Sohail Kasmani. Brand and marketing strategist in London: pharma brand management at Cipla and Alembic, now marketing strategy and AI adoption at ACTAsia. MBA, Brunel University London.
AI customer segmentation: methods, workflow and a worked example
AI improves customer segmentation by finding groups in your customer data that hand-written rules miss and by scoring every customer on what they are likely to do next (buy again, lapse, spend more), with the groups refreshing themselves instead of someone rebuilding a spreadsheet every Monday. For most e-commerce brands that means RFM as a base, predictive scores from the email or commerce platform on top, and an LLM to help interpret the results.
Klaviyo’s predictive analytics switch on once at least 500 customers have placed an order and you have 180 days of order history, and Shopify’s predicted spend tier filter appears once a store has made more than 100 sales. The harder part comes afterwards: a 2025 Gartner survey found personalised marketing produced negative experiences for 53% of customers.
How can AI improve customer segmentation in digital marketing campaigns?
Traditional segmentation is a set of rules a person writes: bought in the last 90 days, lives in Greater London, spent over £100. Rules are transparent, which is why every CRM still runs on them, but they only find the groups you already suspected existed. AI adds methods that look at many variables at once and either discover groupings you didn’t define or predict each person’s next move.
| Method | What it does | Best used for | Watch out for |
|---|---|---|---|
| Rule-based | Groups customers by criteria you define | Operational lists (lapsed, VIP, new) | Only finds what you already assumed |
| RFM | Scores recency, frequency and monetary value of purchases | A fast, explainable base segmentation | Ignores why people buy and what they browse |
| Clustering (e.g. k-means) | Finds natural groupings across many variables | Discovering segments you hadn’t defined | You choose the number of clusters; results can be arbitrary |
| Propensity / predictive | Gives each customer a probability of an action | Churn prevention, next-purchase timing, budget allocation | Scores drift; needs a holdout to prove value |
| Lookalikes | Finds new people who resemble a seed list | Acquisition | Only as good as the seed; the platform’s model is a black box |
| LLM-assisted discovery | Writes the analysis code, interprets and names clusters | Speed and interpretation | Confident nonsense if you don’t check the numbers |
The main AI segmentation methods, explained plainly
RFM (recency, frequency, monetary value)
RFM scores every customer on three things: how recently they last bought, how often they buy, and how much they have spent. Each dimension is usually split into five bands, so a customer who bought last week, orders monthly and spends a lot scores 5-5-5, and someone who bought once two years ago scores 1-1-1. It isn’t strictly AI, but it is the base most AI segmentation sits on, and Shopify builds it in: its segment editor has an rfm_group filter with ready-made groups such as Champions, Loyal, At risk, Almost lost and Dormant.
K-means clustering
K-means takes customers described by several numbers (order count, average order value, discount share, categories bought) and sorts them into a fixed number of groups so that people within each group are as similar as possible. The scikit-learn documentation states plainly that the algorithm “requires the number of clusters to be specified”, that it “responds poorly to elongated clusters, or manifolds with irregular shapes”, and that it can settle on a local minimum depending on where it starts. I’d standardise the variables first (otherwise spend in pounds swamps everything else), try four to eight clusters, and keep the version where each cluster fits one plain sentence.
Propensity models
A propensity model predicts each customer’s probability of a specific action within a set window. Google Analytics 4 is the free example most marketers already have: its predictive metrics include purchase probability (the chance a recently active user buys in the next seven days), churn probability (the chance a recently active user doesn’t return in the next seven) and predicted revenue over the next 28 days. You only get them if at least 1,000 returning users triggered the condition and at least 1,000 didn’t, within a seven-day period over the last 28 days, which rules out a lot of smaller shops. Klaviyo’s equivalent works from order data and gives predicted customer lifetime value (CLV), churn risk and an expected date of next order, with the CLV model retrained at least weekly.
Lookalike audiences
Lookalikes point the other way, at people who aren’t your customers yet. You upload a seed list (your best customers, ideally, rather than everyone who ever bought a £4 sale item) and the ad platform finds similar people. Meta’s Marketing API documentation says you need a Custom Audience of at least 100 people, and the lookalike can cover between 1% and 20% of a country’s population. I’d worry about the seed long before the percentage, which is why the RFM and CLV work feeds straight into acquisition.
LLM-assisted segment discovery
A language model reading raw CSV rows and asked to “segment my customers” will happily invent patterns. LLMs earn their place around the maths: tools such as ChatGPT’s data analysis feature write and run Python code on an uploaded spreadsheet, so the clustering is done by code you can inspect, and the model then helps you name the segments, spot oddities and draft campaign ideas. Shopify’s Sidekick assistant does a narrower version inside the admin, generating a customer segment from a plain language description.
