Lead scoring and churn prediction in 2026: the scores are fine, the follow-through is where revenue leaks

92% of sales managers say qualified leads are regularly dropped every month, and 42% say their teams never make a second attempt. AI gives sellers back 4.8 hours a week and 72% of organisations report low reinvestment of it. The scoring is rarely the constraint.

92% of sales managers say qualified leads are regularly dropped every month, and 42% say their teams never make a second or third attempt after the first touch. Those are the two numbers to sit with before funding a better scoring model.

They come from a Zapier survey run by Centiment across 404 US sales and marketing professionals at manager level or above, fielded 3 to 17 April 2026 at B2B or mixed companies with 50 or more employees. 38% said leads are dropped multiple times a week. 12% said nearly every day.

A score is a ranking of leads by likelihood. It only produces revenue through what someone does with the top of the list. In 2026 the measured evidence says the list is not being worked.

Where the leads actually go

What happens to qualified leads after they are scored

Sales and marketing managers reporting on their own teams. Source: Zapier, survey conducted by Centiment, fielded 3 to 17 April 2026, n=404 US sales and marketing professionals at manager level or above, B2B or mixed companies with 50+ employees. Published 16 June 2026.

• Say qualified leads are regularly dropped each month
• Receive more than 50 qualified leads a week
• Spend three to 10 hours a week on CRM upkeep
• Say teams never make a second or third attempt
• Say broken follow-up pushes reps toward busywork
• Say leads get stuck between marketing tools and the CRM

Source: Zapier, survey conducted by Centiment, fielded 3 to 17 April 2026, n=404 US sales and marketing professionals at manager level or above, B2B or mixed companies with 50+ employees. Published 16 June 2026.

The same survey identifies the mechanics rather than just the symptom. 37% said misaligned tools and teams cause leads to get stuck between marketing tools and the CRM. 68% said team members spend three to 10 hours a week on CRM upkeep. 47% said broken follow-up pushes reps toward busywork over selling. 73% receive more than 50 qualified leads a week, and 11% receive over 500.

Read those together and the picture is specific. It is not that sellers are ignoring good leads out of laziness. The leads arrive in volume, they arrive in the wrong place, the plumbing between the marketing tool and the CRM loses some of them, and the time that could be spent on the second attempt is going on data entry.

None of that is a modelling problem, and none of it improves when the score gets more accurate.

The time that came back, and where it went

The hours AI gave back, and what happened to them

• Saved per seller, per week. 4.8 hrs. Average across sales organisations using AI tooling.
• Report low reinvestment of it. 72%. Organisations that do reinvest are 3.1x more likely to exceed lead-to-opportunity conversion goals.

In the same survey 25% report a return of 50% or more on their AI investment and 20% report a negative return of 50% or more. Roughly as many are losing badly as winning well, on the same class of tooling. Source: Gartner press release, 19 May 2026. Fielded January to February 2026, n=210 chief sales officers and senior sales leaders.

Source: Gartner press release, 19 May 2026. Fielded January to February 2026, n=210 chief sales officers and senior sales leaders.

Gartner surveyed 210 chief sales officers and senior sales leaders in January and February 2026. AI tools are saving sellers an average of 4.8 hours a week. 72% of sales organisations report low reinvestment of those savings into high-value sales activities.

The organisations that do reinvest are 3.1 times more likely to exceed lead-to-opportunity conversion goals and 2.2 times more likely to exceed customer growth goals. Be precise with that first one when you quote it: Gartner measured lead-to-opportunity conversion specifically, not conversion generally, and the 72% figure is "report low reinvestment" rather than "do not reinvest".

The same survey found 25% reporting a return of 50% or more on their AI investment and 20% reporting a negative return of 50% or more. Roughly the same number of organisations are losing badly as are winning well, on the same class of tooling.

That distribution is the argument for scoping this work around the follow-through rather than the model. The tooling is not what separates the two groups. Gartner's own finding is that what separates them is whether the freed time was redirected.

The third of churn that no model predicts

The share of churn that is a failed payment, not a decision

Monthly churn, July 2026 platform data. Involuntary means expired cards, declines and billing failures.

