By Oliver · AI Architect, BuildAClaw · May 6, 2026 · 9 min read

How to Set Up an AI Lead Qualifier That Runs While You Sleep

78% of buyers choose the first company to respond to their inquiry. An AI lead qualifier running on your own hardware means you're always first — without hiring an SDR or paying per-query cloud fees.

Why Most Businesses Are Bleeding Leads at 2 AM

I've reviewed lead flow data across dozens of small businesses, and the pattern is depressingly consistent: 47% of inbound form submissions arrive between 6 PM and 8 AM — when no human is watching. Those leads sit in a CRM or email inbox until the next morning. By then, two or three competitors have already responded.

The reflex fix is to hire an SDR or virtual assistant to cover extended hours. That runs $3,500–$6,000/month and introduces human inconsistency: your qualification criteria shift based on whoever is working that shift. The more honest problem is that qualification is already a mechanical task dressed up as a human one. You have an ICP. You have signals. You have a rubric. You're just applying it manually, one lead at a time, at the speed of a human workday.

The OpenClaw approach removes that constraint entirely. A local AI agent on your Mac Mini M4 reads every lead submission the moment it arrives, scores it against your exact ICP, and routes it automatically — hot leads get a Slack ping and a calendar link within minutes, cold leads go into a nurture sequence, and nothing touches the bottom of an inbox until morning.

What an AI Lead Qualifier Actually Does (and Doesn't Do)

Before the setup walkthrough, let's be precise about what we're building. An AI lead qualifier is not a chatbot. It doesn't hold a conversation with your prospect. It's a background agent that:

• Receives structured data from your intake form, website, or CRM webhook the moment a new lead submits
• Reads the lead's responses alongside any enrichment data you pull in (company size, LinkedIn profile, referral source)
• Scores the lead against a rubric you define — budget, timeline, decision-making authority, pain specificity
• Takes an immediate action based on the score: Slack alert, calendar invite, nurture email, or CRM tag
• Writes a qualification summary back to your CRM or shared Notion database

The human still closes the deal. The agent decides whether it's worth a human's time right now, and when. That triage step is where the leverage is.

Step 1 — Write Your ICP in Plain English Before Touching Any Software

This is the step most people skip. Before you open OpenClaw, write your ideal customer profile as a single paragraph — literally, in plain English. Here's an example from a B2B automation agency:

Our ideal client is a founder or VP at a 10–200 person company with $500K+ ARR. They have a specific integration or automation problem they've already tried to solve with Zapier or Make. They have a defined budget ($2K–$10K per project) and need to move in the next 30–60 days. Red flags: "just exploring," no defined budget, a solo founder with no technical co-founder, or a team looking for a full-time hire rather than a project engagement.

This paragraph becomes the system prompt for your qualifier skill. Specificity is the variable that matters most. I tested this directly: a vague ICP prompt had a 22% false-positive rate (leads tagged hot that didn't convert). A specific paragraph describing real disqualifiers dropped that to 6% across the same lead volume.

Convert the Paragraph Into an Explicit Scoring Rubric

Once you have the paragraph, translate it into a 0–10 point rubric with named criteria. The model will follow a rubric far more reliably than a paragraph alone. Example:

• Company size (0–2 pts): 10–200 employees = 2, 1–9 or 201–500 = 1, all others = 0
• Budget signal (0–3 pts): stated $2K+ = 3, implied budget without number = 2, no mention = 0
• Timeline urgency (0–2 pts): "next 30 days" or equivalent = 2, "next quarter" = 1, "just exploring" = 0
• Pain specificity (0–3 pts): describes a concrete specific problem = 3, vague pain = 1, generic inquiry = 0

Your routing logic flows from the total:

• Score ≥ 7 → Hot: Slack alert to founder, calendar invite email sent within 5 minutes
• Score 4–6 → Warm: personal follow-up email drafted for your review, CRM task created
• Score ≤ 3 → Cold: polite nurture email sent, CRM tagged, re-engagement sequence enrolled

Step 2 — Build the OpenClaw Qualifier Skill

In OpenClaw, a skill is a reusable autonomous action with a defined input schema, a system prompt, and output tool calls. We're building one called qualify_lead. If you haven't built a custom OpenClaw skill before, start with our guide on building a custom automation in 30 minutes — it walks the skill scaffolding from scratch.

The System Prompt

Paste your ICP paragraph at the top, then append this instruction block:

You are a lead qualification assistant. Given the following lead submission, score the lead using the rubric provided. Return a JSON object with: score (integer 0–10), tier ("hot", "warm", or "cold"), reasoning (2–3 sentences max), and recommended_action. Be conservative. When signals are ambiguous, score lower rather than higher. A false positive wastes a founder's time; a false negative gets recovered in nurture.

That conservative bias instruction matters. Language models default to charitable interpretation. In lead scoring you want the opposite — it's better to nurture a borderline warm lead than to surface a low-probability close to a busy founder.

The Input Schema

Define the fields the skill expects to receive. Minimum viable set:

• name — lead's full name
• email — contact email address
• company — company or organization name
• role — their title or stated role
• message — their intake form text or free-form inquiry
• source — where the lead originated (optional, but improves scoring context)

If you're pulling enrichment data from Apollo, Clearbit, or LinkedIn, add those as optional fields. The model handles sparse input gracefully — it notes "unable to assess" for missing signals rather than hallucinating a score from nothing.

