What's left when nobody logs in
I almost never open the platform my company runs on.
That is a strange sentence, because it is where the company lives. Forty four connected tables. 207,000 data points. Trips, guides, boats, vehicles, permits, assignments, and the rules that decide who is allowed to do what. If it vanished tomorrow we would be running a whitewater rafting company off a whiteboard again, which is exactly what we were doing two seasons ago.
I built it. I know every table in it. And I go weeks without looking at it.
I talk to it instead.
The part I did not expect
I did not set out to prove anything about software. I set out to stop losing my summers.
Two seasons ago our operation ran on a whiteboard. Thirty guides, six trips a day, three vehicles, and one dry erase marker. Every change rippled through everything, and the only place the ripples were understood was in my head. I was spending twenty to forty hours a week on planning alone, and I was still the single point of failure for a company that ran fine as long as I never got sick.
So I built the model. Not a chatbot. A database, with the real rules of the operation written into it. What a trip actually requires. Which guides are qualified for what. Which boats are committed. How the vehicles and trailers and shuttles have to line up. What changes when the water changes.
Last season my planning workload went from twenty to forty hours a week down to two to four, on the way to a record year.
Then a guide texted our channel one night, lowercase, no punctuation: i’ll cover the second driver tomorrow.
It resolved who the second driver was, what she was driving, whether he was qualified for it, and what else moved as a result. It wrote the changes. The guides saw their new assignments. I confirmed it from my phone.
I never opened a record.
That was the moment I understood I had built something other than a scheduling tool. I had removed the reason to log in.
The realization
Here is what I think is actually going on, and it took me most of a season to say it plainly.
Nearly every feature in business software exists because a human had to click something.
Someone needed to change a price, so there is a price screen. Someone needed to block a date, so there is a calendar with a right-click menu. Someone needed to add a staff member, assign a certification, set a rule, run a report. Each of those needs became a form, and each form became a feature, and the feature count became the sales pitch. Vendors compete on it. Buyers evaluate on it. Whole categories are ranked by it.
But the features were never the product. They were scaffolding for a human interface. Configuration used to be expensive because a person had to be walked through it, so vendors built five hundred paths to walk. That is what a feature list is: the accumulated cost of a human having to do the clicking.
If nobody is clicking, most of that surface was never doing the work. It was hosting it.
I want to be careful here, because this is where people start predicting things and I am not going to. Screens do not disappear. I still look at mine. But the screen stopped being where the work happens and became where I check that it happened. That is a real distinction and it changes what a system actually needs to be.
What is actually left
Take away the interface and three things remain. Not five hundred.
State that is true. What is real right now. Not what someone typed in March.
Rules that are right. What this business actually requires, including the parts nobody ever wrote down.
Actions that are safe. A way to change things that is typed, permissioned, reversible, and auditable, so a mistake is caught rather than discovered in July.
That is the whole system. Everything else was a way for a person to reach those three things through a screen.
And here is the part that surprised me most: the hard part is not the AI. The reasoning was the easy layer. The hard part is that in almost every business I have looked at, the state is not true and the rules were never written down. They live in the head of the person who has run the place for twenty years, and they cannot tell you what they are, because to them they are not rules. They are just how it is done.
Getting those out of a person is the actual work. It is an interview, not an integration.
The sharpest example I have
Availability.
Every booking and scheduling system on earth stores availability as a number a human typed. A capacity per departure, per slot, per day. It was correct at the moment it was entered, by a person who was busy, guessing about a day that had not happened yet.
On our operation, whether a trip can run tomorrow is not a number. It is a result.
It depends on how many boats are rigged and which are already committed. On which guides are working, and not just how many, because a guide who cannot drive the bus and a guide who can are not interchangeable, and a trainee is not a trip leader. On vehicles and trailers and who is shuttling. On the water. On permits. On a text somebody sent last night.
Change one of those and the real answer changes. Not at the start of the season. That day.
