Relevance AI BDR agent (L3-4)

The steps

1. 01

   Define a narrow ICP segment

   Before touching Relevance AI, you must define an ultra-narrow Ideal Customer Profile (ICP) to prevent the AI from hallucinating or sending generic outreach. A broad ICP is the #1 reason AI agents fail. • Identify a specific segment (e.g., "Series B FinTechs in the UK using Salesforce"). • Create a CSV or Google Sheet with exactly 20 'Gold Standard' leads manually. • Detail the specific pain point this group faces right now. • Tools: LinkedIn Sales Navigator or Apollo.io for lead sourcing. • Owner: Sales Operations or Growth Lead. • Time: 3-4 hours of deep research. • Pitfall: Targeting "all SaaS companies" is too broad; the agent won't know which value prop to pick. • Definition of Done: You have a CSV with 50 leads and a 3-sentence description of the exact trigger (e.g., "recently hired a new VP of Sales") that clarifies why we are reaching out today. • QA Check: Could a human explain the difference between this segment and the rest of the market in 10 seconds?

2. 02

   Build the research & draft agent

   Log into Relevance AI and create a new ‘Agent’. Focus exclusively on one 'Play',the initial outbound sequence to a single persona. • In the Relevance AI Tool Builder, add a "Search Web" step to look up the lead's LinkedIn profile and company news. • Add an "LLM Step" (use GPT-4o) with a prompt like: "Based on {{company_description}} and {{lead_bio}}, identify the top 3 challenges they face. Reference a specific recent LinkedIn post if found." • Configure the system prompt to ignore generic praise ("I see you are a leader...") and focus on utility. • Owner: BDR Manager or RevOps. • Time: 2 hours. • Pitfall: Building a complex multi-path agent too early. Keep it to: Research -> Personalize -> Draft. • Definition of Done: The agent successfully generates a unique, 3-sentence personalization snippet for 5 test leads. • QA Check: Does the personalization sound like it was written by a peer, or does it sound like an AI summary? If the latter, refine the prompt tokens.

3. 03

   Connect email delivery tools

   Now, integrate your messaging platform (e.g., Smartlead, Instantly, or Hubspot) with Relevance AI using an API key or the native integration. • Map the 'Draft' output from Relevance AI to a custom variable in your email tool (e.g., {{ai_intro}}). • Set up a sequence where the first 150 characters of the email are dynamic, while the rest of the value prop is static but highly tailored to the niche. • Ensure the "From" address is a secondary domain (e.g., name@get-company.com) to protect your primary domain reputation. • Owner: IT or RevOps. • Time: 1 hour. • Pitfall: Sending directly from the agent without a warmup period for the email account. • Definition of Done: A test email is successfully triggered from Relevance AI and arrives in your personal inbox with the dynamic fields correctly populated. • QA Check: Check for formatting errors like double spaces or broken brackets in the dynamic text.

4. 04

   Implement Human-in-the-Loop (HITL)

   Crucially, enable the 'Approval' toggle in Relevance AI. For the first 100 leads, the agent must not send anything without a human clicking 'Approve'. • Review each draft for 'hallucinations' (making up fake news or names). • If a draft is bad, don't just fix it manually; go back to the Relevance AI 'Agent Instructions' and update the prompt logic to prevent that error from happening again. • Document common failure patterns (e.g., "Agent keeps mentioning the wrong HQ location"). • Owner: Senior BDR or Team Lead. • Time: 2-3 days (intermittent). • Pitfall: Auto-approving because the first 5 looked "okay." The 6th one is usually where the error is. • Definition of Done: 100 emails sent with a 100% human-verified accuracy rate. • QA Check: Are there any "As an AI language model..." phrases in the drafts? If so, add a negative constraint to the system prompt immediately.

5. 05

   Deploy reply handling logic

   Configure the agent to monitor the inbox for replies. In Relevance, use the "Reply Handling" template. • Categorize replies into three buckets: 'Interested/Meeting Requested', 'Objection/Question', and 'Not Interested'. • For 'Interested', set a trigger to alert the BDR via Slack or email to take over and manually book the meeting. • For 'Objections', give the agent a knowledge base (PDF of your FAQs) to draft a response, but keep it in 'Draft' mode for human review. • Owner: BDR. • Time: 1 hour for setup; daily monitoring. • Pitfall: Letting the AI book the meeting directly. AI scheduling is still brittle; a human should provide the final Calendly link. • Definition of Done: The agent successfully tags a reply as 'Interested' and sends a Slack notification to the team. • QA Check: Test the 'Not Interested' filter with a "Remove me" reply to ensure the agent doesn't try to keep selling to them.

6. 06

   Audit ROI and Cost-per-Meeting

   After 60 days of operation, perform a hard audit of the program’s economics. • Calculate 'Cost Per Meeting' by adding the Relevance AI subscription + LLM token costs + the BDR's time spent monitoring/fixing. • Compare this to the historical Cost Per Meeting of a fully manual BDR. • Formula: (Monthly Software Cost + (Hours Spent * Hourly Rate)) / Total Meetings Booked. • Owner: Head of Sales or CFO. • Time: 2 hours. • Pitfall: Ignoring the 'hidden cost' of the time it takes to fix the agent's mistakes. • Definition of Done: A spreadsheet or dashboard showing whether the agent is at least 20% cheaper than a human-only approach. • QA Check: Is the 'Meeting to Close' rate of AI-sourced leads the same as human-sourced leads? If AI leads are lower quality, the agent needs better ICP filtering.

Next playbooks

• Always-on AI meeting notes (L1)
• ChatGPT for first-draft cold emails (L1)
• Personal AI assistant for repetitive sales admin (L1)
• Zapier + ChatGPT inbound lead summarizer (L1)