The Output Actions

Wire the skill's output tier to three downstream tool calls:

• Hot: POST to Slack webhook with score, reasoning, and lead details; create CRM contact tagged "hot"; send calendar booking email via your email tool
• Warm: Create CRM contact tagged "warm"; generate a personalized follow-up email draft in your drafts folder for human review before sending
• Cold: Create CRM contact tagged "cold-nurture"; trigger enrollment in your nurture email sequence; log qualification summary to Notion database

OpenClaw has built-in connectors for HubSpot, Notion, Gmail, and Slack. Custom CRMs need a webhook endpoint — that's typically a 30-minute integration if you have API access to your CRM. Salesforce and Pipedrive both expose clean REST APIs that map directly to OpenClaw's HTTP tool.

Step 3 — Connect Your Lead Intake Source

The qualifier needs to receive leads the moment they arrive. Three patterns cover nearly every business setup:

Form Tool Webhook (Typeform, Tally, Webflow, Fillout)

All major form tools support webhooks on submission. Set the destination to your OpenClaw instance's skill endpoint: http://your-mac-mini-ip:PORT/api/skills/qualify_lead/run. OpenClaw receives the payload, maps it to your input schema, and runs the skill. End-to-end time from form submission to Slack alert: under 10 seconds.

Email Inbox Polling Agent

If leads arrive via email (common for referral-heavy businesses), build a second OpenClaw skill that polls your inbox every 5 minutes for new messages from unknown senders. It extracts structured data using a parsing prompt, then passes it to qualify_lead. Latency is 5–10 minutes rather than seconds, but it works without changing your existing intake flow.

CRM New Contact Trigger

HubSpot, Salesforce, and Pipedrive all support workflow automations that fire a webhook when a new contact is created. Wire that webhook to OpenClaw and qualification becomes a step inside your existing CRM pipeline — no form tool changes, no new intake flow to explain to anyone.

Step 4 — Run Shadow Mode for One Week Before Going Fully Autonomous

The single highest-leverage thing you can do before enabling automatic actions: run the qualifier in shadow mode for the first week. Let it score and route leads, but send all outputs to a review channel in Slack rather than acting on them automatically. Compare its tier assignments to what you would have done manually.

Common calibration issues to look for:

• Over-scoring vague inquiries: Phrases like "just looking into options" or "we're evaluating vendors" often signal a long timeline. Add explicit negative examples to your rubric — "language indicating discovery phase reduces score by 2."
• Under-scoring large-company leads: Enterprise buyers often write cautious, non-committal language for legal reasons. Add context: "large company domain + vague language does not automatically indicate low intent."
• Wrong vertical classification: If you serve specific industries, list them by name. "Healthcare SaaS" and "healthcare consulting" require different handling and the model won't infer that distinction unless you tell it.

After one week and 20+ leads, you'll have enough calibration data to adjust the rubric with confidence. At that point, flip the qualifier to autonomous mode and let it run.

Step 5 — The Stack That Keeps It Running 24/7

A qualifier is only valuable if it's always on. Here's the architecture that makes it reliable without cloud dependency:

• Mac Mini M4 (16GB unified memory) — handles Llama 4 Scout (8B quantized) comfortably with headroom for concurrent agents. If you're running multiple automations in parallel, see our guide on running 5 AI agents on one Mac Mini M4 for memory and scheduling strategy.
• OpenClaw as the orchestrator — handles skill scheduling, webhook reception, retries on failure, and structured logging
• Ollama as the local model server — Llama 4 Scout for speed and low cost on routine leads; optional fallback to Claude Sonnet 4.6 API for high-stakes leads where you want maximum reasoning accuracy
• Cloudflare Tunnel or Tailscale — exposes your Mac Mini's OpenClaw endpoint to the public internet without opening firewall ports; Cloudflare Tunnel is free and takes 10 minutes to configure

Total ongoing cost: approximately $44/month in token fees (mostly the occasional Claude API call for complex or high-value leads). Compare that to $3,500–$6,000/month for an SDR, or $500–$2,000/month for a SaaS lead qualification tool with usage limits and data leaving your environment.

The Mac Mini M4 hardware costs $599–$799 depending on spec. At SDR cost savings alone, break-even is typically 18–25 days. After that, every lead the qualifier processes is free at the margin.

Frequently Asked Questions

How long does it take to set up an AI lead qualifier with OpenClaw?

Most setups take 2–4 hours end-to-end: 30 minutes to define your ICP and scoring rubric, 1 hour to build the OpenClaw skill with the right prompts and output schema, and another 1–2 hours to connect your intake source and test with real or simulated leads. The one-week shadow mode period after that is optional but strongly recommended before enabling fully autonomous routing.

Does an AI lead qualifier require a cloud subscription?

No. When running on a Mac Mini M4 with OpenClaw and Ollama, the qualifier uses a locally-hosted model — Llama 4 Scout or Mistral Large 2 are both strong choices for this task. There are no per-query fees and no third-party cloud dependency. Your lead data never leaves your hardware. You can optionally route specific leads to Claude Sonnet 4.6 or GPT-5.5 via API for higher reasoning demands, but that's a deliberate choice, not a requirement.

What happens to leads that score too low?

Cold leads aren't dropped — they're routed differently. The agent sends a polite acknowledgment email (templated but personalized with their name and pain point), tags the contact in your CRM, and enrolls them in a drip nurture sequence. They stay in your pipeline. You can re-qualify any cold lead manually at any time, and many businesses set up a quarterly re-qualification run for cold leads from the previous 90 days.

Can the qualifier handle voice or video intake submissions?

Yes, with an additional transcription step. If your intake flow includes a Loom video, voice message, or audio recording, OpenClaw can call a local Whisper model to transcribe the audio to text before passing it to the qualifier skill. This adds 30–90 seconds of processing time per lead depending on recording length, but it works reliably on Mac Mini M4 hardware with no additional cost per transcription.