How to Build an AI Sales Agent with OpenClaw (Replace Your First SDR)

I spent three months looking for an SDR last year. Posted the job, ran interviews, made an offer. She lasted six weeks before taking a better gig.

During those six weeks, she researched about 200 prospects, sent maybe 400 emails, and booked 11 calls. Not bad work, honestly. But the math bothered me: salary, benefits, tools, management time — it was north of $6,000 a month when you added it all up.

So I tried something different. I built an OpenClaw agent that does roughly 80% of what she did, runs 24/7 without coffee breaks, and costs me $18 a month.

This guide walks through exactly how I set it up. The agent researches prospects, writes personalized cold emails, follows up on a schedule, and logs every interaction to HubSpot. It is not perfect — I will be honest about where it falls short — but for early-stage outbound, it punches way above its price tag.

Build Your AI Sales Agent: Step by Step

Here is the full setup, broken into five pieces. You can get a basic version running in under an hour. A production-ready agent takes a day or two of tuning.

Step 1: Define Your Ideal Customer (SOUL.md)

Every OpenClaw agent starts with a personality file. For a sales agent, this file is where you define your ICP (ideal customer profile), outreach tone, and qualification criteria.

# SOUL.md — AI Sales Agent

## Identity
You are a sales development agent for [Your Company].
Your job: find qualified prospects, research them, and send personalized outreach.

## Ideal Customer Profile
- B2B SaaS companies, 20-200 employees
- Series A to Series C funding
- Using at least one competitor product (HubSpot, Outreach, Apollo)
- Based in US or EU

## Outreach Rules
- Never send more than 50 emails per day
- Always reference something specific about the prospect's company
- Tone: professional but human — like a smart colleague, not a robot
- If a prospect replies with "not interested", mark as closed-lost and stop
- Follow-up sequence: Day 0 (initial), Day 3, Day 7, Day 14 (final)

## Qualification Criteria
- Has a sales team of 5+ people
- Annual revenue > $2M (check Crunchbase or LinkedIn)
- Currently hiring for sales roles (strong buy signal)

This file shapes everything the agent does. Be specific about what makes a good prospect for you. Vague instructions lead to vague outreach.

Step 2: Set Up Prospect Research

The agent needs to find and research prospects before it can write to them. You have two paths here:

Path A: Apollo skill — If you have an Apollo.io account, install the apollo skill from ClawHub. It gives your agent access to Apollo's people and org enrichment API. You can search by company size, funding stage, tech stack, and job title.

# Install the Apollo skill
openclaw install apollo

# Configure API key
export APOLLO_API_KEY=your_key_here

# Test: search for VP Sales at SaaS companies with 50-200 employees
openclaw run "Find 10 VP Sales at B2B SaaS companies, 50-200 employees, Series A-B"

Path B: Browser-based research — No Apollo account? The agent can use OpenClaw's built-in browser automation to scrape company websites, check LinkedIn profiles, and pull contact info from public sources. Slower, but free.

I use both. Apollo for the initial list, browser research for the personalization details that make outreach actually work — recent blog posts, product launches, job openings. (There is also a phone-caller skill if you want voice outreach. I tried it once. The prospect thought they were talking to a real person for about 45 seconds. Impressive and slightly unsettling.)

Step 3: Build the Outreach Engine

Install the cold-outreach skill from ClawHub. This one generates personalized cold emails using proven frameworks like AIDA, PAS, and BAB.

# Install cold outreach skill
openclaw install cold-outreach

# Create an outreach skill file
cat > outreach.skill.md << 'EOF'
# Sales Outreach Agent

## Task
For each prospect in my pipeline:
1. Research their company (website, LinkedIn, recent news)
2. Identify a specific "hook" — something relevant happening at their company
3. Write a 3-sentence cold email using the hook
4. Subject line: under 5 words, no clickbait
5. CTA: ask for a 15-minute call, suggest 2 specific times

## Email Rules
- First line must reference something specific (not "I saw your company")
- No bullet lists in emails — write like a real person
- Under 100 words total
- Sign off with my real name and title
EOF

Specificity is what gets replies. "I noticed you just launched a Shopify integration last week" beats "I saw your company is doing interesting things" every time. That is the whole reason the research step exists.

Step 4: Connect Your CRM

Your agent is useless if it cannot log what it does. Here is how to connect the three most common CRMs:

# HubSpot MCP server config (add to openclaw config)
{
  "mcpServers": {
    "hubspot": {
      "url": "https://mcp.hubspot.com",
      "auth": {
        "type": "oauth",
        "clientId": "your-client-id",
        "scopes": ["crm.objects.contacts.write", "crm.objects.deals.write"]
      }
    }
  }
}

Step 5: Schedule the Loop

This is the part that changes everything. Set a heartbeat interval and the agent runs the full pipeline on its own.

# Schedule the agent to run every 4 hours
openclaw schedule --interval 4h --skill outreach.skill.md

# What happens each cycle:
# 1. Check for new prospects in the research queue
# 2. Research top 10 prospects (company + person)
# 3. Draft personalized emails for each
# 4. Send emails (respecting daily limits)
# 5. Check for replies to previous outreach
# 6. Send follow-ups where due
# 7. Log all activity to CRM
# 8. Report summary to Slack

I run mine every 4 hours during business hours (8am-8pm). That gives the agent 3 cycles per day, each processing about 15-20 prospects. At that pace, you are doing 300+ personalized touches per week without lifting a finger.

What a Day in the Life of Your AI SDR Looks Like

Here is what my agent actually does on a typical Tuesday:

By the end of the day, the agent has researched 45 prospects, sent 45 personalized emails, processed 20+ follow-ups, and updated every record in CRM. I spent zero minutes on it. The first time I saw the Slack summary pop up at 9:17am — "Researched 15, emailed 15, 2 replies flagged" — while I was still drinking my coffee, it felt like cheating.

Which AI Model to Use (and How to Save Money)

You do not need the most expensive model for every task. I run a two-model setup that cuts API costs by about 60%:

# Two-model config in openclaw settings
{
  "models": {
    "research": "claude-haiku-4-5",
    "writing": "claude-sonnet-4-5"
  },
  "routing": {
    "prospect-research": "research",
    "crm-logging": "research",
    "email-drafting": "writing",
    "follow-up-writing": "writing"
  }
}

With this setup, roughly 70% of the agent's work runs on the cheap model. Only the actual email copy uses the premium one. My API bill dropped from $22/month to about $9/month after switching.

Frequently Asked Questions

The Bottom Line

An AI sales agent is not magic. It does not close deals, and it does not replace the need for someone who actually understands your customers.

What it does is handle the tedious stuff — research, cold emails, follow-ups, CRM data entry — on autopilot. Around the clock. For less than you spend on coffee in a week.

The agents that actually work are the ones where someone spent real time on the SOUL.md file and kept a human in the loop for anything that required judgment. I am still not fully convinced this scales past Series A companies with simple sales motions. But for $18 a month, you can find out for yourself pretty fast.

Nikhil Kumar (@nikhonit)

Growth Engineer & Full-stack Creator

I bridge the gap between engineering logic and marketing psychology. Currently leading Product Growth at Operabase. Builder of LandKit (AI Co-founder). Previously at Seedstars & GrowthSchool.

Visit LandKitConnect on LinkedIn