If you want to automate team messages with AI, start with the message you repeat every week in the same shape - not with a decision. You find it by tallying, for two days, the messages you forward as they came. Then you test it with three boxes: same shape every time, a record of who told what to whom, and a mistake that another message can fix.
Picture a clinic, a repair shop, a small store. Messages move from hand to hand and through the owner. If that's you, keep reading.
This sits ahead of two things we've already written about: what happens when an assistant is handed a job it can't finish, and the automation that quietly stops. Both are about what goes wrong. This one is about what to hand over first.
1. You're already the relay
"The 3 o'clock wants to move to Thursday, tell the front desk." "The supplier says Tuesday, let the warehouse know." "He paid on Monday, give him the unit."
None of those carries a decision of yours. Someone said something, and you pass it to the person who needs it, in the same words. It doesn't feel like work. It feels like paying attention.
The difference matters when you pick what to hand over first. A decision calls for judgment: who to see first, what to charge, whether to discount. A message has to arrive whole, at the right person, on time. The second can be handed over. The first stays with you for now.
A message is a delivery
That's why the tally starts with messages, not tasks.
2. The two-day tally
You don't have to guess your most repeated message. You count it.
For two working days, each time you forward or repeat a message as it came, put a mark in a note next to a short phrase for it. Skip the answers you thought through. Count only what you pass along as it was told to you.
At the end, add up. The messages with the most marks go to the test.
A half-kept tally still points somewhere
Forget to mark half of them and you still land on the one that wore you out most. That's a starting point.
3. An imaginary clinic, blank fields
To see how it looks, here's an imaginary case: a small clinic with three people, a front desk and the owner. It's no real business, we used no client data, and the numbers are blank on purpose. Wherever you see [N], put your own.
The owner's note after two days, with the three boxes of the test:
| Message I repeated unchanged | Marks in two days | Same shape every time? | Record of who told what to whom? | Fixed by another message? |
|---|---|---|---|---|
| "The [time] appointment moved, tell the front desk" | [N] | [yes or no] | [yes or no] | [yes or no] |
| "The supplier arrives [day], tell the warehouse" | [N] | [yes or no] | [yes or no] | [yes or no] |
| "[Client] paid, hand over [item]" | [N] | [yes or no] | [yes or no] | [yes or no] |
She looks at the marks column and takes the row with the most. She reads its three boxes. If one says no, she drops to the next row.
Each box has a reason:
- Same shape every time. The details change (the time, the name), not the structure. If you rebuild it for whoever asks, it's a conversation.
- A record of who told what to whom. If someone asks tomorrow "who told them that?", "I think the AI did" can't be the answer. You need to open the exact message.
- Fixed by another message. A notice that the appointment moved is corrected with another notice. A charge sent to the wrong person may not be.
The agent's job, in one sentence. "When [someone] writes me that [thing], build this message from the details they gave: [template]. Send it to [person] and keep a copy." That's where the AI comes in: reading free text and filling the template. It doesn't choose who to write to or what to say.
Three yeses make a good candidate. A no, and it stays with you - which is fine. If no row earns three yeses, that business has no first agent yet. There's a message to tidy first, or a decision posing as a message.
The test is our own rule of thumb: start where a mistake is cheap and shows up fast. Nothing was measured behind it. Anthropic's guide to building agents recommends "finding the simplest solution possible, and only increasing complexity when needed" (Anthropic, Building effective agents). It says nothing about messages or cheap mistakes. That part is our position. And if the text never changes, a plain automatic reminder may be enough - no AI. If that's you, use that.
The conversation about the message
Todd: I spent thirty years as a project manager inside large companies, and I now design and build these systems. Before building anything, I ask which message gets repeated word for word. If there's none, there's nothing to hand over yet.
Naty: I had a spa in Venezuela and ran it as the owner. I didn't handle the consultations myself, so what reached me was whatever someone else said the client had said.
Todd: That's the case for delivering the message exactly. The agent doesn't improve it. It keeps it from changing on the way.
Naty: I'd say it's a stance, not a result. I'd start with the message that has a clear answer to "can another message fix it?"
Todd: Same here. If the fix is another message, I'm comfortable. If it isn't, it stays with the person.
4. Three things we haven't answered
We haven't tested this, so what follows are open questions, not findings.
Does the person receiving it know a machine sent it? If your front desk reads "the appointment changed" with no idea where it came from, they may doubt it. Say so in the message: "this is the assistant for [your business]".
What happens when the message changes shape? If the supplier sends the date in another format one day, the agent may send something that reads differently than usual. Read it yourself for the first few days, before you leave it alone.
What if the person receiving it writes back? A reply ("what time?") is no longer a message. It's a conversation, and someone has to answer it. Decide ahead of time who.
When an automation stops without telling you, that's a different problem, covered in the automation that stopped working won't email you about it. And if the message you carry contains a word like "verified," what to ask about it is in the article on "verified". Today's question is only which one goes first. For how much access an assistant needs for any of this, see Give a program a job it cannot finish and it will find another way out.
What you do this week
One note and two working days.
- Open a note and start the tally today. Each time you forward or repeat a message unchanged, a mark next to a short phrase for it.
- At the end of day two, add up and take the message with the most marks. Answer its three boxes: same shape every time, a record of who told what to whom, fixed by another message.
A small business gains more from messages that arrive without always passing through you than from a big promise of automation.
The first agent we propose doesn't decide. It carries the same words to the right person and keeps a copy you can read. You find it with a two-day tally and test it with three boxes.
At Hub365 AI, an AI-Native marketing agency, we help businesses find the first job for automation and build it. We work two ways, and you choose: we do it for you, or we give you the tools, the ideas and the support so you do it yourself.
If you run the tally, tell us which message you repeated most.
Keep reading
- The automation that stopped working won't email you about it - when automation stops without telling you
- Give a program a job it cannot finish and it will find another way out - on permissions
- Your website still promises things that stopped being true, and AI repeats them - on website promises that stopped being true
- It says "verified." Verified against what? - the article on "verified"
- Building effective agents, Anthropic - the original guide
