How does AI personalise customer service responses?

Personalisation in AI customer service uses the customer's history (past contacts, purchases, preferences, account state) to tailor the AI's response to their specific situation. Done well, it feels attentive (the AI remembers what was discussed last time, picks up where it left off). Done poorly, it feels intrusive (the AI uses information the customer would not have offered, in a way that proves the company has it). The line is whether the personalisation serves the customer or shows off the data.

A customer contacts a retailer about a return. The AI's reply addresses them by name, references their order, and offers options based on their account history. The customer feels recognised. The same AI in a different deployment opens with "I see you've had three returns in the past six months, is this one related?" The customer feels surveilled. The data was the same; the framing was different.

What people in the field are saying

Blake Morgan's "Experience is everything..." argues that personalisation is a customer-experience win only when it makes the customer's job easier, not when it makes the company look smart. The distinction is small in design and large in effect.

What kinds of personalisation work?

Recognising the customer (using their name, knowing their account). Picking up context from the prior conversation (you mentioned X last time). Tailoring the answer to their plan or product (the steps are different for Pro accounts). Skipping questions whose answers are already known (do not ask for the order number you already have).

What kinds backfire?

Surfacing information the customer would not have offered (their return history, their cancellation pattern). Predicting needs without being asked (here's a discount, we noticed your usage dropped). Anything that highlights data flow rather than serving a need. Each crosses the line from helpful to surveillance-like.

Where is the practical line?

Personalisation that the customer would say yes to if asked. "May I use your previous order history to suggest a faster path?" is a yes. "I see you've been here 12 times this year" is information the customer did not need volunteered and reads as kept-track-of.

What about timing?

Personalisation works at the moment of need (the AI uses prior context to skip a step). It fails in the opener (greeting with personal facts before the customer has said what they want). The AI should personalise as it answers, not before the customer asks.

What is the practical first step?

List what data the AI has access to and what it uses in responses. For each, ask: would the customer thank you for using this, or would they be surprised you used it? Adjust the policy until every used data point passes the question.