AI Lead Qualification Workflow: Prompts, Scoring Rules and CRM Automation Templates

Overview

AI lead qualification works best as a decision-support workflow rather than a fully autonomous sales gate. The automation should collect and clean inbound information, apply your documented qualification rules, produce a concise explanation, and recommend a route. A person can then approve uncertain or high-value cases before the CRM owner, priority, or follow-up process changes.

A practical workflow has six stages:

1. Capture: receive a website form, email, chat transcript, event registration, or other inbound signal.
2. Normalise: standardise names, company details, job titles, regions, phone numbers, and campaign information.
3. Enrich carefully: use only approved business data sources and label information that was inferred or supplied by a third party.
4. Qualify: ask an AI model to extract buying signals and apply a fixed scoring rubric.
5. Summarise: create a short explanation that preserves the prospect’s words, needs, timing, and open questions.
6. Route and review: update the CRM, recommend an owner or queue, and send uncertain cases to human approval.

Define the rules before writing the prompt. Your ideal customer profile should specify the customers you can serve, the use cases you support, the regions or segments you cover, and the signals that justify prompt follow-up. Keep these rules in a versioned document or configuration table, not only inside an automation step. That makes seasonal planning and later calibration easier.

For broader implementation guidance, compare this workflow with the AI automation testing and monitoring checklist and the AI automation security checklist for business workflows.

Checklist by scenario

Scenario 1: Website form submission

• Capture the submission ID, timestamp, source page, campaign, consent status, and all form fields.
• Check whether the email address or company already exists in the CRM before creating a new record.
• Preserve the original message in a separate field so the AI summary can be audited.
• Ask the model to distinguish explicit answers from assumptions. An empty budget field should remain unknown, not become “low budget.”
• Write the score, score band, reasoning, recommended owner, and follow-up questions to mapped CRM fields.
• Send incomplete, contradictory, or high-impact records to a review queue.

Scenario 2: Inbound sales email

• Extract the sender, company domain, subject, message body, thread identifier, and attachments that the workflow is permitted to process.
• Remove signatures and repeated quoted text where possible, while retaining the current request.
• Match the sender to an existing contact without treating a shared email domain as conclusive proof of identity.
• Ask the model to identify the requested outcome, urgency, product or service area, stakeholders, and unanswered questions.
• Keep the original email linked to the CRM record. The generated summary should aid triage, not replace the source.

Scenario 3: Campaign or event leads

• Store the campaign, event, landing page, and consent context as first-class fields.
• Do not treat attendance, a download, or an email click as proof of buying intent. Use these as signals with documented weights.
• Apply a segment-specific rubric if the campaign targets a different industry, company size, or use case.
• Set a follow-up window appropriate to the campaign and include the reason in the task or notification.
• Review a sample of routed records after the campaign to identify false positives and false negatives.

Reusable lead qualification prompt

Place the stable instructions in the system or instruction field and pass the lead data as a clearly separated input. Replace the bracketed values with your own rules:

You are a sales operations assistant. Qualify the lead using only the supplied data and the rubric below.

Ideal customer profile:
- Supported customer types: [describe]
- Supported use cases: [describe]
- Regions or segments served: [describe]
- Disqualifiers: [list]

Scoring rubric:
- Fit: 0-30 points
- Need: 0-25 points
- Timing: 0-20 points
- Authority or buying involvement: 0-15 points
- Engagement or next-step signal: 0-10 points

Rules:
1. Do not invent facts, company details, budget, authority, timing, or intent.
2. Mark missing information as unknown.
3. Separate evidence from interpretation.
4. Apply disqualifiers only when the supplied evidence supports them.
5. If the input is contradictory, lower confidence and request human review.
6. Recommend a route; do not claim that a person has been contacted.
7. Return valid JSON matching the requested schema and no extra commentary.

Lead data:
[insert structured lead data]

Return:
- the extracted qualification fields
- points by category
- total score and score band
- evidence for each material decision
- confidence level
- recommended CRM route
- follow-up questions
- a concise sales handoff summary

Example scoring rubric

A weighted rubric is easier to audit than a general instruction such as “decide whether this is a good lead.” Adjust the weights to your sales process:

Set score bands only after agreeing on the operational response. For example, a high band might create a sales task, a middle band might enter a review queue, and a low band might receive a non-urgent nurture route. The labels are less important than the action, owner, and service expectation attached to each one.

