IN PRODUCTION · CASE BLIVE
Compliance-first outreach engine — in production
A compliance-first B2B outreach research agent, built and run in-house by Agent Foundry Labs. From a declarative ideal-customer profile it researches prospects, scores fit, and drafts governed, on-brand outreach — read-only and daily-capped by design, with hard compliance rules enforced on every message.
- BUILT BY
- Agent Foundry Labs (in-house, self-built)
- DOMAIN
- B2B outreach / sales research
- SCOPE
- Declarative ICP → web research → fit-scoring & lead-typing → governed, on-brand draft generation; read-only, daily-capped by design
- OUTCOME
- One profile definition → 18 researched, 15 fully-drafted leads; zero account risk; agent running-cost engineered down ~60–70%
- STATUS
- Live, in daily use
- STACK
- Provider-neutral agent runtime · MCP tool layer (web search + read-only browser) · policy-as-data governance · hard compliance rules enforced on every message
The problem
Outbound research is slow, and automating it badly is worse than not automating it at all. Researching a prospect properly — who they are, what they are building, whether they fit — takes real time per lead, which is why most outreach is generic. But the automated alternative carries risks a studio cannot accept: scraped data of uncertain provenance, messages that drift off-brand, and automation patterns that put accounts at risk.
We wanted the research depth without any of that — and we wanted to prove the discipline we sell on a system we run ourselves, every day.
The approach
The engine starts from a single declarative ideal-customer profile. From that one definition it researches prospects across the open web through an MCP tool layer — web search plus a read-only browser — scores each against the profile, types the lead, and drafts governed, on-brand outreach for the ones that clear the bar.
Governance is policy-as-data, not prompt engineering: the compliance rules — what may never be claimed, how messages must read, what the agent may touch — live as data the system enforces on every message, rather than as instructions the model is asked to remember. The agent is read-only and daily-capped by design.
What makes it production-grade
The constraints are structural, which is the point. A read-only tool layer cannot post, connect, or send — so account risk is zero by construction, not by promise. Hard compliance rules are enforced on every message, and every run is traced: research steps, scoring decisions, drafts, and cost per task. That instrumentation is what let us engineer the running cost down by roughly 60–70% — same task, same quality bar, measured before and after — because you cannot reduce a cost you are not measuring.
It is also why this system doubles as our reference build: it carries the same evaluation and observability discipline we ship in client engagements, exercised daily on our own workflow.
The outcome
From one profile definition, the engine researched 18 prospects and fully drafted 15 — governed, cited research rather than scraped lists, with zero account risk throughout. It runs in daily use, and its cost per task is tracked like any other KPI.