Predictive Analytics: Business Growth Guide

Predictive analytics uses past data to estimate what will happen next: which customers will cancel, which leads will buy, which invoices will be paid late, how many units a store will sell. A prediction is only worth building if someone will do something different because of it, and if doing that thing earns more than it costs.

Projects that disappoint often have a workable model and fail for other reasons. The score predicts an outcome nobody can change, the action it triggers costs more than it saves, or it sits on a dashboard the sales and customer success teams stop opening. This guide covers how to check those things before and after you build. Forecasting totals such as revenue and cash is covered in our AI financial forecasting guide, and running models in production in our MLOps guide.

Descriptive, predictive, and prescriptive

Descriptive analytics reports what happened: 212 accounts cancelled last quarter, mostly on the starter plan. Predictive analytics estimates what will happen to specific cases: this account has a 14% chance of cancelling in the next 60 days. Prescriptive analytics recommends an action: call this account, leave that one alone, offer a third an annual plan.

The step from predictive to prescriptive is bigger than most vendor diagrams suggest. A churn model learns who tends to leave. It cannot learn who would stay if you intervened, because historical data rarely contains that comparison. Recommending actions needs evidence about cause and effect, usually from experiments, which is why the section on uplift below matters.

Start with a decision, then check the data

Before choosing a tool, write down four things:

1. The decision the score will change and who makes it, such as which 250 accounts the customer success team calls each month.
2. The outcome and time window, defined precisely. Cancelled within 60 days of the score date, downgraded, and did not renew at term end are three different models.
3. How many past outcomes you have. Models learn from the rare class, so 50,000 customers with 300 cancellations is a small dataset.
4. How many cases the team can act on. If the team can make 250 calls a month, the part of the model that matters is the quality of its top 250.

Vendor minimums give a feel for scale. Salesforce's Einstein Lead Scoring needs at least 1,000 leads created and 120 of them converted in the last 200 days, for each segment you score; if you score all leads together and fall short, it uses a global model built from anonymized data from many Salesforce customers (Salesforce Help). HubSpot's AI lead scores, available in Marketing Hub Enterprise, need a minimum of 50 contacts, 25 converted and 25 not (HubSpot). A tool that agrees to produce a score has told you nothing about whether the score is good, and a model trained on 25 conversions cannot be very precise.

With a few hundred customers and a handful of cancellations a month, simple rules and cohort reports on plan, tenure, seat use, and support volume will usually do as well as a model and are easier to explain. Our churn reduction guide covers those. Agreeing on field definitions first also saves rework later; see our cloud data governance guide.

Why accuracy is the wrong number

Suppose 2% of accounts cancel each month. A model that predicts nobody will ever cancel is 98% accurate and useless. Accuracy rewards getting the common case right, and the common case is the one you don't need help with.

Three numbers, measured at the cutoff you will actually use, tell you more. Precision is the share of flagged accounts that cancelled. Recall is the share of all cancellations that were flagged. Lift is precision divided by the base rate, so a lift of 3 means flagged accounts cancel three times as often as average.

An illustration with hypothetical numbers: a B2B software company has 5,000 accounts paying $250 a month at an 80% gross margin, and about 2% cancel in a typical month, or 100 accounts. The team scores every account at the start of a month and later checks who left.

That is a useful model. It is also one where 88% of the top-5% list and 94% of the top-20% list would have stayed anyway, which matters as soon as acting on the list costs money.

If you plan to multiply probabilities by dollar values, check calibration too. HubSpot defines its "Likelihood to close" property as the percentage chance a contact becomes a customer in the next 90 days, so a score of 22 should mean 22% (HubSpot). Test that by grouping last quarter's contacts by score and comparing each group's predicted rate with what happened. The scikit-learn documentation calls the resulting chart a calibration curve or reliability diagram and describes ways to correct probabilities that run high or low (scikit-learn). A model can rank accounts well and still be badly calibrated. Ranking is enough for deciding who to call first; calibration is needed for anything involving expected dollars.