Source: Recurly Research, churn rate benchmarks, July 2026 data. Observed platform data, 14-month lookback, trials excluded. Number of sites and subscribers not disclosed.

Recurly's platform benchmarks for July 2026 put cross-industry monthly churn at 2.34% voluntary and 1.25% involuntary. For SaaS specifically: 3.22% total, 2.16% voluntary, 1.06% involuntary.

Involuntary churn is a failed payment. An expired card, a declined transaction, a billing address that moved. Roughly a third of measured subscription churn is not a customer deciding to leave, and a churn propensity model built on product usage will not see it coming because the signal is not in the product.

The price sensitivity is the useful detail. Involuntary churn runs at 1.30% below $25 average revenue per customer and 0.18% above $250. If you sell at a low price point, dunning, card-updater services and retry logic are likely to recover more revenue this quarter than any model will, and they are engineering work rather than data science work.

Recurly does not disclose the number of sites or subscribers behind the benchmarks, so it belongs in a document as observed platform data with sample size not stated. It is still measured behaviour rather than a self-report, which puts it ahead of most of what is published on churn.

What the budget says about the priority

Gartner's 2026 marketing survey, fielded January to March 2026 across 401 CMOs and marketing leaders mostly at organisations above $1B, found awareness and conversion accounting for 62.6% of total media spend, while customer loyalty and retention receives less than 15%, a 29% decline since 2024.

That is a measured allocation, and it is a far better fact than the claim that acquisition costs five to 25 times more than retention. We traced that one: it goes back to a 2014 Harvard Business Review article citing older Bain work, and no current primary source exists for it. If you are making the retention case internally, the Gartner allocation figure is both current and harder to argue with, because it describes what the organisation is actually doing.

The same Gartner base found 15.3% of marketing budgets allocated to AI, only 30% reporting mature AI readiness, and 70% saying internal marketing processes lack the maturity to implement and scale AI. The readiness gap is a process gap, by the respondents' own account.

What the data underneath is doing to the decisions

Validity's State of CRM Data Management in 2026, published 25 August 2026 across 500 B2B and B2C marketing professionals, found only 21% saying their CRM data is very well prepared to support AI, and 62% of organisations reporting lost revenue directly due to poor CRM data quality.

The finding worth reading twice: 78% of C-suite and 92% of SVP and VP respondents said they had acted on AI recommendations they later suspected were wrong because of the underlying data. The failure is not that the model was wrong. It is that a decision was taken on it anyway and the suspicion arrived afterwards.

Validity does not state a fieldwork date, which is worth attaching when citing it. And we could not find any 2026 source that quantifies duplicate, stale or incomplete record rates in CRM, despite that being one of the most confidently asserted statistics in the category. The decay figures in circulation all trace to older vendor content.

Where the modelling evidence actually lands

Scoring and retention claims, and whether they trace to a source

The category is sold on numbers, and a large share of them cannot be traced at all.

• Retention receives under 15% of total media spend, down 29% since 2024. Gartner, fielded January to March 2026, n=401 CMOs. A measured budget allocation, and the honest replacement for the usual cliché.
• Uplift targeting beats churn-propensity targeting in a controlled benchmark. 36.9% against 14.3% on the top 30%. Ruf and Handrich, IJMR, February 2026, n=1,980. Semi-synthetic: real attributes, simulated outcomes.
• Removing identifiers costs real model performance. AUC lands at 0.769 to 0.825 on behavioural, transactional, demographic and temporal features alone. Scientific Reports, May 2026.
• Acquisition costs five to 25 times more than retention. Traces to a 2014 Harvard Business Review article citing older Bain work. No current primary source exists for it.
• Predictive lead scoring lifts conversion by a specific percentage. No vendor-independent 2026 study measures lift against rules-based scoring. Every circulating figure traces to aggregators citing each other.
• Poor data quality costs $12.9M a year on average. Gartner's own page states this is research from 2020. It is quoted undated on almost every data-quality statistics page.

Source: Gartner, 8 June 2026; Ruf and Handrich, International Journal of Market Research 68(3), 19 February 2026; Scientific Reports 16:22349, 17 May 2026; Harvard Business Review, October 2014; Gartner data quality research, 2020.

