Writing

• Marketing Workflows
• 2 October 2026

Predictive analytics for the customer journey, minus the vendor fog

Predictive analytics for the customer journey means using your own behavioural data to score each customer on what they are likely to do next (buy, lapse, open, upgrade) and then changing what they experience at that stage of the journey. The AI platforms that do this well are mostly the CRM, email and analytics tools you already pay for, and the deciding factor is how much clean data you have, not which vendor you pick.

The data bar varies a lot. Klaviyo needs 500 customers who have ordered and 180 days of order history before it will predict anything, Google Analytics 4 needs 1,000 returning users who did the thing and 1,000 who didn’t, and Amplitude recommends more than 100,000 monthly users. If you sit below those lines, no amount of AI will predict your customers for you.

What is predictive customer journey mapping, once you take the vendor labels off?

A traditional journey map is a picture of stages (awareness, consideration, first purchase, repeat, lapse) drawn in a workshop with a lot of sticky notes and one person who insists the funnel is actually a flywheel. It describes an average customer who doesn’t exist. The predictive version attaches a probability to each real customer at each stage: this person has a 70% chance of buying again in the next 28 days, that one is drifting towards lapse. At that point the map is a set of scores that trigger different messages, offers or timing, and the poster can come down.

The vendor fog comes from everyone now selling something with “journey” and “AI” in the name. Gartner published its first Magic Quadrant for customer journey analytics and orchestration in 2026, and the definition, as quoted by CSG (one of the vendors named a leader), covers tools that “track and analyze how customers and prospects interact with an organization across multiple channels over time” and “enable organizations to prioritize and orchestrate real-time improvements to the customer experience” (CSG, April 2026). That is a big tent, and a lot of what sits inside it is dashboards with a propensity score bolted on.

Underneath, only a handful of prediction types matter for marketers:

• Propensity to buy or convert within a set window (GA4’s purchase probability looks 7 days ahead).
• Churn or lapse risk, which needs you to define what churn means for your business.
• Predicted customer lifetime value, usually over the next 12 months.
• Timing predictions: expected date of next order, or the best hour to send an email.
• Next best offer or content, which ranks options by predicted conversion.

If a platform demo can’t tell you which of these it is actually running, with what data and over what window, you are still in the fog. Tools that act on predictions by themselves are a separate question, which I cover in generative vs agentic AI in marketing.

Which AI platforms provide predictive analysis for customer behaviour?

For most UK marketing teams the realistic shortlist is below. The prediction types and minimum data requirements come from each vendor’s own documentation, because that is where the real limits are written down (usually a few scrolls below the hero video).

Sources for the table: Google Analytics Help, Klaviyo Help Center, Microsoft Learn, Braze Docs, Salesforce Help, Adobe Experience League (plus its AI features overview) and Amplitude Docs.

The underlying methods differ (Braze uses gradient boosted decision trees, Amplitude a transformer-based sequence model reading each user’s most recent 128 events), but for a brand manager that matters far less than the dull stuff: purchase events that carry a value and currency, a churn definition that makes sense, and someone whose actual job is to act on the scores.

How much data do you need before predictions mean anything?

More than most of us have, frankly, and UK consent rules make it tighter. The thresholds in the table are only the point where the software agrees to switch on, and clearing them says nothing about whether the output is useful, which is why Microsoft grades its churn models: an A means the model got at least 50% of predictions right and beat the baseline rate for churners by at least 10%, while a C means it did worse than 50% overall or worse than the baseline on churners. I’d like every vendor to show a grade like that on the dashboard, next to the confident-looking percentage.

Consent matters because a UK site with a proper cookie banner loses behavioural data from everyone who clicks “reject”. Google’s workaround, behavioural modelling for consent mode, estimates non-consenting users from similar consenting ones and has its own entry requirements: at least 1,000 events a day with analytics storage denied for at least 7 days, and at least 1,000 daily users with it granted for at least 7 of the previous 28 days (Google Analytics Help). So the predictions a smaller UK brand sees in GA4 can rest on modelled users, which is a model feeding a model. Worth knowing before you present it to the board as insight.

