AI-driven customer segmentation has left behind rigid demographic segments (age, gender, postcode) and moved towards segmentation based on behaviour, intent and predicted value. Rather than grouping customers by what they are, models group them by what they are likely to do: buy, churn, recommend or spend more. This article covers the three most profitable levers of advanced personalisation — propensity models, customer lifetime value and lookalike audiences — and explains how to deploy them in compliance with the GDPR and the European AI Regulation.
From Demographic to Predictive Segmentation
Classical segmentation answers the question "who is my customer?" Predictive segmentation answers "what is my customer going to do?" The shift is profound because it moves the focus from the static profile to the behavioural signal: pages visited, purchase frequency, recency, average transaction value and response to previous campaigns. Unsupervised techniques such as clustering (k-means, hierarchical clustering) discover natural groups in the data without prior labels, while the classic RFM model (Recency, Frequency, Monetary value) remains a solid, explainable starting point.
Competitive advantage lies not in having more data but in converting that data into differentiated actions. Good predictive segmentation allows each group to receive a distinct treatment: the high-value customer at risk receives proactive retention; the new customer receives onboarding activation; and the lapsed customer receives reactivation with incentives calibrated to their expected value.
Propensity Models: Anticipating the Next Action
A propensity model estimates the probability that a customer will take a specific action within a given time window: purchasing a product, responding to an offer or, in particular, churning (churn models). Technically these are binary classification problems, solved with logistic regression when interpretability is the priority or with gradient boosting tree models when predictive accuracy is paramount.
The quality of a propensity model is not measured by accuracy alone but by metrics suited to imbalanced classes: AUC-ROC, precision and recall, and above all lift curves, which indicate how many times better than random chance the model identifies propense customers in the top decile. A common mistake is optimising overall accuracy when only 3% of customers churn: a model that always predicts "nobody churns" would achieve 97% accuracy and be completely useless.
Customer Lifetime Value: Prioritising by Predicted Worth
The Customer Lifetime Value (CLV or LTV) estimates the net revenues a customer will generate throughout their relationship with the company. Modelling it accurately allows the golden marketing question to be answered: how much can I afford to spend acquiring and retaining this customer? Probabilistic approaches such as the BG/NBD model combined with the Gamma-Gamma model predict, respectively, how many future purchases a customer will make and what their average monetary value will be.
CLV transforms budget allocation. Instead of spreading investment evenly, effort is concentrated on customers with the highest predicted value and the greatest churn risk. This combination of churn propensity and CLV defines a prioritisation matrix that avoids the waste of spending on retaining low-value customers or on acquiring profiles that will never be profitable.
Lookalike Audiences: Scaling What Works
Lookalike audiences start from a seed — the best customers ranked by CLV or recent converters — and search the available universe of users for those who most closely resemble them in behaviour. Advertising platforms automate part of this process, but the quality of the result depends entirely on the quality of the seed: a small or noisy seed produces wide and imprecise audiences.
Best practice is to build seeds from verified high-value customers, not from simple purchasers, and to refresh them regularly. With the gradual disappearance of third-party cookies, first-party data becomes the central asset: clean integration between the CRM, the Customer Data Platform (CDP) and activation channels is what makes personalisation scale without losing precision.
Recommendation Systems and Next Best Action
Personalisation does not end with grouping customers; it culminates in deciding what to offer each person at each moment. Recommendation systems address that problem through two families of techniques. Collaborative filtering recommends to a customer what similar customers have purchased, while content-based filtering recommends products similar to those the customer has already shown interest in. Hybrid approaches combine both and mitigate the cold-start problem — that is, what to recommend to a new customer for whom no signals yet exist.
The next step is next best action logic: rather than optimising a single campaign, the system decides, for each customer, which is the best possible action among all those available — an offer, a piece of content or simply waiting. This is where reinforcement learning techniques and multi-armed bandit tests come into play, balancing the exploitation of what already works with the exploration of promising alternatives. Well governed, this logic avoids customer fatigue, because it stops bombarding them with the same message and prioritises relevance over frequency.
