The Rise of Intelligent Personalisation

It is worth being explicit about the ladder, because most organisations are several rungs lower than they believe.

This is the argument I make most often with commercial teams, and it is the one that meets the most resistance, because relevance is what personalisation platforms are built to optimise.

A relevant offer may still reach a customer who would have purchased anyway, in which case the organisation has converted margin into discount. It may create unnecessary discount cost across an entire segment because propensity models are, by construction, best at identifying people who were already likely to act. It may arrive at the wrong time, when the customer is mid-journey or in financial difficulty. It may ignore circumstances the data does not capture. And it may raise short-term response while eroding the trust that supports long-term engagement.

Propensity models find people who are likely to say yes. Incrementality models find people whose answer you actually changed. Those two groups overlap far less than most marketing plans assume.

The commercial consequence is significant. McKinsey's research indicates that faster-growing companies generate around 40% more of their revenue from personalisation than their slower-growing peers, and estimates a revenue lift in the range of 5% to 15% and a marketing return improvement of roughly 10% to 30% for organisations that get personalisation right, alongside potential acquisition cost reductions of up to 50%. These are research estimates drawn from company performance data and case work, not guaranteed outcomes, and the upper end of each range reflects the strongest performers rather than the typical one.

Every organisation pursuing personalisation is navigating five boundaries at once, and the boundaries move depending on category, culture and the individual customer.

• Helpful versus intrusive. The same message can be either, depending on whether the customer understands how you knew.
• Convenient versus manipulative. Reducing friction is a service. Reducing deliberation is not.
• Relevant versus repetitive. Frequency destroys relevance faster than poor targeting does.
• Customer value versus commercial extraction. If the interaction only makes sense from one side of the relationship, it will not survive scrutiny.
• Automation versus human judgement. Some moments, particularly hardship, bereavement and complaint, should never be handled by a decision engine alone.

In my experience the organisations that handle this well have written down where their boundaries are, and have made those boundaries a governance artefact rather than an individual judgement made under quarterly pressure.

The measurement set determines the behaviour of the programme, so it should be chosen before the platform is bought rather than after.

• Incremental value against a permanent holdout, reported in currency and not in response rate.
• Customer lifetime value trajectory for exposed versus unexposed cohorts over a period long enough to see decay.
• Discount leakage, meaning the share of incentive spend allocated to behaviour that would have occurred anyway.
• Trust indicators: opt-out rates, complaint volume, contact fatigue and channel deliverability.
• Long-term engagement depth, which is the only metric that captures whether the relationship is improving rather than being harvested.

Differential treatment at scale can become discrimination, even when no protected attribute is used as an input. Proxies are pervasive in behavioural data, and the only reliable defence is testing outcomes across groups rather than auditing inputs.

Vulnerable customers require explicit protection. Systems optimised for response will find people who respond, and some of those people respond because they are in difficulty. Any commercial decisioning system in financial services, gambling, credit or health needs hard suppression rules that cannot be overridden by an optimisation objective.

Data misuse and scope creep are gradual rather than dramatic. Data collected for one purpose migrates to another, each step defensible, and the aggregate is something the customer never agreed to. Purpose limitation should be enforced technically, not just in policy.

Filter bubbles narrow the customer's experience of what you offer. Optimising every interaction on predicted preference produces a shrinking view of the proposition and, over time, a less valuable relationship. Deliberate exploration should be built into the system.

Explainability matters commercially, not only for regulators. If a frontline colleague cannot explain why a customer received a recommendation, they will not stand behind it, and adoption collapses.

There is also a legitimate counterargument that broad-reach marketing remains underrated. Evidence on brand building suggests that reaching light and non-buyers matters for category growth in ways that precision targeting does not capture. Personalisation should be understood as one component of a growth mix, not as a replacement for reach.