Does acting on the score pay?

A retention action pays when the value of the customers it keeps exceeds its cost across everyone who receives it, including those who were never going to leave. The break-even works out as:

break-even save rate = cost per account contacted ÷ (precision × value of one saved account)

The save rate is the share of would-be cancellers the action actually keeps.

Continuing the illustration, assume a saved account stays 18 more months, worth 18 × $200 = $3,600 of gross margin. Two actions are on the table: a call from a customer success manager costing about $50 of staff time per account, and an automatic 20% discount for three months, which costs $150 of revenue per account whether the customer needed it or not.

The one-in-six figure comes from a real campaign. In a mobile operator's retention program described by Nicholas Radcliffe, churn in the targeted group was 25% against 30% in an untreated control group, so the campaign kept about one in six of the customers who would otherwise have left. He described it as "a very positive result" (Radcliffe, 2007).

The same model supports one profitable program and three money-losing ones. Widening the list from 5% to 20% halves precision and doubles the break-even, and a discount sent to everyone flagged mostly pays customers who would have stayed. The table also assumes the same save rate for every customer, which is rarely true, and that assumption is what uplift modeling replaces.

Predicting churn is different from predicting who can be saved

Radcliffe's paper also describes a second operator whose retention campaign raised churn, to 10% in the treated group against 9% in the control group. His explanation is that contacting dissatisfied customers who had stayed mostly out of inertia prompted some of them to leave. His analysis found that targeting the right 30% of the base would have cut churn by about one percentage point instead.

He sorted customers into four groups: persuadables, who stay only if contacted; sure things, who stay regardless; lost causes, who leave regardless; and sleeping dogs, who become more likely to leave when contacted. A churn model puts a mix of all four at the top of its list, because high risk and dissatisfaction go together. An uplift model estimates the difference an action makes for each customer, which requires data from a randomized test of that action. A study in Information Sciences compared the two approaches on a retention experiment in the financial industry and found the uplift models outperformed churn prediction models and made the campaigns more profitable (Devriendt, Berrevoets and Verbeke, 2021).

Start collecting that evidence with the first campaign. Whenever you act on a score, hold back a random slice of the flagged accounts and do nothing for them. Without that control group you cannot measure the save rate the whole business case depends on.

Be realistic about sample sizes. Radcliffe's second operator treated a million customers and held out 100,000. In the illustration above, telling a 12% cancellation rate apart from 10% at the usual standard (80% power, 5% significance) takes about 3,800 accounts in each group. Holding back 50 of the 250 flagged accounts each month, that would take more than six years. A smaller company will get directional reads at best, so hold back a larger share early, pool results across months, and test actions large enough to show up.

Leakage: when a model looks too good

If a churn model scores near perfectly in testing, suspect leakage. Leakage is information in the training data that would not be available at the moment the prediction is made (scikit-learn). In customer models it usually comes from:

• Fields filled in during or after the outcome, such as a cancellation reason, a "renewal at risk" flag a manager sets after the customer says they are leaving, or a downgrade request.
• Snapshots taken at the wrong time, such as days since last login computed today for customers who cancelled months ago, which makes every churned account look inactive.
• Random train and test splits that put the same customer's earlier and later records on both sides.

The fix is to rebuild the training data as of each score date: for each account and month, use only what was known at the start of that month, then test on a later period the model never saw. Our AI sales and RevOps guide applies the same check to lead scores.

Put the score where the decision happens

A churn score on a separate dashboard tends to get checked for a week or two and then ignored. Scores get used when they appear inside the tool where the work happens: a field and a sorted queue in the CRM, a task created for the account owner, a routing rule that sends a high-intent lead to a senior rep.