When the stored guess is low, we turn away trips we could have run. When it is high, we sell something we then have to solve at six in the morning. Both of those are expensive, and one of them is invisible, which is why almost nobody fixes it.
My model computes the real answer. It has all the inputs. It knows the rules. It is right far more often than the number in the box.
And then I open the booking software and type the number in by hand.
Where it stops
That last sentence is the whole problem, and it is smaller and more specific than “AI is coming for business software.”
I went and read what the sales channels actually accept. It turns out the big ones do not read a number off a shelf. They call the operator’s system and ask, in real time, several times during a single checkout, and again every half hour to keep themselves honest. Others take a push, and will accept updated availability, updated pricing, and updated product content whenever the operator’s system has something new to say.
So the demand side is already prepared to have a conversation with software. It is sitting there with an open endpoint, asking a question.
And the operation can be made ready. I made ours ready. It took a season and it was not easy, but it is done and it works.
Both ends are ready. The part in the middle is not.
The system of record between them will not tell me when something changes, and will not let me write the answer back. So the most sophisticated part of my company ends with me retyping its output into a box.
I have written elsewhere about the six layers of an operations stack and which of them are broken. This is the shortest version of it. The wall is in the middle, and it is not a technical limitation waiting on better models. It is a roadmap decision.
What I actually want
Much less than any vendor is selling.
Tell me when something changes. A booking created, moved, or cancelled should arrive as an event, not as an email I read later or a report I export on Monday.
Let me write the answer back. Not a spreadsheet import in the off season. An endpoint that accepts a computed number, with the reasoning attached and a way to express confidence.
Let me author the product once. The description, the price, the policy, the age minimum. I maintain those facts in four places today and I am the only reason they match.
That is two endpoints and a catalog. Not a rebuild. Not a migration. Not a five year roadmap.
I do not need anyone to compute capacity for me. That work is mine. It is where the actual knowledge of a business lives and I would not hand it over. I need the plumbing to carry the answer.
But won’t the labs solve this
It is the obvious objection, so I want to answer it rather than wait to be asked. By labs I mean the top tier model companies, Anthropic, OpenAI, Google DeepMind and the handful of others at that frontier.
They are building two things. Reasoning, which gets better and cheaper every few months. And protocol, a standard way for a model to reach a tool and take an action. Both are real, both are moving fast, and my entire operation runs on top of them.
Neither one knows that a trainee cannot lead a trip. Nobody at those companies has met my operation and they are not going to. They are making the engine cheaper every quarter. They are not making the map.
The protocol point deserves saying plainly too. A universal standard for agents to call tools can exist, and increasingly does, and a given vendor still has to decide to expose an endpoint. No lab makes that decision for them.
I will be honest about what does get commoditized. If what you built is a chat window over your own data, that is going to be free and built into everything, probably soon. I would not spend a year on it. But the better the reasoning gets, the more valuable a correct model of the business becomes, because it is the only scarce input left.
I also do not think being one of the few people who can do this stays true for long. It is a skill and skills spread. That is a reason to move now, not evidence of a deep moat.
Why I am writing this
Because I am fairly sure I am not the only one who has hit this wall, and I have not found many people saying it in these words.
I am not selling you booking software. I built this because I was losing my summers to it, and the more of it I build the more obvious it becomes that the missing piece is not intelligence. It is a system of record that can be talked to.
What I am doing about it is two things. I install this for other owner-operated service businesses, and I am writing down the part that does not scale yet, which is getting the rules out of the person who has them. The install is how I find out whether the method transfers or whether it only worked because I already knew every answer in my own operation.
If you run an operation and you have hit the same wall, I want to hear how you solved it. If you build the software these businesses run on, I would like to compare notes, because the two endpoints above are not a large ask and somebody is going to ship them first.
Here is what an install looks like, and here is how I work with platforms. Either way, get in touch.