B2B marketers have been hearing a lot about A.I. recently. Yet A.I. has been a part of marketing technology for several years already and has simply been known by another name: Predictive Lead Scoring.

Now that marketers are starting to demand more transparency into the inner workings of predictive solutions, it’s important to understand the basic principles of A.I. so you know how and when to deploy predictive lead scoring.

Predictive lead scoring is more than a crystal ball

Predictive lead scoring depends on the same type of deep learning algorithms – i.e., A.I. – that now power auto-tagged photos, speech recognition, and even self-driving cars. These algorithms, in turn, rely on massive quantities of high-quality data, which is why A.I. and predictive lead scoring weren’t possible until several years ago — there simply wasn’t enough data in the world to train algorithms to make accurate predictions.

Predictive lead scoring algorithms - B2B data

A large quantity of high-quality data is essential for A.I. and predictive lead scoring because machines – like humans – learn from data.

For example, a machine learns to tag photos correctly when it’s presented with labeled photographic data.

Predictive lead scoring algorithms - Dog or muffin

If we were interested in teaching a machine how to tag dog photos, we would label the left photo as “dog” and the right photo as “no dog.” A powerful machine would then be able to identify the characteristics of each photo – the colors, shapes, and textures that are present – and use that information to predict whether other photos have a dog in them.

Undergoing this process with only two photos, however, is inherently limiting: There are not many patterns for a machine to observe before determining how to predict whether other photos have dogs in them or not.

Using only the two photos above, in fact, a machine might mistakenly assume that dog photos tend to have three dark dots within a round, lighter-shaded region.

Predictive Lead Scoring Algorithms - Dog or Muffin photos

As you can see in these two photos, however, not all photos with three dark dots within a round, lighter-shaded region are dogs. Sometimes, those photos are other things – like muffins.

For a machine to learn how to tag dog photos correctly, it needs a large number of labeled images representing the many different types of scenes that do and do not contain dogs.

Predictive Lead Scoring Algorithms - Dog or Muffin - More data

In addition, the labeled photos that are presented to a machine need to be high resolution.

Predictive Lead Scoring Algorithms - Dog data

A photo like this one simply does not contain enough detail to help a machine understand what shapes, colors, and textures tend to be present in a photo containing a dog.

Predictive lead scoring prerequisites

The process is quite similar when training a machine to recognize future customers in a marketing database rather than dogs in a set of photos.

Instead of loading photos into a machine, this process begins with loading accounts that are labeled as “wins” or “losses.” Then, leveraging algorithms that are similar to those used for identifying dogs in photos, a machine will be able to look at other accounts and identify future customers. To make these predictions actionable for marketers and salespeople, the machine will then assign predictive lead scores that represent the likelihood that each account is a “win.”

This process has the same prerequisites as photo tagging. First, it requires a lot of data — only over the course of thousands of accounts will patterns become evident that a machine can use to predict whether future accounts will become customers. Second, it requires high-resolution data that reveals granular patterns that can be truly explanatory for predicting customers.

If either condition is missing, however, predictive lead scores will likely lead you astray.

Dog days of predictive lead scoring

That is why many early attempts at predictive lead scoring failed for B2B marketers: the data they used for training their models was often limited in quantity or quality.

Companies attempting to build predictive lead scoring models in-house often encountered limitations because of low fill rates in their CRM and marketing automation systems. Similarly, providers of outsourced predictive lead scoring often struggled to assemble rich datasets when pulling together disparate information on-demand for clients seeking to supplement their internal data. As a result, high-resolution account attributes such as tech stacks, revenue bands, and revenue/headcount growth trajectories could not be factored into predictive lead scoring.

Furthermore, companies were often attempting to predict their next customers when only a small slice of their sales history was present in their CRM. Training a model based on a handful of “won” accounts could not provide a complete picture of the variables influencing which companies were likely to buy, resulting in many missed predictions — just like how training a machine to tag dogs based on just a handful of reference photos could very easily lead to a machine thinking that muffins are dogs.

These factors combined to cause many early predictive lead scoring deployments to fail.


Armed with knowledge about the underlying principles of predictive lead scoring and A.I., B2B marketers can embark on the journey of vendor selection and implementation with more confidence about when and how they can benefit from the technology.

Now that you understand more about what’s happening behind the curtain of predictive lead scoring, read our playbook to learn how you can use predictive lead scoring to run high-converting campaigns.