Lead scoring with AI Builder

AI Builder lead scoring is the embedded machine-learning model in Dynamics 365 Sales that ranks leads by their likelihood of converting to a qualified opportunity. It exists because the manual lead-scoring rules sales operations teams traditionally build — points for industry, deductions for tiny accounts, points for whitepaper downloads — are slow to adjust and quickly become folklore. The AI model learns from outcomes and updates as the data evolves.

Training. The model is trained on the customer's own historical lead data. It looks at the features of leads that did or didn't convert — demographic (industry, size, location), firmographic, behavioural (web visits, email opens, form submissions), and source attribution — and finds patterns. Training is automatic and re-runs on a schedule, so the model stays current.

Features. Out-of-the-box features include standard lead fields, related contact and account attributes, and connected behavioural data from Customer Insights (where licensed). Customers can add custom fields to the feature set and exclude features that introduce noise.

Scoring. Once trained, the model scores every active lead and writes the score back to the lead record. The score is a value 0–100 with an explanation: which features contributed positively, which negatively. Scores refresh continuously as the lead's underlying data changes.

Grades and tiers. Scores are bucketed into A/B/C/D grades for triage. Sales operations decides the thresholds. Grades are what most users actually look at; the raw score is for analysis.

Routing. Routing rules can use the score or grade — A-grade leads route to senior reps, C-grade to nurture. Combined with the sales accelerator, scored leads land in seller queues in priority order automatically.

Model accuracy. The model reports its own accuracy (precision, recall, AUC) against held-out data, with the ability to drill into false positives and false negatives. Below an accuracy threshold, the model marks itself as low-confidence and rounds scores conservatively.

Operating discipline. Two pieces of discipline make the model work well:

1. Outcome data must be reliable. Lost leads with no reason marked, won deals not tied back to a lead — both pollute training.
2. Periodic review. Once a quarter, review which features the model is using and whether they match how the business actually wins. Models drift when the world changes (a new product line, a new geography, an economic shift) and need explicit attention.

Where it fits. AI Builder scoring is suitable for mid-to-large B2B businesses with at least a few thousand historical leads. For very small datasets, manual scoring rules outperform.