I built a system of action by accident, running a rafting company
The best minds in AI are writing about systems of action. I built one by accident, running a rafting company.
Here is what they mean, in plain terms.
Your software is a system of record. It knows what happened: the booking, the payment, the name on the reservation. It is very good at remembering.
A system of action knows what a change means, and does something about it. When a booking jumps from 12 guests to 18, a record just stores the new number. A system of action knows that number broke the boat-to-guide ratio, the van seats, the driver count, and scheduling conflicts, and it starts fixing them.
I couldn’t find one. So I built it, on top of the booking software we already had.
Not a chatbot. A database. 44 connected tables and 207,000 data points, with the operation’s real rules written into it: who is cleared to guide what, how many boats a trip takes, how many seats are in each van, which movements collide. It has run the season, 4,649+ guests so far, 44 seasonal staff, no whiteboard.
The moment it clicked was a text. A guide messaged our channel one night, lowercase, no punctuation: “i’ll take Maia’s driving tomorrow.”
Any AI can read that sentence. Almost none can act on it. Mine could, because the database underneath knew who Maia was, that her driving meant three specific movements with three specific vehicles and times, and what would break if they moved. It resolved the line, wrote the changes to my database, and the guides saw their new assignments. I confirmed it. I never opened a record.
That worked because we own that database. I decide what is allowed to write to it.
The booking software is the wall. It will let an AI read all day. What it will let an AI write, if anything, is a narrow slice, and the operating model is not in it. So anything that lives there stays manual, one record at a time, at human speed, no matter how smart the model gets.
Operators think they need better AI. Software companies think they need better features. Neither is the problem. The intelligence is already here, and the features already exist. What is missing is a place where an agent is allowed to act, with the rules of the business written down well enough that acting is safe.
Once you have that, it flips. The AI stops summarizing and starts moving the operation as fast as the operation actually changes. The AI is the interface. The operating model underneath is the moat that took years to build.
Booking software solved the sale. The operation that starts the second a booking lands is still run by hand, in every service business, not just rafting.
That is the seam. That is the thing worth building.
If you want to see it rather than read about it, here is the board that replaced the whiteboard. The layer that reads the crew’s messages is written up as the sensor layer, and how the whole thing is built covers the database, the conflict engine, and the voice interface.