Personalisation fails at scale for one reason: it is done by hand. A team that manually edits segments and writes personal messages collapses somewhere around 10,000 customers. AI changes the economics - models can rank products, draft subject lines, and pick offers for every individual in real time, and the same system that serves one customer serves a million at identical marginal cost.
Personalisation is a data problem first
Before any model, answer the three standard questions per customer: who (identity), what (behavior), when (recency). You need a clean event stream - page views, purchases, cart actions, email behavior - plus an identity graph that joins the same person across WhatsApp, email, and the app. Without this foundation, AI just writes better versions of the wrong message.
Recommendation engines: start with the two workhorses
Two families cover most use cases:
- Collaborative filtering - "people like you bought X". Works when you have enough purchase volume; the basis for most product recommendations.
- Content-based - recommends items similar to what the customer already viewed or bought. Useful for cold-start and new catalog items.
A hybrid that blends both typically outperforms either alone. Expect a 5-15% lift in average order value on recommended placements, and apply recommendations across the product page, email, and cart - not just the home page.
Personalize the message, not only the product
Subject lines, preview text, and dynamic content blocks are cheap to personalize and highly visible:
- Subject lines tuned to the customer's category affinity - "Your monthly skincare edit" beats "This month's offers".
- Dynamic hero blocks that swap based on segment, season, or weather in the customer's city.
- Countdown timers and stock indicators personalized per product view.
Test these changes with holdout groups - a 10% segment that receives the non-personalized version - so you can attribute the lift instead of assuming it.
Optimize the offer, but cap the discount
AI can also decide who gets which offer by predicting sensitivity: frequent buyers get loyalty perks instead of price cuts; cold contacts get the deeper discount that actually brings them back. The guardrail is a discount floor - set the minimum acceptable margin per customer segment before the model runs, and audit outcomes monthly. Letting a model chase conversion with unlimited discounts is how margins disappear quietly.
Guardrails: consent, bias, and control
Three rules keep AI personalisation safe in the region:
- Consent first - personalisation runs on first-party data the customer knowingly provided; respect opt-outs across channels.
- Bias review - periodically check that recommendations are not repeatedly hiding low-margin or minority-category inventory.
- Human override - every model has a kill switch, and a named owner reviews weekly outputs and changes triggers.
Measure lift against a baseline
Run continuous A/B holdouts. A realistic pattern after six months of AI personalisation: 10-20% lift in revenue per user, 5-10% lift in AOV, and a measurable drop in irrelevant-message complaints. If you cannot show these three numbers, the personalisation is decoration, not engineering.
Want to put automation to work in your business? Talk to Smart Logic.