Polle de Maagt 
Polle de Maagt
In 2016, a hotel emailed to ask how my stay was going and whether it could do anything for me.
Kind. Personal. Proactive.
Except I had not checked in yet. I was still at the office.
The hotel knew I had a booking. It did not seem to know whether I had arrived. I could only guess which trigger it had used: the earliest check-in time, an average arrival time or something else.
The intention was good. The message was wrong.
The missing five percent.
“95% right is 100% wrong” was how I described it at the time. It was a figure of speech, not an accuracy measurement.
The point was that the missing detail could change the meaning of everything else.
YouTube recommended Dora episodes after my daughter got hold of my iPad. It recognised the viewing history, not who had been watching.
Retargeting a dress intended as a surprise gift could reveal it when the recipient used the same computer. Matching the browsing history was not the same as helping.
Booking.com sent me tourist guides to Amsterdam when I was Dutch and staying in the city’s hotels almost every week. It knew the destination, not the reason for the visit.
An account is not always one person. A purchase is not always for the buyer. A destination is not always a holiday.
Let people correct the picture.
In 2016, I argued that personalisation would get worse before it got better. Companies would have to pass through “artificial stupidity” on their way to artificial intelligence.
That was too generous to the idea that customers must endure the learning process.
Testing matters. But there are other choices: ask a question, use a confirmed event, test with a smaller audience or give a useful answer that does not pretend to be personal.
There are also ways to let people correct what a system thinks it knows. In October 2025, Spotify announced that listeners could exclude individual tracks from their taste profile, extending an existing playlist feature. Its examples included sleep sounds and children’s songs. Excluding a track reduces the effect of past and future plays on recommendations.
That is not proof that personalisation is fixed. It is a concrete response to the same problem: what I played is not necessarily what I want you to infer about me.
Confidence is not context.
A food retailer might want to answer “What should I eat tonight?” with something personal.
But purchase history does not reveal everything. Who is eating? What is already in the fridge? Is a dietary preference still current? Was the last purchase for someone else?
The system can ask instead of guessing. It can explain what it used. It can offer an ordinary recommendation when it lacks the information that would make a personal one useful.
Knowing something about me is not the same as understanding this moment.
More data is not always the missing ingredient. Sometimes the missing ingredient is a question.
Choose one automated message your organisation sends. List what the system knows, what it assumes and what happens if that assumption is wrong. Replace the riskiest assumption with a confirmed event, a question or a less personal fallback.
Related: The interface trains you back, on when we adapt to the system instead of the other way around.
Sources: adapted from my “Personalisation: it will get worse before it gets better”, Medium, 4 July 2016. The hotel and recommendation examples are from that original, not claims about those services today. Updated example: Spotify, “Spotify Gives You Even More Control With the Ability to Exclude Tracks From Your Taste Profile”, 1 October 2025. Spirit level made with Midjourney.