A few habits help:

• Show the two or three main drivers with each score, such as "seat use down 40% in 60 days, open billing ticket." People act on reasons they can check.
• Size the flagged list to the team's capacity. If the team can make 250 calls, flag 250.
• Log what was done for every flagged account, including nothing. Without that record you cannot separate the model's effect from the team's.

Once the score and the action have proven themselves, workflow automation can handle the routing. To estimate what a retained account is worth for the break-even math, see our customer lifetime value guide.

When the world changes under the model

Models learn patterns from a period that eventually ends. From mid-March 2020, Instacart's model for predicting whether an item would be on a store's shelves degraded as panic buying emptied stores. A metric tracking how many items were found fell from 93% to 61%. Engineers retrained the model, rescored every hour instead of every three hours, and brought the metric back to about 85% (Fortune).

Churn models face a slower version of this problem because outcomes arrive late: you learn whether a March prediction was right in May. Watch the inputs, which move first. If you flag accounts by a score threshold rather than a fixed count, a sudden change in the size of the flagged list is an early warning. Then compare predicted and actual rates by score band as outcomes arrive and retrain when the gap persists. Our MLOps guide works through a population stability index calculation for input drift.

Predictions about people come with legal duties

Most marketing and churn models are ordinary data processing, subject to privacy law on the data they use (see our data privacy compliance guide). Models used for credit and hiring decisions carry extra duties:

• Credit in the US: when a creditor takes adverse action, Regulation B requires a statement of the specific principal reasons and says that failing to reach a qualifying score on the creditor's scoring system is not a sufficient reason (12 CFR 1002.9). The CFPB withdrew its 2022 and 2023 circulars applying this to complex algorithms on May 12, 2025 (CFPB). The regulation did not change, so a black-box credit model still has to produce reasons an applicant can act on.
• Hiring in New York City: Local Law 144 bars employers from using an automated employment decision tool unless it has had a bias audit, the results summary is published, and candidates get notice. The city began enforcing it on July 5, 2023 (NYC DCWP).
• The EU: GDPR Article 22 gives people a right not to be subject to decisions based solely on automated processing that have legal or similarly significant effects, with limited exceptions (GDPR Art. 22). The EU AI Act classes AI used for credit scoring of individuals (fraud detection excepted) and for recruitment as high-risk (Annex III); after the 2026 Digital Omnibus, those obligations apply from December 2, 2027.

If a score built for marketing ends up deciding who is offered payment terms or financing, check whether credit rules now apply to it.

Build, buy, or configure

Scores built into the CRM, such as Einstein and HubSpot's, are the quickest to switch on and already sit in the workflow. You get limited control over the outcome definition and features, and you still need to run the checks above on your own data.

Warehouse and AutoML tools such as BigQuery ML, Amazon SageMaker Canvas, and Vertex AI let an analyst train a model on data already in the warehouse, with the outcome and time window defined by you. Custom models in Python, using scikit-learn or a gradient-boosting library, give the most control, and you own deployment, monitoring, and documentation.

For a first churn or lead model, regularized logistic regression or gradient-boosted trees on a clean, well-defined table are reasonable starting points. Whichever route you pick, defining the outcome, building as-of snapshots, and checking for leakage will take most of the time.

A first project, in order

1. Pick one decision with an owner and a capacity, such as the 250 accounts the success team can call each month.
2. Define the outcome and window, and build an as-of training table with no fields filled in after the score date.
3. Train on older months and test on the most recent ones. Report precision, recall, and lift at the cutoff you will use, and compare them with a simple rule the team already trusts.
4. Work out the break-even save rate for each possible action before launching any of them.
5. Launch with a random holdout among flagged accounts and log every action taken.
6. Review monthly: realized save rate, calibration by score band, and input drift. Retire the model if it does not beat the rule.

For how predictions feed broader planning, see our guides on AI business strategy and AI for strategic decisions.

This guide is general information, not professional advice. The figures in the illustrations are hypothetical. Check legal obligations for credit, hiring, and personal data with qualified counsel in your jurisdiction.