A lead score is not a truth about a person or company. It is a compressed prediction for one named decision. Without that decision, the number is decoration. With it, the score becomes a model that can be tested, calibrated and switched off
Building a small model
Suppose a software company wants to decide which enquiries receive specialist review within two hours. Historical cases suggest four useful features:
- problem fits the services offered: 0, 20 or 40 points
- a responsible decision maker is involved: 0 or 20 points
- realistic start within six months: 0, 10 or 20 points
- a concrete operational reason is described: 0 or 20 points
An enquiry with 70 points gets rapid review. This looks orderly, but it is only a hypothesis. The weights come from experience, not natural law. They must be tested against real outcomes
Defining a correct outcome
“Deal won” may be months away and depends on price, sales execution and capacity. “A suitable specialist meeting takes place” can be a better first target. It is closer to the decision and can be measured sooner. Later, the team can test whether that intermediate result actually relates to good projects
The target cannot be changed retrospectively to make the model look good. Its definition, period and exceptions should be recorded before evaluation
Two error types, two costs
A false positive receives rapid attention even though it does not fit. That costs time. A false negative waits even though it is valuable. That can cost revenue and trust. The threshold of 70 is therefore not a purely statistical choice. It depends on team capacity and the cost of each error
If 40 of 100 highly ranked enquiries really fit, precision is 40 per cent. If 50 suitable enquiries existed overall and the model found 40, recall is 80 per cent. Both measures matter. A system that marks only one very safe lead can have perfect precision while delivering almost no value
Calibration makes numbers honest
If enquiries scoring 80 fit in four cases out of five, the number has a useful interpretation. If only half fit, the scale looks more precise than the model. Calibration groups scores and compares prediction with reality
With small datasets, broad classes are often more honest: high, medium and low priority. A decimal place does not create insight. It only hides uncertainty
Missing data is not zero
“Budget unknown” must not automatically mean the same as “no budget”. Missing information can have its own value or trigger a later question. Otherwise the system penalises early exploratory enquiries and favours people who already speak the supplier’s language
Company size is rarely a neutral substitute for potential. It can systematically disadvantage particular sectors or young businesses. Every feature needs a plausible business relationship, not merely a chance correlation in a small dataset
An auditable operating sheet
- name the decision and target variable
- document the business reason for every feature
- store score, rules and version for each enquiry
- allow human overrides with a reason
- review precision, recall and calibration regularly
- check relevant groups for unusual differences in outcome
- define fallback behaviour if data or model fails
When machine learning does not help
With only a few hundred inconsistently recorded enquiries, a complex model often learns the habits of individual salespeople. A small rules model is more transparent and easier to change. Machine learning becomes interesting when there are enough clean examples, a stable target and a process for continuous monitoring
The lead qualifier gathers information; the score prioritises one clear next step. Neither should reject people finally unless its impact, errors and route of appeal are understood. The number is a working instrument. Accountability remains with the people who design the sales process