AI lead scoring: data, testing and human oversight

A lead score helps organise sales follow-up. It does not prove that someone will buy. Before using AI, define the outcome you want to predict: qualification, an appointment or a sale within a stated period. Compare the model with a simple rule your team understands.

1. Distinguish fit from interest

Fit concerns the profile: need, sector, relevant offer and ability to work together. Interest concerns observed signals: an explicit enquiry, visit or reply. Repeat visits are not confirmed buying intent; a suitable prospect may show no digital signal.

Document criteria and limitations with sales staff. Avoid automatically rejecting enquiries because they differ from historical customers. Check what the data represents and whom it covers poorly.

2. Prepare usable historical data

Correct duplicates, incomplete statuses and dates. Define an observation period and target outcome. A late sale might be incorrectly labelled a failure if the window is too short. Requirements vary by software and model; check versions and vendor documentation.

Data leakage happens when the model receives information unavailable at the decision point. Do not use a final sale status to predict the outcome before contact, for example. Test on a later period than the one used to build the model and examine market changes.

3. Assess errors that matter to the business

Select a threshold compatible with follow-up capacity. For a sales target, a false positive is a prioritised prospect that does not buy within the stated period; a false negative is a sale the score failed to prioritise. Definitions change when the target is an appointment or qualification.

Examine the precision of priorities, missed opportunities and differences between segments. Compare with the simple rule and work without scores. An overall indicator may hide errors for a language or new offer. Record limitations before deciding to deploy.

4. Define CRM actions

A score can prepare a task or suggest priority. Specify its owner, timeframe and how to correct a status. Plan for missing data, delayed calculations and failed synchronisation. Staff should be able to check and challenge recommendations.

Do not automatically add every prospect to an email sequence. Check contact rules, preferences and objections. A low score does not justify indefinite retention. Define suitable retention and deletion procedures.

5. Govern personal data

The FDPIC explains that the FADP applies to personal data processed by AI. Check purposes, information supplied to people, access, contracts, retention and transfers. Human involvement alone does not guarantee compliance.

An impact assessment is required when processing may create high risk under the FADP conditions. Also assess whether rules on automated individual decisions or other markets apply to your use. Do not assume every sales recommendation falls under identical obligations.

6. Organise the pilot and 2027 reviews

Fictional example: an SME tests a score in observation mode without changing enquiry assignment. Each week, the team compares suggested priorities with observed qualifications and records errors. It also checks French and English enquiries. The test may lead to changed criteria or abandoning the model; it promises no additional sales.

For 2027, assign a model owner and set reviews, stop thresholds and a procedure for returning to a simple rule. Connect monitoring to defined outcomes rather than the number of scores calculated.

Continue with our guide to chatbots for sales enquiries.