit — rev ops sampled rows and scored accuracy.
3. Hallucination playbook — if news predated 30 days, row auto-flagged.
4. Rep feedback column — bad hooks fed back into prompt changelog.

What they'd do differently

1. Cache company news from day one — duplicate searches were 40% of early API spend.
2. Train reps on the confidence column — some sent low-confidence hooks anyway and blamed the tool.
3. Keep Apollo for phone numbers — custom pipeline didn't solve mobile enrichment; they kept a lighter data vendor for that one field.

Takeaway

Enrichment SaaS sells convenience. Custom pipelines sell specificity. For outbound teams where the opening line is the product, search + structured extraction often beats a $1,500/mo seat — if you enforce source citations and human review gates. Start with the workflow guide.