# AI lead qualification, step by step

> AI lead qualification means scoring every inbound request by fit and intent, then handing it to sales in priority order. A sound setup has three parts: an automation that collects every channel into the CRM in one flow, a scoring model sales can explain, and a review step where a person makes the final call. GDPR and Türkiye's KVKK both limit fully automated decisions.

URL: https://corrplus.net/blog/ai-lead-qualification/
Updated: 2026-10-26
Corrplus Team

## What is the difference between lead qualification and lead scoring?

Lead qualification is the yes-or-no decision on whether a request deserves a sales conversation. Lead scoring is the number that supports that decision by putting requests in order. One is the check at the door; the other is the ticket number in the queue.

Scores usually sit on two axes. HubSpot's documentation separates them clearly. A **fit** score looks at attributes such as company size, industry or job title. An **engagement** score looks at behaviour such as site visits, email clicks and form submissions. Together they answer the question "is this the right customer at the right time?"

What AI adds is a third input: free text. The message box on a contact form, a WhatsApp thread or an email often carries the most useful information. "We need a quote for three branches and will decide by the end of the month" fits no form field. A language model can read that text and turn need, urgency and budget signals into structured fields.

## When should you use rules, and when AI-assisted qualification?

The three approaches are rarely alternatives. They are layers. Which one you start with depends on your lead volume, your data history and how clearly your sales team can state its criteria.

| | Rule-based score | Predictive score (inside the CRM) | Qualification with a language model |
|---|---|---|---|
| Input | Form fields, behaviour events | Past won and lost records | Free text: messages, emails, chats |
| Explainability | High, every point has a rule | Medium, the model can show which attributes weigh most | High if you ask for a rationale, low if you do not |
| Setup needs | A workshop with sales and CRM configuration | Enough labelled history and the right CRM tier | An automation flow, prompt design, a test set |
| Best fit | Low to medium volume | Medium to high volume, with a steady sales cycle | Requests with long, varied text |
| Main risk | Rules go stale | Repeats past bias | Misreads text or invents field values |

Our practical advice: start with a rule-based score, because sales understands it and can correct it. Use the language model first only to turn free text into structured fields, and let the rules compute the score. A predictive score makes sense once a few months of clean data have built up.

At low volume, for example a [construction and building materials](/industries/construction-building-materials/) supplier receiving a handful of requests a week, fast and complete routing matters more than the score. In high-volume [ecommerce and retail](/industries/ecommerce-retail/), the real gain is separating wholesale or B2B requests from individual orders.

## How do form, WhatsApp and email leads reach the CRM in one flow?

Collection comes before scoring. If requests live in three places, even the best model works with half the picture. The fix is a [workflow automation](/glossary/workflow-automation/) that connects every channel to the same flow.

A typical flow has five steps:

1. **Capture.** The web form sends a webhook. The WhatsApp Business Platform notifies you of an incoming message. A shared sales inbox is read at regular intervals.
2. **Match.** The flow looks up the contact in the CRM by phone or email, so requests from different channels land on one record.
3. **Enrich.** A language model reads the free text and fills fields such as need, product group, urgency and location. Every field gets a "not stated" option so the model does not fill gaps by guessing.
4. **Score.** The CRM rules calculate the score.
5. **Route.** High scores reach the right sales rep as a notification. Low scores receive an automatic reply and a slower follow-up cadence.

You can build this with n8n, Make or Zapier. In Make, each flow is a "scenario" that starts with a trigger module. Whichever tool you choose, log every step so you can see where a lead was lost. The same structure connects naturally to [customer onboarding automation](/solutions/customer-onboarding-automation/): once a qualified lead becomes a customer, the welcome flow starts from the same record.

WhatsApp deserves a separate note. Meta requires businesses to get an opt-in before messaging people, and the opt-in must clearly name the business. A user writing to you first opens a conversation, but it is not enough for later campaign messages.

