200+ BDRs save 2-4 hours a week.
GTM analysis time cut 50%.

The best upsell signals lived on data nobody could move.

The signals that mattered most — usage patterns, transaction volumes, feature adoption — sat inside client environments that were extremely confidential and could not be moved to a central location for training. Some of that data lived in air-gapped segments under strict regulatory control.

CRM was no help either. Records were inaccurate, which ruled them out as training data. That left the BDR team with no reliable way to see which of 10,000+ accounts were actually ready for an upsell conversation.

Three constraints were non-negotiable:

• Confidential data could not be moved off-site
• Air-gapped environments with strict regulatory controls
• CRM data too inaccurate to train on

A model trained where the data lived, not where it could be moved.

The team built a federated learning architecture: models trained locally at each data-holding site, with only model updates — never raw records — shared onward. The whole system was deployed into an air-gapped environment, satisfying the strictest security requirements in the sector.

Access ran on strict rule-based controls, with comprehensive logging of every user action. Because CRM couldn’t be trusted, the team reasoned from first principles instead — identifying the indicators that actually signal upsell readiness, then training the model on target customers who showed those patterns.

The result: a working system that surfaced upsell candidates to the global BDR team without a single confidential record leaving its network.

200+ BDRs, hours back every week.

The federated approach delivered real productivity gains without touching the constraints that mattered most. BDRs got hours back every week, the GTM team cut its reporting load in half, and confidential data never left the network it started in.