Build a Lead Enrichment Agent That Prioritizes Your Pipeline

What this workflow does

Every form fill arrives as a name, an email, and maybe a company. Someone — usually nobody — is supposed to figure out whether it's a Fortune 500 evaluator or a student doing homework. This workflow builds an agent that does that research automatically: within minutes of a lead arriving, it enriches the record from multiple sources, scores fit against your ICP, drafts a one-paragraph briefing, and routes accordingly — hot leads to a rep with context, medium to nurture, junk to the archive.

Outcome: speed-to-lead for qualified prospects drops from hours or days to minutes, reps stop triaging manually, and your nurture segmentation runs on real firmographics instead of self-reported dropdown fields.

Prerequisites

• An automation platform (n8n, Make, Zapier) or an agent framework if you have engineering support
• An LLM API with web search capability, or an enrichment API (Clay, Apollo, Clearbit-class) as the structured-data backbone — the best setups use both
• CRM access with API write permissions (HubSpot, Salesforce, etc.)
• A written ICP definition with explicit scoring criteria
• A Slack channel or CRM queue for the hot-lead handoff

The workflow, step by step

Step 1: Write the ICP as a rubric, not a vibe (2 hours, one time)

The agent can only score against what you specify. Convert your ICP into weighted, checkable criteria, for example:

• Company size 50–2,000 employees (30 points; partial credit at edges)
• Industry in [list] (20 points)
• Persona: title indicates [function] at [seniority] (25 points)
• Intent signals: pricing page visit, competitor comparison content, or "evaluating" language in form fields (15 points)
• Tech stack includes [complementary tools] (10 points)

Define bands: 75+ = hot (route to rep), 40–74 = nurture, under 40 = archive. Also define hard disqualifiers (student emails, embargoed countries, competitors) that zero the score regardless.

Step 2: Build the trigger and enrichment chain

Trigger: new lead created in CRM (webhook). Enrichment sequence:

1. Structured pass: hit the enrichment API with the email domain — company size, industry, funding, tech stack. Fast, cheap, reliable where it has coverage.
2. Agent pass: for whatever the API missed (and it will miss things, especially smaller companies), the LLM researches with web search:

Research this lead: [name], [title if known], at [company/domain].
Find: what the company does (one sentence), employee count estimate,
industry, any recent news (funding, launches, layoffs), and what this
person's role likely owns. Cite a source for each fact. If you cannot
verify something, write "unverified" — do not guess. Return JSON
matching this schema: {...}

3. First-party pass: pull the lead's own behavior from your systems — pages visited, content downloaded, email domain quality, form answers.

Step 3: Score and brief

Feed everything to a scoring call with the rubric from Step 1:

Score this lead against the rubric below. Show your work: points per
criterion with the evidence used. Then write a 3-sentence briefing for
a sales rep: who this is, why they likely came to us, and one relevant
conversation opener based on verified facts only.
Flag confidence LOW if key fields were unverifiable.

Write the score, band, briefing, and evidence back to CRM fields. Keeping the evidence trail in the CRM is what lets humans audit and trust the scores.

Step 4: Route

• Hot: post to the sales Slack channel and assign in CRM — lead name, score, briefing, link. Include the confidence flag.
• Nurture: enroll in the segment-appropriate sequence using the enriched industry/persona fields.
• Archive/DQ: suppress from sequences, tag with the DQ reason.
• LOW confidence at any band: route to a human review queue instead of auto-acting. This single rule prevents most embarrassing failure modes.

Step 5: Capture the verdicts

Add one required field for reps on hot leads: "Was this routing correct?" (yes / should've been nurture / junk). Two seconds per lead, and it's the loop's fuel.

Failure modes and fixes

• The agent hallucinates company facts into briefings. Enforce the citation requirement, keep temperature low, use structured output, and make anything unverifiable render as "unverified" in the briefing rather than silently dropped or guessed.
• Personal Gmail signups score as mysteries. Route free-email leads down a different path: skip company enrichment, score on behavior only, and lean on progressive profiling in nurture.
• Reps ignore the scores. Usually an early-accuracy problem that poisoned trust. Run the agent in shadow mode for 2–3 weeks first — scoring but not routing — and share the accuracy stats before going live.
• Costs creep on junk volume. Order the chain so cheap disqualifiers (free email + no company found + DQ patterns) short-circuit before expensive enrichment and LLM calls run.
• Privacy exposure. Enrich from legitimate business sources only, honor regional rules (GDPR lawful basis for enrichment is your legal team's call, not the agent's), and document what sources the agent may touch.

Turning it into a loop

Monthly, join three tables: agent scores, rep verdicts, and actual outcomes (opportunities created, closed-won). Then ask the model: "Given these 400 scored leads and their outcomes, which rubric criteria predict conversion and which are dead weight? Propose specific weight changes." Apply the changes, version the rubric, and track precision (what share of "hot" leads did reps validate?) cycle over cycle.

The compounding effect: your ICP definition stops being an annual offsite artifact and becomes a living model, recalibrated monthly against what actually closes. Most teams discover within two cycles that at least one criterion they were sure about predicts nothing — the loop pays for itself on that discovery alone.