Multi-label by design
A company can qualify for A and C but not B without overwriting any result.
Case study 05 / B2B lead generation
5M+ companies processed.
98% accurate.
A production multi-agent pipeline qualifies companies across industries and overlapping niches while scaling model capability with the complexity of each case.
5M+ companies processed92% lower model costHuman review for rare exceptions
01
98%
classification accuracy in production
02
−92%
model cost through complexity routing
03
5M+
companies processed across industries
04
3 tiers
of progressively capable agents
The classification problem
Classify 5M+ companies.
Separate 40 niches.
A B2B outreach program needed to qualify tens of thousands of companies against dozens of narrow, sometimes overlapping sub-niches. Generic provider filters routinely pulled adjacent but incorrect businesses into the same list.
The issue was classification precision. One company could legitimately qualify for Segment A and C but not B. A shared classification field silently overwrote that reality and contaminated campaigns.
The hardest errors came from companies offering many different services. Their broad language looked relevant to several niches, which made disciplined context and confidence routing more important than simply choosing a larger model.
The system
Route by complexity.
Cut model cost 92%.
Every tier has a high-confidence exit. Most records resolve from the company description. Only harder cases earn website tools, web search, broader context, or human judgment.
Qualification run / confidence escalation
Every tier has a confident exit
5M+ processed98% accurate
Data and context architecture
Isolate each company.
Preserve every verdict.
A junction table stores one independent company-to-segment verdict per row. Each execution receives one company, one segment definition, the relevant qualifying and disqualifying criteria, and only the evidence gathered for that case.
Context managed tightly
Every tier sees only the criteria, evidence, and tools required for its decision.
State carried forward
Confidence, sources, and the precise reason for escalation travel with the record.
Production engineering
Process millions.
Contain four failure modes.
01Cross-item contaminationAgents were isolated into strict one-company executions so one record could never leak into another.
02Runaway tool useExplicit stopping conditions and tool budgets taught agents when a blocked page was a valid conclusion.
03Silent field lossCriteria were preserved across every escalation so later agents always knew exactly what they were evaluating.
04Index misalignmentFiltered branches were keyed to actual records rather than fragile array positions.
Cost follows complexity
Start with cheap models.
Spend 92% less.
A flat architecture would send every record through the most capable agent with website tools, web search, and a large context window. This system starts with the smallest useful model and narrowest useful context. Capability expands only after a low-confidence result.
Pass 1
Text only
Description, segment definition, and strict output schema.
Escalation
Tools on demand
Website access and external search appear only when the evidence requires them.
Measured result
−92%
Model cost versus treating every company as a maximum-complexity case.
Human review became an exception path.
It was reserved mainly for scraper-blocked websites and genuinely unusual companies the evidence could not resolve. The automated system produced accurate results around 98 percent of the time.