STARTING LINE

One of our credit letters opened with “Dear Valued Customer.”

The news in the second paragraph was that we were cutting her credit line.

Nobody at MBNA wrote that letter to be cruel. Somebody wrote it years earlier, ran it past Legal, dropped it into the system, and moved on. Then the company built a brand promise on top of it and never revisited it.

One of our “computer-generated” letters ended up in front of our CEO after one of his close friends called to complain. It’s an understatement to say that he was not happy.

I’ve created a five-day email course called the Five Rooms that gives founders and senior executives 15 questions that catch the sentences that cost them trust in high-profile rooms. As I think about this story, it fits that course because Customer letters are often a locked closet inside of at least one of those rooms.

It took longer to collect all the letters than I expected because there were 19 different owners. By the end of that day (or maybe even the next day), 1,400 letters from across the bank were sitting on my desk, and I was the one who had to rewrite them.

None were wrong the day they shipped. Each matched the company we were when we wrote it. But the way we talked about ourselves evolved in the annual report, with affinity groups, and in every piece of marketing we produced. The letters never got the memo.

Your service reps answer out of a macro library somebody built in 2021. Your sales team runs sequences a departed SDR wrote and nobody has opened since. If you’re using AI, a growing share gets drafted by a model that learned your tone from those same templates. All of it goes out with your name on it. None of them gets the four rounds of edits your investor letter gets.

In February 2024, Air Canada's website chatbot told a customer named Jake Moffatt he could book a full-fare ticket to his grandmother's funeral and claim the bereavement rate afterward. That was never the airline's policy. Air Canada argued the chatbot was responsible for itself. A British Columbia tribunal disagreed and made the airline pay.

That defense will get tried a lot over the next few years, and it will keep failing. If it goes out under your name, it's yours, whoever or whatever typed it.

Klarna went further than most. Its AI assistant was handling the work of 700 agents within a month of launch. By May 2025 the company was hiring people back. “We focused too much on efficiency and cost,” CEO Sebastian Siemiatkowski said. “The result was lower quality.”

Nothing was broken. It was fine. “Fine” is what your customers quietly downgrade you for.

Three weeks ago, Gartner reported that 87% of customers say a company using generative AI for service has to give them a way to reach a human being. Your customers already assume the writing is automated. What they're telling you is that they want a door out of it.

The writing your company sends most is the writing nobody in your company has read for a long time.

Try this: Email whoever owns customer communications and whoever owns sales operations. Ask each one for the five templates their team sends most often, and for just two things: the date each one was last edited, and the name of the person who approved it.

Don't ask to see the copy yet. Ask for the dates and the names.

If something your customers receive every day was last touched in 2021 by a person who left in 2023, you've found the closet, and you haven't read a sentence yet.

In The Five Rooms, Day 4 walks into the Customer's room, where the mistake is what you write on purpose: the deck where every sentence starts with “we.” This is the closet in that room. The writing you approved once and stopped seeing. Same room, different door.

Subscribe here: fiverooms.carrd.co​

If you'd rather have an outside ear on whatever comes back before you carry it into your CEO's office, my inbox is open.

THREE GREAT POSTS YOU MIGHT HAVE MISSED

1. Do you sign your work? Carla Johnson says the bigger danger of the tools we use is that they make acceptable work so easy that we stop pushing for anything better.

2. The lost art of watching people work: Ideo's Molly Zeins uses HBO's The Pitt to make a simple point: expertise transfers when work is observable.

3. The two questions behind perfect content. Joe Pulizzi is one of my favorite creators. This post led me to buy Theo of Golden, and I'm loving it. The post is a great summary of his takeaways and food for thought about the content you put out in the world as a leader.

PROMPTS ARE QUESTIONS TOO

Here's a prompt worth stealing. Paste it once, and it runs on everything the model writes for you afterward.

The Undermine List

For everything you write for me from now on, end your response with a section called The Undermine List.

In it, name every factual claim in what you just wrote: numbers, dates, names, quotes, descriptions of what a person or a company did, and anything stated as settled fact. For each one, give me two things.

First, where it came from. A source you can point me to, your training data, or an inference you made to fill a gap.

Second, what it costs if it's wrong. Rank them worst first.

Then flag separately any claim I could not defend if the person it describes read it and disagreed.

Do not soften the list to make the draft look better. If a sentence is doing work its evidence can't support, say so plainly.

The boundary line (don't cut this part):

Run that on your next board letter, and it will find things. It will not find the sentence that's accurate and still lands wrong in the room, because that isn't a fact problem. The model checking its own work has the same disability you do. It already knows what it meant.

I welcome your comments or suggestions for future issues. Drop me a note here. If you found this on LinkedIn or had it forwarded to you, you can subscribe by clicking the button below.