Source: Gartner, 8 June 2026; Ruf and Handrich, International Journal of Market Research 68(3), 19 February 2026; Scientific Reports 16:22349, 17 May 2026; Harvard Business Review, October 2014; Gartner data quality research, 2020.

Three 2026 findings are worth knowing before anyone commits to an approach.

On uplift versus propensity, Ruf and Handrich published in the International Journal of Market Research in February 2026 on 1,980 matched customers and found targeting the top 30% by uplift cut churn 36.9% against 14.3% by churn propensity, with response-propensity targeting producing no effect at all. That dataset is semi-synthetic: real customer attributes with simulated outcomes. It is a controlled benchmark, not a field result, and presenting it as a measured 36.9% reduction in practice would misrepresent it.

The counterweight arrived three months later. A May 2026 paper in Applied Sciences found uplift models outperform traditional baselines in a causally controlled regime while traditional baselines remain economically superior in confounded proxy settings, and that the highest-yielding causal policy frequently suffers severe targeting instability. Uplift wins in clean conditions and can lose money in realistic ones.

On model performance, published 2026 results on the same problem span a range that should end any discussion of a benchmark AUC. A January 2026 Scientific Reports paper reported ROC-AUC 0.864 with precision 0.821 and recall 0.418, missing most actual churners. A February 2026 Frontiers in Artificial Intelligence paper reported AUC 0.932 on a different public telecom dataset. A May 2026 paper that removed identifiers and trained only on behavioural, transactional, demographic and temporal attributes reached AUC 0.769 to 0.825. Every one of those ran on older public datasets rather than 2026 data, which is itself worth knowing.

The privacy result is the practical one: strip the identifiers and you lose real performance. If your consent posture is heading that way, plan around the lower band.

What to count before building anything

What to count before building a better score

Two measurements, both possible in a day, both of which change what gets funded.

• Second and third contact attempts on last quarter's top-scored leads. If the top of the list is not worked twice, the score is not the constraint and a more accurate one changes nothing.
• Time to first response on the highest-scored leads. The routing between the marketing tool and the CRM is where 37% of managers say leads get stuck.
• Last quarter's cancellations split into voluntary and involuntary. If a third were payment failures, retry logic and card updating recover more revenue this quarter than any model, and in weeks.
• Whether the freed seller hours went back into selling. 72% of organisations report low reinvestment. The 3.1x conversion difference sits between those who redirected the time and those who did not.
• A benchmark conversion rate to score yourself against. The widely used B2B conversion tables come from one firm's 2019 to 2025 client data with no disclosed sample size. Use your own history.

Source: Zapier and Centiment, fielded 3 to 17 April 2026, n=404; Gartner, fielded January to February 2026, n=210; Recurly Research, July 2026 platform data.

Source: Zapier and Centiment, fielded 3 to 17 April 2026, n=404; Gartner, fielded January to February 2026, n=210; Recurly Research, July 2026 platform data.

Take last quarter's highest-scored leads and measure four things. How many received a first contact attempt. How many received a second. How many received a third. And how long the first response took.

If the top of the list is not being worked twice, the score is not the constraint and a better one changes nothing. That measurement takes a day and it settles the question.

On the churn side, split last quarter's cancellations into voluntary and involuntary before doing anything else. If a third of them were payment failures, the highest-return work this quarter is retry logic and card updating, not modelling. It is less interesting and it recovers revenue in weeks.

For the modelling side of this, including the precision collapse that shows up after go-live, see lead scoring and churn prediction: what works. The identity work that has to come first is covered in customer analytics in 2026. Our data foundation practice resolves the customer record these scores are built on, and the AI readiness assessment is where we run the follow-through count above before a model is scoped.

Frequently asked questions

How many scored leads never get worked?

Enough that scoring is rarely the binding constraint. Zapier, using Centiment, surveyed 404 US sales and marketing professionals at manager level or above between 3 and 17 April 2026 and found 92% reporting that qualified leads are regularly dropped each month, with 38% saying it happens multiple times a week and 12% nearly every day. In the same survey 42% said their teams fail to make a second or third attempt after the first touch, and 37% said misaligned tools cause leads to get stuck between marketing systems and the CRM.

