Introduction
Enterprise B2B sales teams have long struggled to operationalize engagement data at scale. Most GTM systems still rely on contact-level lead scoring models—often outdated, narrowly trained, and poorly aligned to how enterprise buying actually works. But as our buying groups become more complex and multi-threaded, it’s become clear: we need to model the group, not just the lead.

At Kinetik, we’ve been experimenting with a modeling technique known as Multi-Instance Learning (MIL) to solve this challenge. Unlike conventional supervised learning that treats each contact in isolation, MIL allows us to classify an entire buying center based on aggregated behavior across all its members. The results have been promising. Here’s what we’ve learned so far.

Why Traditional Lead Scoring Fails in Enterprise B2B
Legacy lead scoring models were designed for high-volume, transactional sales environments. In enterprise B2B tech, however, opportunities typically involve:

• Multiple stakeholders across functions
• Dozens of touchpoints across weeks or months
• Varied types of content, channels, and campaign sources

This diversity cannot be captured by one lead’s score. In many cases, no single individual ever exhibits the “perfect” score—but the group, in aggregate, signals strong buying intent.

How MIL Solves the Problem
Multi-Instance Learning flips the structure of traditional supervised learning. Instead of learning from labeled instances, the model learns from labeled bags of instances.

In our application:

• Each “instance” is a contact-level touchpoint (e.g., a webinar, content download, MQL)
• Each “bag” is a buying center—a primary CSN and its related CSNs
• The label is a classification of that group’s aggregate engagement: High, Medium, or Low

This allows the model to identify latent patterns that emerge only when multiple contacts interact with marketing and sales over time.

Engineering the Engagement Cube
To train the model, we structured buying center behavior into a three-dimensional grid:

• Interest: Recent, high-quality engagement (e.g., demo requests, in-region webinars). Measured using a weighted activity score emphasizing recentness and directness of engagement.
• Intent: Sustained behavioral signal strength (e.g., number of MQLs, response frequency). Computed over a rolling 90-day window to reflect sustained intent.
• Ideal Prospect: Fit to win (e.g., industry, historical win rate, company size). Derived from firmographic alignment, past conversion rates, and opportunity velocity.

Each dimension is independently scored and normalized, then segmented into High, Medium, or Low based on percentile thresholds across the account universe. This creates a 3x3x3 classification cube (27 segments), allowing GTM teams to prioritize and tailor outreach based on a buying center’s precise position within the cube.

This framework enables scenario-specific GTM motions. For example, a group with High Interest, Medium Intent, and Low Fit may trigger a fast-lane nurture path focused on qualification, while a High-High-High profile may go directly to seller outreach.

Lessons from the Field
Through our experimentation, we’ve learned:

• Data architecture matters: Without structured buying center aggregation—linking primary and related CSNs based on match scores—the model cannot converge on useful insights. We used a match threshold of 70+ to define valid related CSNs.
• Not all signals are equal: Simple activities like email opens often mislead the model. Instead, we emphasized high-conversion signals such as demo requests, pricing page visits, event attendance, and custom ABM campaign engagement. These outperformed by a factor of 10–20x in predictive lift.
• Classification works best for interpretability: We initially tested a regression model that produced continuous scores, but classification proved far more actionable for GTM users. The 27-cube segmentation was far easier for SDRs and RevOps teams to operationalize.
• Subset training accelerates iteration: We trained on representative CSN subsets (∼300 groups) to shorten runtime from 7 days to ∼6 hours, enabling daily model evaluation. This agile experimentation loop allowed us to tune scoring logic and feature weights quickly before scaling.
• Cross-functional input is essential: Feature engineering benefited significantly from expert feedback on what behaviors actually signal buying readiness. Purely data-driven modeling underestimated GTM context without human guidance.

Next Steps and Broader Implications
We’re now scaling our MIL implementation to support end-to-end journey mapping, engagement scoring, and even predictive alerts based on cluster movement across the cube. In parallel, we are layering role inference models to bring more precision to contact-level orchestration.

Enterprise GTM teams don’t just need more data. They need models that respect how buying happens.

Multi-Instance Learning has brought us closer to that truth—and it’s only the beginning.

If you’re working on engagement scoring or buying group orchestration in your GTM engine, I’d love to connect. Let’s compare notes.