JSON output schema

Structured output makes the automation easier to map and validate. A simplified schema might look like this:

{
  "lead_id": "string",
  "fit": {"score": 0, "evidence": ["string"]},
  "need": {"score": 0, "evidence": ["string"]},
  "timing": {"score": 0, "evidence": ["string"]},
  "buying_involvement": {"score": 0, "evidence": ["string"]},
  "engagement": {"score": 0, "evidence": ["string"]},
  "total_score": 0,
  "score_band": "high|medium|low|review",
  "confidence": "high|medium|low",
  "recommended_route": "string",
  "follow_up_questions": ["string"],
  "handoff_summary": "string",
  "needs_human_review": true,
  "rubric_version": "string"
}

Validate numeric ranges, required fields, allowed labels, and maximum summary length before updating the CRM. If validation fails, log the error and route the record to an operator instead of silently writing partial data.

Zapier and Make.com mapping checklist

1. Trigger: choose the form, email, inbox, event, or webhook event and capture its unique ID.
2. Search: look up the contact and account using approved matching fields. Define what happens when there are multiple matches.
3. Prepare: map source data into a compact object with clear field names and remove unnecessary sensitive content.
4. AI step: pass the rubric version and structured lead object to the model. Require the agreed JSON format.
5. Validate: check the response for parse errors, missing values, score limits, and review flags.
6. CRM update: write to dedicated fields such as AI score, qualification band, evidence, summary, route, confidence, and processed-at time.
7. Approval path: pause or notify a named reviewer when confidence is low, data conflicts, or a disqualifier is detected.
8. Notification: send only the context needed for the next action, with a link to the CRM record and original source.
9. Logging: retain the workflow version, prompt version, model configuration, outcome, and error state according to your internal retention rules.

What to double-check

Data quality: Confirm that required form fields, email parsing, deduplication, and company matching behave as expected. Test blank, very long, multilingual, and malformed inputs.

Scoring consistency: Give the workflow a small set of previously reviewed examples. Compare its category scores and routes with human decisions, then document why disagreements are acceptable or require a rule change.

Context preservation: A short summary is useful only if it retains the prospect’s problem, desired outcome, timing, and next question. Keep a link or reference to the source record.

Privacy and access: Send only necessary information to the selected automation and AI services. Review connector permissions, CRM field visibility, retention settings, and internal approval requirements using your organisation’s policies. For technical implementation patterns, see the OpenAI API integration guide for business automation.

Human ownership: Every route should have a named queue or owner. “Needs review” is not a useful outcome if nobody is responsible for reviewing it.

Operational limits: Check the current limits, retry behaviour, rate controls, payload sizes, and connector capabilities of the tools you use. These details can change, so avoid hard-coding assumptions into the prompt.

Common mistakes

• Using a vague prompt: “Is this a good lead?” produces an opaque decision. Give the model explicit categories, weights, evidence rules, and output fields.
• Scoring confidence as intent: High confidence means the input is clear, not that the prospect is ready to buy. Store these concepts separately.
• Inventing missing data: A job title or company domain may suggest a possibility, but it does not prove authority, budget, or fit.
• Overwriting the CRM: Keep AI-generated fields separate from salesperson-owned fields until the process has been tested and approved.
• Ignoring duplicates: Creating a new contact for every form or email can fragment history and send conflicting assignments.
• Routing without an explanation: A score without evidence makes correction difficult. Store the material signals behind the recommendation.
• Automating the exception away: Contradictory, sensitive, or high-impact cases should have a visible human path.
• Changing the rubric silently: Version the scoring rules and include that version in each result so later reviews can distinguish process changes from model variation.

When to revisit

Review this AI lead qualification workflow before seasonal planning cycles, after changing your ideal customer profile, and whenever a new campaign targets a different segment. Revisit it when sales ownership changes, CRM fields are renamed, a connector or model is replaced, or lead volume makes the existing review queue impractical.

Use outcomes, not only model output, to recalibrate the rubric. Sample routed leads and compare the recommendation with salesperson feedback, accepted opportunities, disqualified records, and leads that received no timely follow-up. Look for repeated errors: a campaign that inflates engagement scores, a region that is incorrectly disqualified, or a role that is mistaken for buying authority.

Before each update, run the workflow against a fixed test set containing clear, incomplete, contradictory, duplicate, and disqualified examples. Record the prompt version, rubric version, mapping changes, and approval decision. Then publish the change only after confirming that CRM updates, notifications, retries, and review routes still work.

Start with one controlled source, such as a website form, and a small set of CRM fields. Once the scores and routes are understandable, add email, event, or chat inputs one at a time. This keeps the automation measurable and gives the sales team a practical way to correct the system as its rules evolve.