The honest map
Lead generation has six steps. Agents win at four and lose at two.
Step 1 — ICP definition. Human. Agents can't read your contracts well enough to know who you actually serve.
Step 2 — Prospect discovery. Agent. Crawling Google Places, NPI Registry, LinkedIn signals, etc. is mechanical. Cost: $0.03/business via DeepSeek and Places API.
Step 3 — Enrichment. Agent. Decision-maker name, title, verified email, LinkedIn, company size, tech stack. Cost: $0.10-0.50/prospect depending on depth.
Step 4 — Sequencing. Agent for execution, human for design. Agents shouldn't invent the messaging from scratch; they should use approved templates with personalization.
Step 5 — Reply handling. Human. Cold replies are political and contextual; an agent that auto-replies to a "stop emailing me" with a clever line will lose you the relationship.
Step 6 — Meeting booking. Agent or hybrid. If you have a clear qualification rubric, an agent can book; if it's judgment-heavy, a human takes the call.
That's the map. Now the details.
What an enrichment agent actually finds
For a B2B prospect at a 50-person construction company, a well-built enrichment agent will surface:
- Decision-maker name + title (typically COO or VP Marketing at that size)
- LinkedIn profile + tenure
- Verified work email (with confidence score; we drop sends below 0.85)
- Estimated company employees + revenue band
- Technologies detected on their site (CMS, ad pixels, analytics, helpdesk)
- Last website activity (if your tracking pixel ever fired for them)
- Touchpoint history across all channels
What a vanilla LLM cannot do without enrichment APIs: get the verified email reliably. You need a real verifier (Hunter, NeverBounce, ZeroBounce) in the loop, not just LLM guessing.
What an outbound sequencer actually runs
Six-step sequence is the floor for B2B (one-step is too thin):
1. Email day 0 — short, specific, references something true about them 2. Email day 3 — different angle on the same problem 3. LinkedIn connect day 5 — no message, just connect 4. Email day 7 — direct ask for 15-minute call 5. LinkedIn message day 10 — soft check-in after connection 6. Final email day 14 — "should I close the loop?" exit
Agent's job:
- Personalize lines 1-2 of each email using enrichment data
- Track opens, clicks, replies, bounces
- Pause the sequence if a reply comes in
- Classify replies (interested, not now, never, wrong person, OOO) for human follow-up
- Promote next step automatically when prior step is delivered
What we won't let the agent do: write the *body* of cold emails from scratch. We use approved templates with structured variable slots. Agents fill the slots, humans approve the templates.
Where this breaks
1. List quality kills everything. A perfect agent on a bad list still delivers nothing. Spend 80% of effort on the ICP definition + list build; 20% on the sequence.
2. Hot leads can't wait for batches. If someone fills out a form, an agent should respond within 60 seconds — not when the next cron tick runs. Get the trigger architecture right or you lose the warm leads.
3. Reply classification mistakes are expensive. An "interested" misclassified as "not now" is a missed deal. An "never email me" misclassified as "ambiguous" is a complaint waiting. Run human review on the first 500 classifications before trusting the automated path.
How ANKR ties this together
Our brain links:
- The prospect database (the pool you target)
- The enrichedProspect layer (what you know about each)
- The prospectSequence (the play you're running)
- The sequenceEvent stream (what happened — opens, clicks, replies)
- The touchPointLog (everything that's ever been done with that prospect, across channels)
- The agentRun monitor (live view of which agent is doing what right now)
You see the whole funnel at the same level of fidelity, and every agent action gets audit-logged. When a sales rep takes over for a hot reply, they read a full history, not a stale CRM note.
How to start
If your team currently runs lead gen by Apollo + a spreadsheet, the agentic upgrade is high-leverage because the execution gap (between identifying a prospect and finishing a sequence) is large.