Customer Context Becomes The Safeguard Against AI Personalization That Feels Intrusive

Every personalized offer rests on a conclusion about why a customer did something. A booking record may show five trips to Paris, but not all companies collect data on the reason for travel and whether those trips were for business, holidays or visits to an ill relative. The next offer usually assumes holidays. Companies are setting rules that infer things about a customer and their context, but risk alienating customers if they get this wrong.

The problem is familiar to Eulin Goh, the former Head of Research & Insight at British Airways, where she led the team responsible for delivering customer research and market insight to power commercial and experiential decisions across the airline. Earlier in her career, she worked agency-side in market research at TNS, giving her a view of consumer measurement from both the supplier and client end of the relationship. What she returns to is the offer that lands correctly and still unsettles the customer. While getting it wrong is a problem, so can being hyper-accurate.

"When something is hyper-personalized, people start to think, 'does this company know too much about me?' And it starts to make them question the value exchange," Goh says. The company has worked something out about the customer, and the customer can tell. Where that becomes a problem depends on the relationship. A customer with 100 purchases behind them accepts a level of specificity that a first-time buyer reads as intrusive.

Good targeting gets hidden

Companies that read a customer correctly face a second decision, which is whether to let the customer see it. One answer is to bury the accurate offer among irrelevant ones so nothing looks deliberate.

Target ran into this in the 2000s. It built a pregnancy prediction score from purchase patterns, then surrounded the resulting baby coupons with offers for things like lawn mowers so the targeting would look random. "With Target, they’re personalizing, but in a non-overt way, because it could get a bit too weird to receive such specific comms," Goh says.

Target realized the coupons felt invasive, which is why it disguised them. Concealing an offer inside a paper mailer was simple enough. A company that also knows what a shopper browsed and where they were when they did it has far more on display, and disguising a single offer leaves the rest visible.

Customers compare every system to a phone call

Consumers judge personalization on whether it makes their lives easier and is relevant to them. Anything a company does with what it knows has to meet that need. If doesn’t, it creates frustration and usually a desire to speak with a human, particularly in customer service scenarios where the comparison happens in real time. Goh points to a common breakdown where a chatbot verifies a customer, collects the details of the problem, then passes the conversation to a live agent who starts from the beginning.

"It feels like I've now got a two-step process to try and sort this out, and it hadn't learned any of that information. How is this any better as an experience than if I had just gone straight to speaking to an agent in the first place?" she says. The chatbot gathered everything it needed. It just never passed any of it to the agent, so the customer explained the problem twice. Data only helps if it reaches the person handling the request.

The record is slower than the customer

Three questions come before any conclusion gets used. The first is whether the company is allowed to combine the data at all. Companies hold so much data on their customers, but whether they are allowed to join it all up lies in the permissions they have from customers. "Are we allowed to do those data joins? In some scenarios they're not, because it's been collected for specific purposes," Goh notes. Permissions are key, particularly as companies also seek to join their data with data from other eco systems.

The second and third questions are how old the information is and the wider context. Customers change jobs and move house, and the record catches up slowly and companies are reliant on customers notifying them.

A business flyer who leaves a job may stop flying and see their frequent flyer status drop, changing their benefits. The tier is accurate. It says nothing about how long the person has been flying with them. "That person may have been a loyal customer of that brand for such a long time and their current status doesn’t tell the whole story of their relationship or feelings towards the brand," Goh adds. This is where different data points are helpful, such as length of membership. Any system that scores customers on recent activity will misread someone whose circumstances changed faster than the file.

Test it before everyone sees it

The fix is a rule about when a human looks at the decision before it reaches the customer. Coming up with personalized offers is now the easy part. AI can produce four versions of a campaign in the time it used to take to write one, which raises the value of the person who reads them before they go out. "You put data into AI and it comes out with three or four different treatments. Do those feel instinctually correct?" Goh says.

That read catches the obvious misses. Everything else gets settled by sending the offer to a small group and watching what happens, the same way marketers have tested website changes for years. "Test it on a small scale, and then if it does seem to work, we can deploy it. We have confidence in it," Goh adds. A wrong call that reaches one segment can be pulled back quickly, before the rest of the base ever sees it.

Customers ask an AI first

Where buyers start has changed. Across France, Germany and the UK, more than a third of consumers now research products with AI before deciding what to buy, which makes reviews and editorial coverage the material those tools draw on when a brand comes up. "As a brand, are you in the content areas which are feeding those LLMs?" Goh says.

They pull from much the same sources, so the answers tend to look alike. What separates one brand from another has to be established before anyone types a query. Repeat customers go straight to the brand, while everyone else meets it inside a shortlist they did not build.

Three-year plans hold little value when the tools change within six months. Goh does not read that as a reason to tear up what already works, and existing programs keep running while a company works out where AI discovery fits alongside them. "Businesses may not be ready to absorb some of these other things going on in the marketplace, but they need to have a point of view on it," she says.