I received an email from Eurostar that began: “We’re sorry, we’ve had to make some changes to the timings of your Eurostar train to Paris Gare du Nord.” A small flutter of annoyance. We’ve all had that email. I read on.

My departure from London St Pancras: unchanged, 12:01. My arrival in Paris: 15:24 instead of 15:30. I will arrive six minutes earlier. Six minutes gagnées, same departure time, and an apology in the bargain.

Whether the culprit was an AI or plain old automation, the effect is the same. It certainly reads as though a change was spotted and the “changes to your train times” template dispatched, apology thrown in, with no check on whether the change was actually bad news. A template cannot read what it sends. It knows something changed; it cannot tell an improvement from a disruption, so it apologises for both, just in case. (It also greeted me as “minter dial,” lower case. Make of that what you will.)

An Apology Is a Currency

Why does this matter? Because an apology is a currency. When a company apologises for giving me six minutes of my life back, that apology is worthless, and so, by extension, is the next one. The day Eurostar genuinely ruins my trip, the same words will arrive with the same “sincerity”, which is to say none. Automation done wrong doesn’t just waste our time reading nonsense; it devalues the very empathy companies claim they want to show. This is what I refer to as “organisational empathy”… understanding what one action does or implies for the whole organisation and later actions.

When It Stops Being Funny

And this is the gentle end of the spectrum. When automation goes wrong with real stakes, it stops being funny. There are many tales of automation gone wrong. I bet you’ve got your own examples? The problem is that so much of the automation goes unchecked inside the organisation (like a black box). And it’s going to get worse with AI. Here’s an example where the consequences were real.

In 2024, Air Canada’s chatbot invented a refund policy that didn’t exist, telling a grieving customer he could claim a bereavement fare after travelling. When the airline refused to honour its own bot’s promise, a Canadian tribunal held Air Canada liable. The chatbot was helpful, articulate and wrong, a combination now being shipped at scale.

Or think back to Facebook’s Year in Review in 2014, which cheerfully framed users’ photos with party balloons, including, for Eric Meyer, a photo of his daughter who had died that year. He called it “inadvertent algorithmic cruelty.” The machine wasn’t malicious. It was indifferent.

The pattern is the same each time. We automate the message but not the judgment. We build systems that detect events and trigger words, without the one thing that matters: knowing when to speak, when to stay silent, and what the words will mean to the human receiving them. Empathy is not a tone of voice you can template. It demands an understanding of the other person’s context, and a six-minute improvement is not an apology-shaped event. That doesn’t mean empathy can’t be created or encoded into AI (i.e. so that the AI endeavours to understand cognitively — and recognise — the client, emotion, context, and needs).

So here is my modest proposal for the engineers and marketers behind these systems: before you automate an apology, automate the question “should we be apologising at all?” If the answer is no, say nothing. Or better still: recognise the news for what it really is: good news!

What’s the most absurd automated message you’ve ever received? Invitation open! 🚄


Sources

Pin It on Pinterest