Data Governance and Input Quality
No model surpasses the quality of its input data. Before discussing algorithms, a serious personalisation strategy requires a data governance policy: unified and deduplicated customer identity, traceable consents by purpose, shared definitions of variables and a clear lineage showing the origin of each data point. The Spanish Data Protection Agency (AEPD) emphasises the principle of minimisation: process only the data necessary for the declared purpose, not accumulating everything "just in case." Far from limiting personalisation, this principle tends to improve it, because it forces focus on the signals with genuine predictive power and discards noise.
| Lever | Question it answers | Typical technique |
|---|---|---|
| Propensity | What will the customer do? | Classification (logistic, boosting) |
| CLV / LTV | How much is the customer worth? | BG/NBD + Gamma-Gamma |
| Lookalike | Who else should I acquire? | Similarity against a value seed |
Compliance: GDPR and the EU AI Regulation
Personalisation is not a lawless territory. The GDPR requires a legal basis for processing (informed consent or balanced legitimate interest), transparency about the logic applied and, under Article 22, safeguards against automated decisions that produce legal or similarly significant effects on individuals. Profiling must be explainable, and the customer retains the right to object and to request human intervention.
Added to this is the European AI Regulation (AI Act), which entered into force in 2024 and is being applied on a phased basis. Although most marketing models are considered limited risk, the regulation prohibits manipulative techniques or the exploitation of vulnerabilities and requires transparency. The Spanish Data Protection Agency (AEPD) offers specific guidance on data processing in advertising and profiling that is worth following before deploying any model in production.
Steps for Implementing AI-Driven Segmentation
- Unify the data: consolidate first-party data in a CRM or CDP with consistent customer identity and traceable consents.
- Define the business objective: retention, cross-selling or acquisition; the model is designed for a specific action, not in the abstract.
- Build and validate the model: with separate training and test sets and metrics appropriate to the class imbalance.
- Activate in the channel: connect model scores to email, advertising or web content to differentiate the message by segment.
- Measure with a control group: reserve a group that does not receive the treatment in order to measure real incremental impact, not apparent impact.
Common Errors
The first is data leakage during training: including variables that are only known after the event to be predicted, which inflates metrics and causes the model to collapse in production. The second is confusing correlation with causation and acting on customers who would have purchased anyway, without measuring incrementality. The third is ignoring bias: a model trained on historically biased data perpetuates and amplifies that bias, with regulatory and reputational risk.
Frequently Asked Questions
How much data is needed to get started?
Less than most people think. A simple RFM or propensity model works with a few thousand customers and several months of history. Data quality and cleanliness matter more than raw volume; a small, well-labelled dataset outperforms a large and noisy one.
Does AI-driven personalisation comply with the GDPR?
It can, provided it rests on a valid legal basis, the customer is informed of the logic applied, rights of objection and human intervention are respected and the processing is documented. The key is designing compliance in from the start, not bolting it on at the end.
How do I know whether the model is actually adding value?
Through a control group. Only by comparing the behaviour of those who receive personalised treatment against those who do not can the incremental effect be measured. Without a control group you conflate the value of the model with the value of the channel itself.
What about the disappearance of third-party cookies?
It reinforces the importance of first-party data. Lookalike and personalisation strategies shift towards first-party data managed in a CDP, with clean consents, which is more robust and durable than third-party tracking.
Conclusion
AI-driven segmentation and personalisation generate value when built on three concrete levers — propensity, customer lifetime value and lookalike audiences — and measured with control groups that demonstrate real incremental impact, not apparent impact. Well-governed first-party data is the central asset of this strategy, especially in a world without third-party cookies. And none of this is viable outside of compliance: the GDPR and the European AI Regulation define the boundaries of legitimate, transparent and customer-respectful personalisation. At Summum Artificial Intelligence we design explainable, measurable models, integrated with your CRM and compliant with regulations, so that personalisation moves beyond a buzzword and becomes demonstrable growth.