Does giving sellers time back improve conversion?

Only if the time goes back into selling, and usually it does not. Gartner surveyed 210 chief sales officers and senior sales leaders in January and February 2026 and found AI tools saving sellers an average of 4.8 hours a week, while 72% of sales organisations report low reinvestment of those savings into high-value sales activities. The organisations that do reinvest are 3.1 times more likely to exceed lead-to-opportunity conversion goals and 2.2 times more likely to exceed customer growth goals. Note the specificity: it is lead-to-opportunity conversion, not conversion in general.

How much of subscription churn is a payment failure rather than a decision?

Roughly a third, and no model fixes it. Recurly's platform benchmarks for July 2026 put cross-industry monthly churn at 2.34% voluntary and 1.25% involuntary. For SaaS specifically it is 3.22% total, 2.16% voluntary and 1.06% involuntary. Involuntary churn, meaning failed payments and expired cards, falls sharply with price: 1.30% below $25 average revenue per customer against 0.18% above $250. Recurly does not disclose the number of sites or subscribers behind the benchmarks, so quote it as platform data with sample size not disclosed.

Is retention actually underfunded relative to acquisition?

Yes, and there is now a 2026-fielded figure for it instead of the usual cliché. Gartner's 2026 marketing survey, fielded January to March 2026 across 401 CMOs and marketing leaders, found awareness and conversion accounting for 62.6% of total media spend while customer loyalty and retention receives less than 15%, a 29% decline since 2024. That is a measured budget allocation. The widely quoted claim that acquisition costs five to 25 times more than retention traces to a 2014 Harvard Business Review article citing older Bain work and has no current primary source.

What does poor CRM data do to a scoring programme?

It moves the failure from the model to the decisions taken on the model's output. Validity's State of CRM Data Management in 2026, published 25 August 2026 across 500 marketing professionals, found only 21% saying their CRM data is very well prepared to support AI, and 62% of organisations reporting lost revenue directly due to poor CRM data quality. More pointed: 78% of C-suite and 92% of SVP and VP respondents said they had acted on AI recommendations they later suspected were wrong because of the underlying data. Validity does not state a fieldwork date.

Does uplift modelling beat scoring by churn propensity?

In a controlled benchmark, yes. In realistic conditions, the picture is mixed and worth knowing before you fund it. Ruf and Handrich, publishing in the International Journal of Market Research in February 2026 on 1,980 matched customers, found targeting the top 30% by uplift cut churn 36.9% against 14.3% when targeting by churn propensity. That dataset is semi-synthetic: real customer attributes with simulated outcomes, so it is a benchmark rather than a field result. A May 2026 paper in Applied Sciences found uplift models outperform traditional baselines under clean causal conditions while traditional baselines remain economically superior under realistic confounding, and that the highest-yielding causal policy frequently suffers severe targeting instability.

How good are churn models in 2026, and does privacy change it?

Published results span a wide range on the same problem, which tells you the headline numbers are dataset artefacts rather than capability statements. A January 2026 paper in Scientific Reports reported ROC-AUC 0.864 with precision 0.821 but recall only 0.418, meaning it missed most actual churners. A February 2026 paper in Frontiers in Artificial Intelligence reported AUC 0.932 on a different public telecom dataset. A May 2026 paper that deliberately removed identifiers and trained only on behavioural, transactional, demographic and temporal attributes landed at AUC 0.769 to 0.825. Stripping identifiers costs real performance, which is the number to plan around.

What should we fix before building a better score?

Count the follow-through first. Take last quarter's highest-scored leads and measure how many received a second contact attempt, how many received a third, and how long the first response took. Zapier's April 2026 survey found 68% of managers saying their teams spend three to 10 hours a week on CRM upkeep and 47% saying broken follow-up pushes reps toward busywork rather than selling. If the leads at the top of the current score are not being worked twice, a better score changes nothing, and the work is in the routing and the sequence.

The work behind this

Six engagements in the case library carry this capability, on both sides of the problem: finding the accounts worth attention, and catching the ones about to leave.

Every one names the client where we are permitted to and states the measured outcome: Lead scoring and churn prediction, 6 engagements.