The other data problem is fragmentation. Salesforce’s State of Marketing 2026 survey of 4,450 marketing decision makers across North America, Latin America, Asia-Pacific and Europe (fieldwork October to November 2025) found only 58% had complete access to service data, 56% to sales data and 51% to commerce data (Salesforce, February 2026). A churn model that can’t see complaints logged with customer service is predicting with one eye shut. Salesforce sells the fix, of course, but in most of our organisations the CRM, the ecommerce platform and the service inbox have never been formally introduced. If your journey data is that thin, fix how you group customers before you try to predict them; my guide to AI customer segmentation covers the groundwork.

Why the customers most likely to churn are often the wrong ones to target

Eva Ascarza of Harvard Business School combined two field experiments, using data from a wireless services provider and a membership organisation, with machine learning to test the standard practice of sending retention offers to the customers at highest risk of leaving. Her paper, “Retention Futility: Targeting High-Risk Customers Might Be Ineffective”, published in the Journal of Marketing Research in February 2018, concluded that “customers identified as having the highest risk of churning are not necessarily the best targets for proactive churn programs” (Harvard Business School).

The better metric, she argued, is lift: how likely the retention effort is to change what the customer does. Simulations on both datasets showed that targeting by lift reduced churn more effectively than targeting by risk, and the paper won the American Marketing Association’s 2018 Paul E. Green Award (AMA, April 2019).

My reading is that plenty of high-risk customers have already decided to leave, so the 15% voucher just subsidises their exit, while some medium-risk customers would stay if given a reason. The churn features from Klaviyo, Braze and Microsoft all output a risk score, and none of their churn documentation offers a lift score out of the box. So the discount goes to the people who were leaving anyway, and six months later we are in a quarterly review explaining that churn went up despite the voucher, a sentence nobody wants on a slide.

You can approximate lift without a data science team: hold out a random control group within each risk band, send the retention message to the rest, and compare. The band where the gap between treated and untreated is biggest is where your budget belongs, and that may well be the middle band rather than the top one.

How to explore AI solutions for predictive customer journey mapping: a step-by-step process

This order keeps the software decision until you know what you are asking it to do.

1. Pick one journey moment where a prediction would change an action. “Second purchase within 60 days of the first” is a good one for ecommerce; “renewal in the 30 days before contract end” for subscriptions.
2. Write down what you would do differently for a high, medium and low score. If the answer is “send the same email”, stop here and save the licence fee.
3. Check your data against the vendor thresholds. Count customers with the target behaviour and without it, check the history length, and check that purchase events carry value and currency.
4. Use the prediction you already own first. GA4, Klaviyo, Braze’s preview model and Salesforce’s Einstein Engagement Scoring are already switched on or cheap to test for many teams.
5. Define success before launch, with a holdout. A random 10-20% of each score band gets the default journey, so the comparison measures lift. My piece on measuring the ROI of AI in marketing covers the payback arithmetic.
6. Run it for at least one full buying cycle, then compare treated against held-out customers band by band.
7. Only then look at a dedicated journey orchestration platform, and bring the results to the demo so the vendor has to beat a real number.

A worked example (illustrative numbers)

Say a UK skincare brand on Shopify has 4,000 customers who have ordered and two years of history, so it clears Klaviyo’s bar comfortably. It uses Klaviyo’s expected date of next order to time a replenishment email three days before each customer’s predicted date, instead of the old blanket “30 days after purchase” email, and holds out 15% of customers in each churn-risk band.

After one cycle, suppose the low-risk band shows almost no difference between treated and held-out customers (they would have reordered anyway), the medium band shows a clear gap, and the high-risk band a small one. The sensible move is to keep the email for everyone because it is cheap, put the discount only into the medium band, and stop discounting loyal customers who were going to pay full price. The pattern is made up, but the test is the one Ascarza’s work implies.

Prompts for using a general AI assistant in the mapping stage

General assistants like Claude or ChatGPT won’t predict your customers, but they are useful for turning an export into hypotheses you can then test, as long as you use anonymised or aggregated data.

