Season 1 • Topic Idea 2.
The Morning After
Here’s how it usually goes. The owner of a small coffee roaster has a slow month and a good idea: win back the customers who drifted away. She opens the AI assistant built into her email platform and types one line: “Send a win-back offer to our inactive customers.”
It sounds clear. Everyone knows what an inactive customer is.
The assistant looks at what the platform already has. There’s a ready-made list called “inactive”: anyone who hasn’t opened a marketing email in 90 days. That fits the word, so it uses it. It drafts a warm note, “We miss you! Here’s 20 percent off your next bag,” and queues it for that list. The owner reads the draft, likes the tone and approves it before lunch.
She meant something else. To her, an inactive customer is someone who hasn’t ordered in six months. She never said so, because it never occurred to her that the word could mean anything else.
The platform’s list didn’t hold the people who had stopped buying. It held her most loyal subscribers: people whose coffee ships every month on autopay, and who never open marketing email because they don’t need to. They were the last people in the business who needed winning back.
The next morning the inbox is full. Subscribers want to know why they’ve been paying full price. Some ask for the discount on their subscription. A few ask whether someone cancelled their subscription, since the email said the shop missed them. The part-time assistant who answers email spends the week on it, one careful reply at a time.
The customers who really had stopped ordering got nothing.
When the owner looks back, there’s nothing to point at. The AI didn’t misread the request. It picked a meaning that fit, from the tools it had, and it didn’t ask which one she had in mind.
“I Know That’s What I Asked For”
I spent years as a developer, and the reply I heard most when I showed someone what I’d built was this: “I know that’s what I asked for, but it isn’t what I want.”
It took me a long time to stop hearing that as a complaint. Giving directions that can’t be misread is hard. Ask someone for directions and you’ll find out quickly: “turn left at the old gas station” works fine until it reaches the person who never knew there was a gas station.
I write requests like the coffee owner’s all the time. “Clean up the old files.” “Make it more professional.” Each one sounds clear to me, because I know which meaning I have in mind. The person or the AI on the other end gets the words, not the meaning. Whoever does the work has to pick one, and if nobody asks, nobody knows anyone chose. The word costs me nothing to write. The cost shows up later, on someone else’s desk, and by then it looks like an ordinary fix, not like my word.
Four Lines Before I Send a Request
Now, before I hand over anything that matters, I look for the word that could mean two things. It’s usually a plain one: inactive, urgent, recent, customers, done. I write down the other meaning, pick the one I mean, and put it in the request with one example: “customers who haven’t ordered in six months, like someone whose last order was in April.”
Then I ask the other side to say it back: “Before you begin, tell me in one sentence what you think I’m asking, and what you’re assuming.” With an AI, that takes seconds. I add one more thing: “and exactly which list, filter or definition you’ll use.” That part matters, because an answer that just repeats my word back (”your inactive customers”) tells me nothing. In the coffee shop, the answer would have named the platform’s “inactive” list, people who haven’t opened an email in 90 days, and the owner would have caught it before anything went out. A wrong reading caught there costs a sentence. Caught after sending, it costs a week.
The price is a minute or two per request. The harder part is giving up a habit: approving the first draft because it sounds right, which is the one thing every wrong draft manages to do. It won’t catch everything. Some words only look clear because I’m the one who wrote them, which is exactly why the last line asks the other side to show me their reading instead of trusting mine.
Some people will find it fussy, like asking a coworker to repeat instructions back. That’s what it is. It also changes the conversation in a useful way: instead of arguing afterward about what the request meant, the two sides settle it before anyone has spent an hour on the wrong version. It’s cheaper than the week after.
The Two Readings Check
Copy this. Run it on any request where a wrong guess would cost more than a sentence.
THE TWO READINGS CHECK
Request: ______________________________
1. Which word in this request could mean two different things?
(Look at words like “inactive”, “urgent”, “recent”, “customers”, “done”.)
2. What’s the other meaning?
3. Which one do I mean? Write it into the request, with one example.
4. Before work starts, ask: “Tell me in one sentence what you think I’m
asking, what you’re assuming, and exactly which list, filter or
definition you’ll use.” Check the answer against line 3.
The coffee owner’s request wasn’t careless. It was normal. The same thing happens between people at work every day: a request goes out, it comes back done, and it isn’t what anyone wanted. Nobody files it under “one word meant two things.” It goes in as a revision, the person who cleaned up moves on, and the word stays exactly as unclear as it was, ready for next month.
Line worth keeping: The request was one word away from clear. Someone else paid for that word.
Try the check on one request this week. Which word turned out to mean two things?
Want the technical leadership view?
→ The companion Expression Layer post on LinkedIn The Expression Layer (LinkedIn)
Want the architecture view?
→ The companion Expression Layer article on Substack The Expression Layer
→ Free Subscribers can download the companion Leadership Brief: LB-002: When a Request Can Be Read Two Ways: A Worksheet for Catching Ambiguity Before It Becomes Rework.
It turns the idea into a management tool. Inside:
• warning signs
• questions to ask
• economics
• Monday action
• worksheet
If you’re trying to find out how much of your rework starts as a request that could be read two ways, start here.