A worked example: segmenting a UK e-commerce customer file
This is an illustrative example, not a real company. Picture a UK homeware brand on Shopify with about 40,000 past customers, three years of order history, one email platform, a Meta ads account and a spreadsheet that someone updates when they remember.
The data
Export one row per customer (not per order) with: customer ID, first and last order date, number of orders, total spend, average order value, share of orders that used a discount code, number of product categories bought, returns rate, email engagement in the last 90 days, marketing consent status, and first-party postcode area. Remove names and email addresses before any of it goes near an AI tool; the ID is enough to join results back later.
The steps
- Clean the file: merge duplicate customers, remove test orders and staff accounts, and decide how to treat refunds.
- Build RFM scores and look at the distribution, in particular how much of total revenue comes from the top band, before designing anything around it.
- Run k-means on the behavioural variables (order frequency, AOV, discount share, category breadth, returns rate), trying four to eight clusters.
- Add predictive scores such as Klaviyo’s churn risk and expected next order date.
- Cross the clusters with the RFM bands and name each resulting segment in plain English.
- Agree one action per segment, and a holdout group for each, before anything is launched.
What comes out
| Segment (illustrative) | Profile | What you do with it |
|---|---|---|
| Home builders | Frequent, full-price, buy across three or more categories | Early access to new ranges, no discounts, seed list for lookalikes |
| Sale hunters | Most orders use a code, low AOV, buy only during promotions | Fewer emails outside sale periods; test whether they buy without a discount at all |
| One-room refreshers | One or two large orders in a single category | Cross-category recommendations timed to expected next order date |
| Serial returners | High returns rate, especially in lighting and textiles | Better size and colour information, exclude from lookalike seeds |
| Drifting loyalists | Previously frequent, churn risk now high | Win-back sequence, a reason to come back rather than just 15% off |
In a file like this, watch for two things. One cluster may turn out to be a data problem rather than a customer type (the classic is a “segment” that is really everyone whose postcode field says “London London”), and one segment may turn out to be worth far more than the team thought. The second finding is the one that should change the budget.
What are the most effective AI workflows in marketing for customer segmentation?
Masha Shunko (University of Washington business school) and Serguei Netessine (Wharton) argue in HBR’s “Stop Automating Old Processes. Design New Ones Instead.” (September 2026) that managers redesigning a workflow for AI should decide what should change, where human authority should remain, and how the process should respond when the model is uncertain or wrong. For segmentation, that means the model proposes groups, a marketer decides which get a budget, and there is a rule for when a score looks wrong. The general version is in how to build an AI marketing workflow.
Route 1: the LLM-assisted version
Use a business account, not a personal one. OpenAI states that it does not train its models on data from ChatGPT Business, Enterprise or the API by default; your data protection officer will want to know which account you used. Upload the pseudonymised file, then work through prompts like these.
Prompt 1 (profiling the data) The attached CSV has one row per customer of a UK e-commerce brand. Before any analysis, describe each column and flag missing values, duplicates, impossible values (negative spend, future dates) and any column that could reveal health, ethnicity, religion or other special category information. Wait for my confirmation before continuing.
Prompt 2 (RFM and clustering) Calculate RFM scores (1-5 quintiles) using today's date as the reference. Then run k-means on order_count, avg_order_value, discount_share, category_count and returns_rate after standardising them. Try k = 4 to 8, show the silhouette score and cluster sizes for each, and recommend one value of k with your reasoning. Show me the Python code you ran.
Prompt 3 (interpretation and activation) For the chosen solution, give each cluster's means against the overall average, a plain-English name and a one-sentence description. Flag any cluster that looks like a data artefact. Then suggest one campaign objective per segment and the metric that would show it worked against a 10% holdout. No discounts for segments that already buy at full price.
Then check by hand: look up ten customer IDs from each segment in Shopify and see whether they match the description. If the “Home builders” segment is full of people with one order, something went wrong in step two.
Route 2: the no-code and low-code version
- In Shopify, filter customers by rfm_group and predicted_spend_tier, or describe the segment to Sidekick and edit the ShopifyQL it produces.
- In Klaviyo, build segments on predicted CLV, churn risk and expected date of next order once you meet the 500-customer threshold.
- In GA4, if you meet the 1,000-user thresholds, create predictive audiences (likely seven-day purchasers, likely seven-day churners) and share them with Google Ads.