## What do GDPR Article 22 and KVKK say about automated decisions?

This section is not legal advice. Review your setup with your legal counsel.

GDPR Article 22 gives people the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal or similarly significant effects for them. If you receive requests from people in the EU, it applies to you.

[KVKK](/glossary/kvkk/), Türkiye's personal data protection law, gives people in Article 11 the right to object to an outcome against them that results from analysis carried out exclusively by automated systems. Article 10 adds a duty to inform at the point of collection, so the privacy notice on your form should also cover the fact that requests are classified automatically.

In practice this becomes three rules for lead scoring:

- The score sets priority. It does not exclude anyone.
- A person decides when a low-scoring request is closed or declined.
- Strip unnecessary personal data (ID numbers, full addresses) before text reaches the model, and record which region the model runs in.

## How does the sales team come to trust the score?

A score sales does not use is just a number in the CRM. Trust comes from two things: seeing the reasoning and being able to correct it.

**Reasoning.** Put a short explanation next to each score: "Fit high: corporate domain, target industry. Engagement medium: one visit, pricing page viewed. Text: three branches, decision by month end." Asking the model for the sentence each field came from makes this possible.

**Feedback.** A rep who disagrees with a score should flag it in one click. Review the flags monthly and update the rules. Tools such as HubSpot also support score decay, so an old visit does not count as much as today's request.

**Measurement.** Compare the score with outcomes. Do high-scoring leads turn into meetings and deals clearly more often than low-scoring ones? If not, the problem is the criteria, not the model.

## Where should you start?

The first step is an inventory, not a model: which channels requests come from, who records them where today, and which ones get lost. That picture often shows a collection problem to fix before any scoring.

At Corrplus we start with a discovery call. We then write the criteria down with the sales team, connect the channels into one flow and bring in the language model first only to read free text. We build in short cycles, and at the end of each one the sales team tries a working version on real leads. See our [AI sales assistant](/services/ai-sales-assistant/), [workflow automation](/services/workflow-automation/) and [customer experience and digital marketing](/services/customer-experience-digital-marketing/) services. In ecommerce, [ecommerce personalisation](/solutions/ecommerce-personalization/) covers what happens after a lead is qualified.

## Frequently asked questions

### How much historical data do I need for lead scoring?

None for a rule-based score: you start from criteria your sales team already knows. A predictive score needs enough won and lost examples. HubSpot's AI-generated scoring, for instance, asks for at least 50 contacts, 25 converted and 25 not converted.

### Can the AI reject a lead on its own?

Technically yes, but we advise against it. GDPR Article 22 and KVKK Article 11 limit outcomes based solely on automated processing that affect a person. Park low-scoring leads in a slower follow-up queue instead of deleting them, and let a person make the final decision.

### Can I follow up automatically on WhatsApp?

You can reply when the user messages you first. For later marketing messages, Meta requires an opt-in that clearly names your business. Record where and when the opt-in was collected in the CRM.

### Do I need a new CRM for AI lead scoring?

Usually not. Most CRMs, including HubSpot, Salesforce and Zoho, have score fields and a rules engine. The AI layer can sit in front of or beside the CRM, built with an automation tool such as n8n or Make.

## Sources

- [HubSpot Knowledge Base: Understand the lead scoring tool](https://knowledge.hubspot.com/scoring/understand-the-lead-scoring-tool)
- [Regulation (EU) 2016/679 (GDPR), EUR-Lex](https://eur-lex.europa.eu/eli/reg/2016/679/oj)
- [KVKK: Personal Data Protection Law No. 6698 (English text)](https://www.kvkk.gov.tr/Icerik/6649/Personal-Data-Protection-Law)
- [Meta for Developers: WhatsApp getting opt-in](https://developers.facebook.com/documentation/business-messaging/whatsapp/getting-opt-in)
- [Make Help Center: Scenarios](https://help.make.com/scenarios)