How to Build a Unified AI Lead Qualification System as a Solo Founder

You’re getting leads from three places: website forms, inbound emails, and live chat.
Each one arrives in a different format, gets scored (or not scored) differently, and ends up in different places — or worse, the same place with conflicting information.
You waste hours every week trying to figure out which leads are actually worth talking to, and you still miss the hot ones because you were busy building.

Here’s the thing: you can stop doing detective work.
One AI agent can qualify leads coming from forms, email, and chat — using the exact same scoring logic every time — so you get a single, trustworthy number and reason no matter where the lead came from.

This setup costs $30–55/month and takes 4–8 hours to build the first time. After that it runs itself.

Let’s build it.

Why separate channels are quietly killing your pipeline

Most solo founders treat channels independently:

• Form → lands in Google Sheet or Typeform dashboard

• Email → sits in Gmail or Help Scout

• Chat → lives in Tidio / Crisp / Intercom

You end up with three different mental models for “is this person serious?”.
That means:

• You over-prioritize chat leads because they feel “hotter” (even when they’re students)

• You ignore form leads because they look cold on paper

• You double-work the same person who filled a form and then emailed you two days later

Result: inconsistent qualification → inconsistent follow-up → fewer real conversations.

A unified system fixes exactly that: same brain, same scorecard, same output — no matter the channel.

The architecture — three pieces only

1. Single scoring brain

2. Channel normalization layer

3. Deduplication & conflict resolver

That’s it. No complex microservices, no vector database, no $400/month enterprise platform.

1. The single scoring brain

Everything funnels into one prompt that lives in Claude, Grok, or OpenAI.

Your scoring brain is just a well-written prompt that takes structured input and spits out:

• Score 1–10

• One-sentence reason

• Recommended next action (ignore / low-priority / schedule call / urgent)

Example prompt skeleton (copy-paste and customize):

text

You are a ruthless B2B lead qualifier for a solo SaaS founder.

Company: [Your product name]
Ideal customer: [paste 3–5 bullet points of your real ICP — company size, role, pain, budget signals]

Score this lead from 1 (completely irrelevant) to 10 (paying customer tomorrow).

Inputs:
- Name: {{name}}
- Company: {{company}}
- Email domain: {{email_domain}}
- Stated pain / message: {{message_text}}
- Budget mentioned: {{budget_mentioned_yes_no}}
- Timeline mentioned: {{timeline}}
- Source channel: {{form / email / chat}}

Output ONLY in this exact JSON format:

{
  "score": number 1-10,
  "reason": "one short sentence explaining the score",
  "next_action": "ignore" OR "low-priority follow-up" OR "schedule discovery call" OR "urgent — reply today",
  "confidence": "high" OR "medium" OR "low"
}

Be brutal. If they are a student, consultant, or tiny agency with no budget — score 1–3.

Cost: Claude 3.5 Sonnet or Grok-2-mini ≈ $15–25/month at moderate volume.

You call this prompt from Zapier or Make.com whenever a lead arrives (after normalization).

2. Channel normalization — make everything look the same

You need to turn wildly different inputs into the same fields the scoring prompt expects.

Typical sources and how to normalize them:

Practical setup most solos use right now (2026):

• Zapier (free tier or $20/mo Starter) or Make.com ($9/mo Core)

• One “Parse with AI” step using Claude or OpenAI API

Example Zap:

1. Trigger: New Typeform entry

2. Action: Claude – “Extract name, company name, pain points, budget hints, timeline from this form response”

3. Output: structured row with name, company, message_text, budget_mentioned, etc.

4. Next step: send to scoring brain

Same pattern for email and chat.

Time to build the three zaps: 2–4 hours total.

3. Avoiding double counts & score conflicts

The fastest way to break trust in your own system: the same person submits a form, then chats two days later, and you get two different scores.

Solution: one central source of truth + email as the unique key.

Simplest 2026 solo stack:

• Airtable (free tier or $20/mo Plus) or Google Sheets

• Unique key = email address (lowercase, normalized)

Flow:

1. New lead arrives (any channel)

2. AI normalization step finishes

3. Check: does this email already exist in Airtable / Sheet?

   • Yes → update existing record with new data + rescore once

   • No → create new record + score

4. Store: score, reason, timestamp, source channel, raw message

Tools that handle this cleanly:

• Airtable has “Find record” + “Update / Create” actions in Zapier

• Google Sheets: use “Lookup” + conditional paths in Make.com

Most common conflict patterns and fixes:

• Person fills form → emails same day → score changes from 6 → 8 → keep the latest score (usually better data)

• Chat visitor gives fake email → real email later → treat as new until email matches

Step-by-step: build it this weekend

1. Write your ICP description and scoring prompt (1 hour)

2. Create a simple Airtable base or Google Sheet with columns: email, name, company, score, reason, next_action, last_source, last_updated

3. Build three Zaps / scenarios (one per channel) — each ends with:

   • Parse message → check for existing email → score or update → save (3–5 hours)

4. Send 10 real past leads through the system manually (paste into a test zap)

5. Tweak prompt until scores match what you would have decided (1–2 hours)

6. Turn zaps on

Total cost:

• Claude API: $15–25/mo

• Zapier Starter or Make Core: $9–20/mo

• Airtable: $0–20/mo → $24–65/mo

Mistakes that burn most solo founders

• Vague prompt → scores all over the place → fix: add 3–5 real example leads + scores in the prompt

• No deduplication → duplicate records everywhere → fix: force email as key from day one

• Forgetting to capture channel → you can’t see which source brings better leads → add “source” field

• Never reviewing low-confidence scores → AI quietly starts hallucinating → spend 15 min/week looking at 3–5 lowest-confidence leads

• Trying to score too many criteria → prompt becomes inconsistent → stick to 5–7 signals max

When this stops working

This exact setup handles 50–400 leads/month comfortably.

Signs you’ve outgrown it:

• 500+ leads/month → latency and cost go up; consider n8n self-hosted or Relay.app

• Complex multi-person buying committee → simple 1–10 score isn’t enough

• You want real-time bidding / enrichment → migrate to Clay + GPT + HubSpot

Until then, this is stupidly effective for the money and time invested.

Do this tomorrow morning

1. Open a doc and write your real ICP in 4–6 bullets (company size, role, pain, budget signal, red flags)

2. Paste the prompt skeleton above and customize it with your ICP

3. Create one Airtable table or Google Sheet with the columns listed

4. Build just the form → parse → score → save zap first (the easiest one)

Once that works on 5 test leads, add email, then chat.

You’ll have your first unified score by the end of the week — and you’ll stop wondering which inbox tab has the real money.

Let me know when you hit the first conflict or weird score — we can tighten the prompt together.