- Upload your top CLV segment to Meta as a Custom Audience and build a 1% lookalike from it.
- Review monthly: segment sizes, and whether each still has a live campaign.
Top providers of AI-powered customer segmentation tools for e-commerce: what to look for
Most “best tools” lists are affiliate pages, so here is the checklist I’d use, then a few examples checked against their own documentation. Look for predictions built on your order data rather than just clicks, data thresholds you actually meet, automatic sync to the channels you use, scores exposed as fields a human can audit, clear data processing terms, and a way to hold out a control group.
| Tool | What it offers for segmentation | Limits |
|---|---|---|
| Shopify | Built-in RFM groups, predicted spend tier, Sidekick segment generation | Predicted spend tier needs 100+ sales; segmentation stays inside Shopify’s data |
| Klaviyo | Predicted CLV, churn risk, expected next order date, predicted gender | Needs 500+ customers with orders and 180 days of history |
| Google Analytics 4 | Purchase probability, churn probability, predicted revenue audiences | 1,000 positive and 1,000 negative returning users required |
| Meta | Lookalike audiences from a seed of 100+ people | You can’t inspect the model |
| Bloomreach Engagement | Purchase and churn prediction templates; Autosegments powered by its Loomi AI on the premium tier | Built for larger retailers; enterprise pricing and implementation effort |
Bloomreach’s predictions each produce a customer attribute holding a probability between 0 and 1 that you can filter on, and its Autosegments feature sits on the premium tier. On Klaviyo’s predicted gender, I’d leave it switched off unless you have a better reason to use it than “the field was there”, because guessing someone’s gender and sending them a different homepage on that basis is a fairly reliable way to get a complaint.
How do AI workflows support personalised marketing campaigns?
The hand-off works when each segment reaches the campaign team as a package: the audience (synced, not re-exported by hand), the objective, what makes this group different, the message angle, the exclusions (recent purchasers out of the win-back flow, say), and the holdout. Without that last one we end up crediting the segmentation for sales that would have happened anyway, which is how every AI project reports a heroic uplift while the profit line stays exactly where it was. Measuring the ROI of AI in marketing goes through holdouts properly.
McKinsey’s Next in Personalization 2021 report found 71% of consumers expect personalised interactions and that personalisation most often drives a 10 to 15% revenue lift, but that research is five years old and the newer evidence is less comfortable. Gartner’s survey of 1,464 B2B buyers and consumers in North America, the UK, Australia and New Zealand found personalised marketing generated negative experiences for 53% of customers, who were 3.2 times more likely to regret a purchase and 44% less likely to buy again. Audrey Brosnan, Senior Director Analyst in the Gartner Marketing Practice, says “passive personalization tactics alone no longer suffice”, and Gartner found what it calls course-changing personalisation outperforms traditional “next best action” recommendations, pointing CMOs towards interactive experiences such as quizzes.
So use segments to decide what to leave out as much as what to add. The “Sale hunters” in the example above don’t need a fourth email this week; the “One-room refreshers” might welcome a short quiz about the next room they’re planning, which also gives you data they chose to share.
Top services for personalising e-commerce marketing
Personalisation services fall into four rough groups: email and SMS platforms with built-in predictions (Klaviyo), customer data and engagement platforms (Bloomreach), ad platforms using your first-party audiences (Meta, Google), and the commerce platform’s own tools (Shopify). Ask each vendor whether its segments sync to the other three, because a segment that lives inside one tool only works in one channel.
Where does AI customer segmentation go wrong?
Data quality
Clustering algorithms will find structure in anything, including your errors. Duplicate customers, unlinked guest checkouts, test orders and refunds counted as sales will all produce confident-looking segments. We then build a campaign for them, launch it, and blame the creative, the send time, the subject line, probably Mercury being in retrograde.
UK GDPR and profiling
Automated segmentation of individual customers is profiling in the legal sense. The ICO’s guidance defines profiling as automated processing of personal data “to evaluate certain personal aspects relating to a natural person”, including their preferences, interests and behaviour (the guidance was updated on 31 March 2026 for the Data (Use and Access) Act 2025). The practical rules:
- People have “an absolute right to object to direct marketing”, including profiling for it, so an opt-out has to remove someone from the model’s outputs as well as the send list.
- The same ICO page warns that you may be using special category data if you profile people to target marketing, “for example, drawing inferences about people’s race, political opinions or health from other information”. A segment built on purchases of mobility aids or pregnancy products is a health inference, whatever you call it in the CRM.
