Write in one notebook
I run a design and automation practice for about two dozen client organizations. Most of the day-to-day operations are run by an AI system I built, working from written rules. Two weeks ago I found out it had spent a week telling one of those clients they owed me a decision they had already made.
Nothing was down. No data was lost. The system did exactly what it was told. The problem was older and duller than that, and I think most small organizations have the same one sitting somewhere in their business right now.
// TWO COPIESTwo copies of the same fact
For each client I keep a register: every open item, with a stable ID, what it is, who is holding it, and what happens next. When a client and I say “number 29,” we both mean the same thing, forever. That part works.
I also kept a to-do list. Mine, for the week. Every line on it pointed at a register entry, so in theory the two agreed. In practice the list had grown to 324 lines pointing at 141 register entries, and every one of those lines was written by hand, on a different day, by whoever was in a hurry at the time.
Here is what that looks like when it fails. One entry for a client build was edited a dozen times in eleven days as the work moved. Next to it sat a single plain-English sentence, the one that goes into that client’s Friday status email. That sentence never moved. Nobody was editing it because nobody at my desk reads it. The client does. So for eleven days the email told them, politely and confidently, that a decision was waiting on them. They had made it on day two.
The week before, three client questionnaires came back finished and sat in a folder unread, while my dashboard still listed them as open asks. Same failure in a different costume. The information had arrived. It just had no home, so the system behaved as if it had never arrived at all.
// THE RULEThe rule
Two hand-written copies of the same fact will eventually disagree. Not might. Will. And the disagreement shows up in the copy that faces outward, because that is the one the people inside never look at.
This is not an AI problem. It is the oldest problem in small-business operations. The CRM says one thing, the spreadsheet says another, and the email you actually sent the customer says a third. Each was true on the day it was written. Each was written by hand. The customer only ever sees one of them.
When I sat down to fix mine, there were two honest options. The first was to keep both records and build a nightly check that compares them and flags the differences. That is the tempting one, because it keeps everything and nobody has to change how they work. It is also how you end up with a third document: the list of disagreements.
I picked the second option. Write in one notebook. The other one is a printout.
// IN PRACTICEWhat that means in practice
The register is now the only place a client’s work gets written. When something moves, it moves there. Every night, my weekly to-do list is regenerated from the registers, one line per open step, and if I edit a line on the to-do list by hand, it gets overwritten before morning. On purpose. The to-do list is not allowed to know anything the register does not.
The Friday email reads from the same register. So does the planning pass that books my week. So does the board I look at every morning. One notebook, three printouts, and none of the printouts can drift because none of them are written in.
Two details turned out to matter more than I expected:
- The ID never changes, the shape of the work does. Item 29 stays item 29 through a dozen edits, a scope change, and a client call. What lives under it is a short checklist that gets ticked. Progress reads off the ticks, not off a fresh copy of the description.
- A copy you cannot write in cannot lie by omission. The old to-do list looked authoritative because it was detailed. The detail was the problem. It was the last place anyone updated, so it was the most convincing wrong document I owned.
Where the AI actually fits
This is the narrow, useful shape of AI in operations, and it is not “remembering things.” A system that depends on an assistant’s memory is two notebooks again, just with one of them invisible.
Its job is routing. Every fact that comes in, from an email, a call, a form, a questionnaire, has exactly one place it lives, and the AI’s first job is to put it there. Its second job is to make it come back out where it is needed: a task on my calendar with the steps written in, a line in the report I ask for on Monday, the plain sentence in a client’s Friday email. The rule it works under is one line long: an input is not handled until it has a home and, if a person needs to act, an output.
Chat is not a home. An inbox is not a home. A folder someone will get to is not a home. If a fact only lives in one of those, it is lost, and it will stay lost until a customer finds it for you.
// THE QUESTIONThe question to take from this
Pick any fact your business tells customers. A delivery date. What they still owe you. What you are waiting on from them. Now ask: which copy of that fact is the one you write in, and which copies are printouts?
If you can name the notebook, and everything else is generated from it, you are fine. If you cannot, you have two notebooks, and one of them is quietly telling a customer something you would not say to their face.
You do not need an AI to fix this. You need to pick the notebook. The AI is what makes the printouts cheap.
If you want help finding the notebook in your own operation, get in touch. It usually takes one conversation to name it, and the printouts follow.