ML Pipeline Health Scoring for a Mid-Market Sales Team
Forecast accuracy was stuck at 58% because pipeline reviews relied on rep intuition and manager judgment. A machine learning model scoring each opportunity on 12 objective signals improved accuracy to 83% and identified $1.4M in at-risk deals before they slipped.
Cost of problem
$1.4M in undetected at-risk pipeline per quarter
Savings achieved
$1.4M in at-risk deals identified and recovered
Time saved
60 min/week per manager in pipeline review prep
ROI
6x ROI in year one
Build time
8 weeks
The Problem
Weekly pipeline reviews were theater. Reps sandbagged or bluebirded depending on quota pressure. Managers added their own gut-feel adjustments. The VP of Sales was running three parallel spreadsheet models trying to get to a number she could defend to the board. Forecast accuracy averaged 58% over 6 quarters — meaning nearly half of 'committed' deals either didn't close or closed at different values.
What Was Built
A Python ML model (gradient boosted trees via XGBoost) trained on 3 years of historical opportunities in Salesforce. Features include stage velocity, activity recency/frequency, multi-threading depth, champion engagement score, and competitive signals from call notes (via lightweight NLP). The model outputs a 0-100 health score and a risk flag per opportunity, written back to Salesforce as custom fields visible in the forecast view.
How It Was Implemented
Training data was pulled via Salesforce SOQL, cleaned to remove data entry anomalies, and enriched with CRM activity history. The model was validated on a held-out 12-month window (walk-forward validation to avoid lookahead bias). The scoring pipeline runs nightly via a cron job and updates Salesforce in batch. Feature importance is exposed to RevOps managers so they can explain why specific deals are flagged.
The Outcome
Forecast accuracy improved from 58% to 83% within two quarters of deployment. The model flagged 23 deals as high-risk in Q1 that had been marked 'commit' by reps — 19 of them slipped or churned. Proactive intervention on 4 of those deals saved an estimated $1.4M in at-risk ARR. Pipeline review meetings dropped from 90 minutes to 30 minutes because managers came in with data, not opinions.