- The ICO’s list of processing that needs a data protection impact assessment includes “any profiling of individuals on a large scale”, “combining, comparing or matching personal data obtained from multiple sources” and tracking behaviour online. Joining order data to GA4 and ad platform data ticks at least two.
- If you rely on legitimate interests, the ICO says it can apply to direct marketing only where PECR doesn’t require consent, and the use must be proportionate, low in privacy impact and unsurprising to people.
I worked in pharma brand management, where health data rules are the first question rather than the last, so I’d run the special category check before the clustering, not after the segments have been named and presented to the board.
Segments nobody acts on
This failure has nothing to do with the algorithm. We commission a segmentation, produce a lovely deck with personas called things like “Conscious Curators”, get a round of applause, and then keep sending everyone the same newsletter. A segment is only worth building if it changes a decision, even if that decision is to leave someone alone. I’d cap the first version at five segments, give each one an owner and a live campaign within 30 days, and retire any segment that goes three months without one. For wider context, what is AI in marketing covers the basics, and the ACTAsia case study covers AI adoption in a small charity marketing team.
The useful question to take into your next planning meeting is which one decision you would make differently if you knew each customer’s churn risk tomorrow. If nobody can name it, the segmentation can wait; if someone can, the RFM export takes an afternoon.
FAQs
What is AI customer segmentation?
AI customer segmentation uses machine learning to group customers by behaviour and predicted actions rather than only by rules a marketer writes. Common methods are clustering algorithms such as k-means, which find natural groupings across many variables, and propensity models, which score each customer’s likelihood of buying, lapsing or spending more. Platforms such as Klaviyo, Shopify and Google Analytics 4 now include these features out of the box.
How much data do you need for AI segmentation?
It depends on the tool. Klaviyo’s predictive analytics need at least 500 customers who have placed an order, 180 days of order history and recent orders. Shopify’s predicted spend tier needs more than 100 sales. GA4’s predictive audiences need at least 1,000 returning users who triggered the predicted action and 1,000 who did not within a seven-day period over the last 28 days.
Is RFM segmentation still useful when you have AI?
Yes. RFM (recency, frequency, monetary value) is quick, explainable and works on small customer files where machine learning models lack enough data. It also makes a strong input for AI methods: crossing clusters or churn scores with RFM bands gives segments that are both predictive and easy to explain. Shopify includes ready-made RFM groups such as Champions, At risk and Dormant in its segment editor.
Can I use ChatGPT to segment my customers?
You can use ChatGPT’s data analysis feature to write and run clustering code on an uploaded spreadsheet, then help interpret and name the results. Remove names and email addresses first, use a business account whose data terms your organisation has approved, and ask to see the code it ran. Check the output against real customer records, because a language model can describe patterns that are not in the data.
Does customer segmentation need a DPIA under UK GDPR?
Often, yes. The ICO lists large-scale profiling, combining personal data from multiple sources, and tracking online behaviour among the types of processing that require a data protection impact assessment. A segmentation that joins order history with website behaviour and ad platform data will usually meet at least one of those criteria. Customers also keep an absolute right to object to profiling for direct marketing.
Sources
- Klaviyo Help Center, Understanding Klaviyo’s predictive analytics, accessed September 2026.
- Shopify Help Center, Shopify-based customer segment filters, accessed September 2026.
- Shopify Help Center, Creating customer segments, accessed September 2026.
- Google Analytics Help, [GA4] Predictive metrics, accessed September 2026.
- Meta for Developers, Lookalike Audiences (Marketing API), accessed September 2026.
- scikit-learn, Clustering: K-means, accessed September 2026.
- OpenAI Help Center, Data analysis with ChatGPT, accessed September 2026.
- OpenAI, Enterprise privacy at OpenAI, accessed September 2026.
- Bloomreach, Predictions, accessed September 2026.
- Bloomreach, Autosegments FAQ, accessed September 2026.
- Harvard Business Review, Masha Shunko and Serguei Netessine, Stop Automating Old Processes. Design New Ones Instead., 14 September 2026.
- McKinsey & Company, The value of getting personalization right, or wrong, is multiplying, 12 November 2021.
- Gartner, Gartner Survey Reveals Personalization Can Triple the Likelihood of Customer Regret at Key Journey Points, 3 June 2025.
- Information Commissioner’s Office, What is automated individual decision-making and profiling?, updated 31 March 2026.
- Information Commissioner’s Office, Plan direct marketing, updated 28 April 2026.
- Information Commissioner’s Office, When do we need to do a DPIA?, updated 13 October 2025.
- Information Commissioner’s Office, Legitimate interests, updated 23 March 2026.
3 Comments
Comments are closed.