If You Could Not Explain the Intervention to the Customer, Do Not Send It

A simple test for personalization: say out loud why this person is seeing this, and listen to how it sounds.

Published 2026-10-29 · 5 min read

If You Could Not Explain the Intervention to the Customer, Do Not Send It

Two Ways To Get Personalization Wrong

The first way is familiar: the message is irrelevant. It recommends something the person just bought, or greets a ten-year customer as a newcomer. It is annoying and mostly harmless.

The second way is harder to notice from inside a company. The message is relevant, and the customer's reaction is discomfort. How did they know that? They wonder what else the company knows.

Teams rarely catch the second kind, because by every internal measure the message worked. It was targeted correctly, and it may even have converted. The cost shows up later and elsewhere, as lost trust.

The Sentence Test

There is a quick check that catches most of these cases. Before an intervention ships, write the sentence that would explain it to the person receiving it:

"You are seeing this because ..."

Then read it as the customer.

"You are seeing this because you looked at delivery times twice" sounds reasonable. Most people would shrug.

"You are seeing this because your pattern of visits suggests you are worried about money" does not sound reasonable. If that sentence would make the customer uncomfortable, the intervention has a problem, however accurate the inference behind it.

The test checks whether the reasoning could be said to the person's face, which is a separate matter from whether it is correct.

What Makes An Explanation Acceptable

Explanations that pass tend to share three properties:

They refer to what the person did, here. Actions in this product, in this visit or recent ones. They do not involve data from elsewhere or conclusions about who the person is.

They are proportionate. A small observation leads to a small response. Looking at shipping leads to shipping information. It does not unlock a different price.

They leave the person in control. The response is an option the person can ignore, with no penalty for doing so.

Explanations that fail usually break the first property. They move from "you did this" to "you are this kind of person", and that step is the one people object to.

Inference Is Not The Problem. Reach Is.

None of this argues against reading behaviour. A good shop assistant notices when someone is looking for the fitting room. Nobody finds that intrusive, because the observation is limited to what happened in the shop and is used to help.

The same assistant commenting on the customer's finances would be out of line, even if the guess were right.

Digital products need the same sense of reach. This is why Fluence reads behavioural patterns, such as timing, navigation, and returns, and never the content a person types. It is also why every inference carries a confidence level: a guess should be held as a guess.

Make The Explanation Part Of The Record

The test becomes stronger when it leaves a trace. For each intervention a system makes, the log should be able to answer:

which signals led to it,

what was inferred, and with what confidence,

which action was chosen, and which were rejected,

what happened next.

A record like this serves three audiences. The team can see what worked. A compliance or privacy officer can check that the reasoning stayed inside agreed limits. And if a customer asks why they saw something, there is a true answer to give.

Under LGPD and GDPR, people have rights concerning automated decisions made about them. A system that can explain itself is in a much better position than one that has to reconstruct its reasoning afterwards.

A Rule Worth Adopting

Before the next personalization feature ships, add one line to the review: write the sentence the customer would hear. If nobody on the team would be comfortable saying it out loud, the feature is not ready, however well it performs.

See how we handle security and privacy.