I've attached an anonymised export of 12 months of orders: customer_id, order_date, order_value, product_category, acquisition_channel. For customers who made exactly one order, compare them with customers who made a second order within 90 days. List the five biggest differences in first-order behaviour, with the numbers, and say which differences could be causal and which are probably just correlated.

Here is our current journey map (stages and triggers pasted below). For each stage, tell me which prediction would change the action we take (propensity, churn risk, predicted CLV, next order date), what data we would need to produce it, and what we would do differently for high, medium and low scores. Flag any stage where a prediction would not change anything.

The second prompt tends to be the more useful one, because it finds the stages where a prediction adds nothing, which is often most of them. To turn this into a repeatable process, see how to build an AI marketing workflow.

Top AI tools for predictive analytics in customer behaviour: which one fits your situation?

A ranked list would mean pretending a Shopify brand and a mobile bank want the same thing, so here is how I’d match tool to situation:

• Ecommerce brand with a few thousand repeat customers: Klaviyo’s predictions if you’re already on it, plus GA4’s purchase probability, and test both before buying anything new.
• App or SaaS product with heavy event data: Amplitude if you have the 100,000 monthly users it recommends; Braze if messaging is the main lever and churn is the problem.
• Organisation already on Microsoft Dynamics: Customer Insights churn prediction, mainly because the A/B/C grade tells you whether the model is any good.
• Large email programme on Salesforce: Einstein Engagement Scoring for engagement and unsubscribe risk, bearing in mind it only looks at commercial sends.
• Enterprise with multiple channels and a data platform team: Adobe Journey Optimizer, where Adobe’s own documentation says send-time optimisation “may increase email click rate and push open rate in the range of approximately 2% to 10%”, and recommends optimising most emails for clicks rather than opens.

What UK data protection rules apply when you score customers?

Predicting behaviour from personal data is profiling under UK GDPR, which the ICO defines as automated processing “to evaluate certain personal aspects relating to a natural person” (ICO). The ICO’s checklist says people must be told how they can object to profiling, “including profiling for marketing purposes”, so your privacy notice needs to say you score customers, and your opt-out has to actually work.

Most marketing uses (who gets which email, when) are unlikely to count as decisions with legal or similarly significant effects; the ICO’s example of those is the automatic refusal of an online credit application. The line gets closer when predictions decide prices, credit, or who is excluded from an offer. The rules are also moving: the Data (Use and Access) Act 2025 reframed the automated decision-making provisions “from a prohibition with exceptions to a right of challenge with safeguards”, and the ICO consulted on draft guidance from 31 March to 29 May 2026, stressing meaningful human involvement and data protection impact assessments (Covington, Inside Privacy, April 2026). The same summary notes that special category data such as health generally needs explicit consent in automated decisions. Having worked in pharma brand management at Cipla and Alembic, I’d keep health-condition data out of any propensity model unless legal has signed it off in writing.

If you do one thing this quarter, take the journey moment from step 1 and count how many customers sit on each side of it. Compare that count with the Klaviyo, GA4 or Microsoft thresholds and you’ll know whether the next budget line should be a tool or a data clean-up. For how that kind of groundwork looks inside a small organisation, there’s my case study on AI adoption in marketing at ACTAsia. Which moment in your journey would you actually change if you knew who was about to leave?

FAQs

What is the difference between customer journey analytics and predictive analytics?

Customer journey analytics describes what customers did across channels and stages, such as where they dropped off. Predictive analytics scores what each customer is likely to do next, such as buy within 7 days or lapse. Journey analytics tells you where the problem sits; predictive scores tell you which individuals to treat differently at that point. Most platforms now offer both, but they need different data and should be tested separately.

Can a small business use predictive analytics for customer behaviour?

Yes, if it clears the data thresholds. Klaviyo needs 500 customers who have ordered and 180 days of order history, and Microsoft Customer Insights needs 500 customer profiles. GA4 needs 1,000 returning users who did the target action and 1,000 who didn’t. Below those levels, I’d expect simple rules based on time since last purchase to do about as well, and they cost nothing extra.

