Every sales leader I talk to has the same frustration.

Your CRM is full of data. Call notes. Email threads. Follow-up records. Meeting outcomes. Win-loss reasons. All of it sitting in text fields that every lead scoring system ignores.

But your current lead scoring system cannot read any of it.

Why This Paper Fills a Blind Spot

E-commerce recommendation is simple: users see something, click or buy, done. High-stakes sales is different. Automotive. Real estate. Enterprise B2B. The decision cycle runs weeks or months. A lead touches multiple sales reps, multiple test drives or demonstrations, and multiple follow-ups. The conversion signal is spread across dozens of unstructured CRM entries, not a single click.

Three problems have made AI lead scoring fail in these domains until now: sparse supervision signals, a semantic gap in unstructured CRM logs, and the inability to capture relative lead priority. Sales reps don’t think in absolute probabilities. They think in comparisons: “this lead is hotter than that one.”

The LLM-based hierarchical preference framework solves all three at once. It reads call notes. It understands that a lead in “test drive completed” matters differently than one in “initial inquiry.” And it learns from pairwise comparisons — the same mental model sales reps use when they tell each other who to call first.

For a Chief Revenue Officer or VP of Sales, this is the first paper that takes their actual problem seriously instead of treating sales like e-commerce with longer time horizons.

Methodology, Explained Simply

The core question the model answers is not “will this lead convert?” but “should you call this lead before that one?” The model learns from pairwise preferences: given two leads, it predicts which is more likely to convert. This mirrors exactly how sales reps actually prioritize their pipeline.

The funnel is hierarchical, not flat. A lead at “test drive completed” and a lead at “initial online inquiry” are in fundamentally different states. The model scores leads within each funnel stage separately, then propagates preference information across stages. A lead in later stages naturally ranks higher, but within each stage, the model identifies the strongest candidates.

The LLM processes unstructured CRM data that traditional models ignore: call notes, email threads, meeting summaries, follow-up records. No special cleaning required. No manual tagging. The raw text your reps already type into Salesforce or HubSpot becomes the training signal.

The training happens on your existing CRM data. The 25,000-lead automotive dataset in the paper used pipeline stages, call records, and conversion outcomes that any company with a CRM already logs.

Results and Practical Insights

The advantage persists across rank thresholds. At NDCG@3 (top three leads) and NDCG@5 (top five), the model maintains its lead over every competitor. This is not a case of one great prediction and mediocrity after. The model ranks consistently well across the top of the funnel.

The ablation results confirm both innovations independently matter. Removing the hierarchical funnel design drops NDCG by 0.024. Removing pairwise preference training drops it by 0.033. Together, the combination drives the full 8.3 percent improvement.

Key Takeaways for Revenue Leaders

Thanks to All Authors