Lead scoring and churn prediction in 2026: what the evidence supports, and the failure mode nobody demos

There is no credible published figure for what predictive lead scoring does to conversion. What does exist is a documented case of a churn model with 0.90 AUC whose precision collapsed to 0.03 in weekly operation. Here is what actually decides it.

If you are being sold a conversion uplift figure for predictive lead scoring, ask where it came from. We went looking for a credible one and could not find a single uplift statistic in the public record that traces to research rather than to content marketing.

That absence is the most useful thing to know before funding this work, because it means the case has to be built on your own numbers.

What the evidence actually supports

What the evidence on lead scoring actually supports

Sorted by whether the claim survives tracing to a primary source.

• AI-enabled next best actions correlate with commercial growth. 2.6x more likely, Gartner survey of 227 chief sales officers, May 2026. A correlation in a cross-sectional survey, not a measured uplift.
• Data quality is the binding constraint. Only 21% of marketers say CRM data is very well prepared to support AI; 62% lost revenue directly to poor CRM data quality.
• Churn models degrade without retraining. Documented in a 2025 peer-reviewed study on 8,878 real B2B customers.
• Predictive lead scoring increases conversion by a specific percentage. No verifiable figure exists in the public record. Every circulating number traces to content marketing.
• You need a minimum number of churn events to model. No authoritative published threshold exists. Class imbalance and drift matter more than row count.

Source: Gartner, May 2026; Validity State of CRM Data Management, August 2026; Oliveira et al., ICEIS 2025.

The strongest published finding is a correlation, not a causal uplift. Gartner's May 2026 survey of 227 chief sales officers found organisations providing sellers with AI-enabled next best actions were 2.6 times more likely to achieve commercial growth. Useful, and worth knowing, but it does not tell you what a model would do for you.

The failure mode nobody demos

A 2025 study built a churn model on 8,878 real SME business customers of an HR technology company, using CRM data across nine months. The best model reached 0.9029 AUC. On the strength of that number it would have been approved anywhere.

The same churn model, tested two ways

• Held-out test set. 0.9029 AUC. Linear-kernel SVM on 8,878 SME business customers. The number that gets a project approved.
• Rolling biweekly evaluation. 0.03 to 0.08 precision. Recall held at 0.59 to 0.72, so it still caught churners. Almost everything it flagged was a false alarm.

The authors attribute the collapse to the need for constant retraining. A strong backtest tells you very little about weekly operational usefulness.

Source: Oliveira et al., Predicting B2B Customer Churn, ICEIS 2025, 27th International Conference on Enterprise Information Systems.

Then the authors evaluated it the way it would actually run, on a rolling biweekly basis. Precision fell to between 0.03 and 0.08. Recall held up, so the model was still catching churners, but almost everything it flagged was a false alarm. A customer success team working that list would have stopped trusting it within a month.

This is the gap between a backtest and an operational system, and it is where most of these projects quietly die. The authors put it down to the need for constant retraining, which is a polite way of saying the model was never really finished.

What actually decides whether this works

Three things, in order, and none of them is the algorithm.

• One resolved customer record. If the same account appears separately in billing, support and product, every feature you engineer is averaging across people who are not the same person.
• A churn definition someone owns. Twelve months of inactivity is a bookkeeping rule. A useful definition is one where the business agrees the customer is gone while there is still time to act.
• A play attached to the score. The score identifies the list. The campaign recovers the revenue. Those are different pieces of work, and the second one is usually missing.

The published research agrees on the shape of this. A systematic review covering 61 studies from 2020 to 2024 names class imbalance, concept drift and evaluation misalignment as the persistent blockers, and notes limited real-world deployment despite technical maturity. The technology is not the constraint.

What it looks like when it works

What the model found, and what the business did with it

Two engagements. In both, the model produced a list and the campaign produced the revenue.

Source: Thinklytics engagement records. Figures are the clients' own.

Across our own engagements the pattern is consistent. The model finds a list nobody had, and the money comes from what the business does with it. A specialty apparel retailer moved from a twelve-month inactivity rule to a 90-day predictive window, flagged 94,000 at-risk members in the first scoring run, and recovered 68 of every 100 members who received outreach, worth $3.1 million in retained annual loyalty revenue. A B2B software company's rebuilt churn model went from 78 to 84 correct predictions per 100 and flagged $2.6 million of at-risk ARR in its first quarter, which the customer success team used to save 340 accounts.

In both cases the interesting number is not the accuracy. It is what someone did on the Monday after the list arrived.

What we would do first

Before any modelling, run a two-week check. Take last year's churned accounts and ask whether the signal was visible in your data 90 days before they left. If it was not, the problem is instrumentation, not modelling. If it was, count how many accounts your team could realistically work in a week, because that number sets the precision you need and therefore the whole design.

Our AI readiness work starts with exactly that check, and the data foundation practice handles the identity resolution underneath it. For the broader pattern on why models stall on the data layer rather than the model layer, see what production AI automation requires from data.

Frequently asked questions

Does predictive lead scoring actually increase conversion?

There is no credible published figure for it, and we looked hard. Every uplift statistic in circulation traces back to content marketing rather than research. The one defensible finding is Gartner's May 2026 survey of 227 chief sales officers, which found that sales organisations providing sellers with AI-enabled next best actions were 2.6 times more likely to achieve commercial growth. That is a correlation in a cross-sectional survey, not a measured uplift, and it should be quoted that way.

Why do churn models fail after they go live?

Precision collapses. A 2025 study of 8,878 SME business customers at an HR technology company built a model reaching 0.9029 AUC in testing. On rolling biweekly evaluation the same model's precision fell to between 0.03 and 0.08 while recall held at 0.59 to 0.72. In plain terms, almost every account it flagged was not actually churning. The authors attribute it to the need for constant retraining. A strong AUC in a backtest tells you very little about weekly operational usefulness.

What data do we need before a churn model is worth building?

One resolved customer record, a definition of churn everyone agrees on, and enough behavioural history to see the signal. The systematic review of 61 churn studies published in 2025 names class imbalance, concept drift and evaluation misalignment as the persistent blockers, not model choice. In our own work the binding constraint is almost always identity: if the same customer appears three times across billing, support and product, no model built on top of that will behave.

What is the right early warning window?

Short enough that someone can still act, long enough to be reliable. A specialty apparel retailer we worked with counted a member churned only after twelve months of inactivity, which is a bookkeeping rule rather than a warning system. Moving to a 90-day predictive window let them reach 94,000 at-risk members in the first scoring run. The window should be set by how long the intervention takes, not by how far ahead the model can technically see.

Should we build lead scoring or churn prediction first?

Whichever one already has a play attached to it. A score with no defined action is a reporting exercise. Before building either, find out what the team does the morning a list arrives. If there is no campaign, no outreach sequence and no owner, build that first, because accuracy will not change an outcome nobody acts on.

How accurate can a churn model realistically be?

In published research, best-in-class gradient boosting models reach around 91% accuracy and 0.948 AUC in banking contexts. Treat those as ceiling figures from controlled studies on clean datasets. The more useful question is precision at the volume your team can actually action. A model that flags 500 accounts a week at 40% precision is more useful than one flagging 5,000 at 15%, because the second one trains your team to ignore it.

What does retention work actually recover?

It depends entirely on the intervention, not the model. In one retail loyalty engagement the model flagged 94,000 high-risk members, targeted win-back campaigns recovered 68 of every 100 members who received outreach, and the client attributed $3.1 million in retained annual loyalty revenue to it. The model found the list. The campaign recovered the revenue. Those are separate pieces of work and only one of them is data science.