How to design AI lead qualification without losing control

Qualification begins with a commercial policy

AI cannot repair an undefined sales policy. Before scoring anything, write down which problems the team can solve, what evidence indicates fit, which projects are excluded and who decides ambiguous cases.

Separate hard rules from judgment. A prohibited geography, unsupported stack or missing legal basis can be deterministic. Understanding whether a vague brief is close to your experience may benefit from a language model, but the model should explain the match.

Preserve source and raw context

Every lead should retain a permitted source identifier, capture time and link to the original request. Normalized fields help reporting, but they should not erase the context a person needs to verify the result.

Retention must follow the source terms and privacy requirements. Some marketplace data can be cached only briefly, while your own notes and actions may be stored separately.

• Source and external identifier
• Published or received time
• Normalized budget, location and required skills
• Original link and retention deadline
• Internal status, notes and owner

Use a two-stage score

First apply deterministic validation: duplicates, explicit exclusions, missing required fields and source freshness. Then ask AI to classify the remaining request against a compact rubric.

Store the category, score and short reasons separately. A score without reasons is difficult to correct, and long unstructured reasoning is expensive to review. The owner should be able to change the rubric and compare new results with previous decisions.

Measure errors, not only average score

Review false positives that waste attention and false negatives that hide valuable work. Build a small evaluation set from accepted, rejected and ambiguous leads and rerun it when the prompt, model or policy changes.

Track cost per analyzed lead and per qualified lead. Token cost can be reduced by normalizing text first, sending only useful fields and using a cheaper model for obvious classifications.

Keep the external reply manual

Qualification and drafting can save time without giving the system authority to contact a client. A useful review card contains the source link, match reasons, risks, suggested next step and a draft that cites only verified experience.

The person sending the message remains responsible for price, availability, claims and platform rules. This also creates feedback that improves the qualification policy.

Start with a manual check button

A pilot should begin with one official source and a manual run. Examine the first 50 to 100 results, correct exclusions and confirm retention behavior before enabling a schedule or Telegram alerts.

Background monitoring is the last step. It should be low frequency, rate-limit aware and easy to disable without losing the internal review history.