---
title: "How much CRM data do you need for predictive lead scoring?"
description: "The documented data minimums for predictive lead scoring: Einstein's 1,000-lead threshold, Dynamics' 40/40 rule, what the research used, and what actually determines whether your model works."
date: 2026-08-23
canonical: https://ax1om.ai/resources/how-much-crm-data-predictive-lead-scoring
---

Short answer: fewer records than you think, more conversions than you hope. Predictive lead scoring does not need big data. It needs enough examples of the outcome you want to predict, in fields populated before that outcome happened.

The longer answer is worth having, because "do we have enough data?" is the first question every team asks, and almost nobody answers with sources. The vendor thresholds are published. The research numbers are published. Here they are in one place, followed by what actually determines whether your model works.

## The documented thresholds

**Salesforce Einstein Lead Scoring: 1,000 leads, 120 conversions.** Salesforce's Spring '26 setup guide states the requirement for building a model on your own data: at least 1,000 leads created in the last 200 days, of which at least 120 converted, per segment. Below that floor, Einstein does not decline to score you. It quietly uses a global model built from anonymized data across many Salesforce customers, and keeps whichever model performs better. So the practical meaning of the threshold is not "scoring available: yes/no." It is "whose conversion patterns is the score based on: yours, or everyone's?"

**Microsoft Dynamics 365 Sales Insights: 40 and 40.** Microsoft's documentation sets the floor at 40 qualified and 40 disqualified leads created and closed within the training time frame. That is a much lower bar, and it shows in what the products are willing to predict: Dynamics predicts lead qualification, a faster and more frequent event than a closed sale.

**Practitioner folklore: about 20 closed deals is where expectation starts.** In Ops communities, the working rule of thumb is blunter: below roughly 20 closed deals, any model will struggle, and rules will beat it. The consensus in the how-to literature lands in the same place from the other direction: a tuned rules-based system stays competitive until somewhere between 500 and 1,000 closed outcomes, after which learned weights reliably beat hand-assigned ones.

## What the peer-reviewed case study used

The one peer-reviewed B2B case study worth reading, González-Flores et al. (2025) in Frontiers in Artificial Intelligence, trained on 16,600 cleaned CRM records with 11.84% positives, roughly 1,960 positive examples, gathered over four years. We reviewed that study in detail, including the questions its near-perfect AUC should raise, in [our journal-club review](/blog/predictive-lead-scoring-peer-review). For this question, the relevant point is simpler: a real mid-size B2B company accumulated about 2,000 positive outcomes in four years of normal operation. Volume was never their constraint. Field hygiene was.

## Why conversions are the number that matters

Notice that every threshold above is really two numbers, and the binding one is never total records. It is the count of the rarer class: conversions.

A model acquires the difference between converting and non-converting leads by seeing examples of both. Non-converting examples are abundant. Converting examples are scarce, and they carry nearly all of the information. Ten thousand leads with 30 conversions is a worse training set than 800 leads with 200 conversions.

A useful planning heuristic from the statistics literature is events per variable: roughly 10 to 20 positive examples for every feature the model seriously considers. With 120 conversions, a model can support around 6 to 12 real features. That is not a limitation to fight. It matches what the research found anyway: a handful of fields carry most of the signal.

Two more things move the effective requirement up or down:

- **Field population.** A field empty on 40% of records contributes less than its importance chart suggests. Sparse data raises the number of records you need, sometimes steeply. That is why we surface [dataset health](/learn/dataset-health) before training, not after.
- **Leakage-free fields only.** The count that matters is conversions with honestly usable fields. If your strongest predictors turn out to be [fields written after the outcome](/learn/target-leakage), your effective dataset is smaller than it looks, and your evaluation numbers are larger than they deserve to be.

## What history length does and does not buy you

Four years of history sounds strictly better than one. It is not always. Conversion patterns drift: pricing changes, ICP changes, rep turnover, a pandemic. Old outcomes teach the model a market that may no longer exist, which is why validation should run forward in time and why retraining on a schedule matters more than maximizing history depth.

The practical guidance: bring all the history you have, but expect the model, and any honest vendor, to weight recent patterns and to tell you when older data stops helping.

## If you are under the floor today

Below roughly 50 conversions, the recommendation is:

1. **Run simple rules on fit,** and keep the rule count small enough to argue about in one meeting.
2. **Fix the fields you will need later.** Source, consistently tagged. A conversion outcome that is actually recorded. Timestamps preserved. The teams that get value from ML scoring in year two are the ones that cleaned intake fields in year one.
3. **Define the conversion event precisely.** Qualification, opportunity created, and won business are different targets with different data requirements and different meanings. The Einstein threshold counts conversions to account and contact; the research study predicted qualification; what pays your bills is the win. Decide which one you are collecting toward.

ax1om trains on your own conversion history, wins and losses alike, and works from as few as 50 converted opportunities, sharpening as data grows. Before training, it checks [how many records and conversions you actually have](/learn/records-trained), screens fields for leakage, and reports dataset health, so "do we have enough data?" gets answered by inspection rather than by hope. If not, it tells you that too.

The threshold question has a documented answer. The better question, the one the thresholds are a proxy for, is whether your CRM has been recording the outcome you care about, in fields that existed before the outcome did. Get that right, and the record count takes care of itself.

---

**Sources:** Salesforce, "Set Up Sales Cloud Einstein," Spring '26 setup guide · Microsoft Learn, "Configure predictive lead scoring," Dynamics 365 Sales · González-Flores, L., Rubiano-Moreno, J., & Sosa-Gómez, G. (2025). *Frontiers in Artificial Intelligence*, 8, 1554325.