How accurate are AI churn predictions?

It depends on your data, and most vendors don’t show you. Microsoft is an exception: it grades churn models A, B or C, where A means at least 50% of predictions correct and at least 10% better than the baseline at identifying churners. Even an accurate risk score may target the wrong people, because Eva Ascarza’s research found high-risk customers are not always the ones retention offers change.

Is GA4 predictive analytics free?

Yes, GA4’s predictive metrics (purchase probability, churn probability and predicted revenue) come with standard GA4 properties, but your property must qualify. It needs purchase events with value and currency, 1,000 returning users who triggered the condition and 1,000 who didn’t over a 7-day period in the last 28 days, and model quality sustained over time. Predictions update once a day.

Sources

1. Google, Analytics Help, “[GA4] Predictive metrics”, accessed October 2026: https://support.google.com/analytics/answer/9846734
2. Google, Analytics Help, “[GA4] Behavioral modeling for consent mode”, accessed October 2026: https://support.google.com/analytics/answer/11161109
3. Klaviyo Help Center, “Understanding Klaviyo’s predictive analytics”, accessed October 2026: https://help.klaviyo.com/hc/en-us/articles/360020919731
4. Microsoft Learn, “Predict transaction churn: Dynamics 365 Customer Insights”, accessed October 2026: https://learn.microsoft.com/en-us/dynamics365/customer-insights/data/predict-transactional-churn
5. Braze Docs, “Predictive Churn”, accessed October 2026: https://www.braze.com/docs/user_guide/brazeai/predictive_suite/predictive_churn
6. Salesforce Help, “Prerequisites for Einstein Engagement Scoring”, accessed October 2026: https://help.salesforce.com/s/articleView?id=sf.mc_anb_prerequisites_einstein_engagement_scoring.htm
7. Adobe Experience League, “Send-time optimization: Adobe Journey Optimizer”, accessed October 2026: https://experienceleague.adobe.com/en/docs/journey-optimizer/using/delivery-settings/send-time-optimization
8. Adobe Experience League, “Work with AI: Adobe Journey Optimizer”, accessed October 2026: https://experienceleague.adobe.com/en/docs/journey-optimizer/using/get-started/ai-features
9. Amplitude Docs, “Predictions: Use Amplitude’s AI to help maximize lift”, accessed October 2026: https://amplitude.com/docs/analytics/predictive/predictions
10. Harvard Business School, Eva Ascarza, “Retention Futility: Targeting High-Risk Customers Might Be Ineffective”, Journal of Marketing Research 55(1), February 2018: https://www.hbs.edu/faculty/Pages/item.aspx?num=54936
11. American Marketing Association, “Ascarza’s ‘Retention Futility: Targeting High-Risk Customers Might be Ineffective’ Wins 2018 Award”, 24 April 2019: https://www.ama.org/2019/04/24/escarzas-retention-futility-targeting-high-risk-customers-might-be-ineffective-wins-2018-award/
12. Salesforce, “75% of Marketers Have Adopted AI, Yet Still Use It to Send Generic Campaigns” (State of Marketing 2026), 19 February 2026: https://www.salesforce.com/news/stories/state-of-marketing-2026/
13. CSG, “CSG Named a Leader in the First Gartner Magic Quadrant for Customer Journey Analytics and Orchestration”, 1 April 2026: https://www.csgi.com/insights/csg-named-a-leader-in-the-first-gartner-magic-quadrant-for-customer-journey-analytics-and-orchestration
14. Information Commissioner’s Office, “Rights related to automated decision making including profiling”, updated 17 August 2026: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/individual-rights/rights-related-to-automated-decision-making-including-profiling/
15. Covington, Inside Privacy, Jadzia Pierce, “UK ICO Consults on Draft Automated Decision-Making Guidance and Sets Expectations for ADM in Recruitment”, 24 April 2026: https://www.insideprivacy.com/united-kingdom-2/uk-ico-consults-on-draft-automated-decision-making-guidance-and-sets-expectations-for-adm-in-recruitment/