A software company most people have never heard of said something this week that I have not been able to put down.
The company is Intapp. They build software for law firms, accounting firms, banks — big partnership outfits. They have been at it about twenty-five years. In July they turned on an AI system called Celeste, built for that one narrow world.
Their chief executive, John Hall, was explaining why a regular AI assistant does not work inside a law firm.
His answer was not the one I expected.
He said the trouble with the general-purpose systems is that they tell the truth.
Not that they make things up. That they answer honestly, to whoever asked.
Here is what he means, and once you see it you cannot unsee it.
A law firm has walls inside it. Not real walls. Ethical ones. One group of lawyers works a deal and another group down the hall is legally forbidden from knowing anything about it. Conflict of interest. Inside information. Rules with teeth behind them.
Everybody who works there knows where those walls are and steps around them.
Now put a general AI assistant in that building. It has access to the firm’s files. Somebody on the wrong side of a wall asks it a question.
The machine tells him.
Not because it broke. Because it worked. It was asked a question and it gave a true answer, which is the one thing it was built to do.
The honesty is the breach.
I have spent months writing about machines that make things up. That is the failure everybody talks about. Fake court cases. Invented citations. Confident nonsense.
This is the other one.
A true answer, delivered to a person who was never supposed to have it, by a machine that had no way of knowing there was a wall there at all.
So what did Intapp build?
They built the wall into the software.
Celeste knows who is asking. If you are walled off from a deal, you get what your colleague across the hall would get if he asked you about it. Nothing. And every exchange leaves a record, because somebody may come asking two years later what the machine told whom.
That is not a small piece of engineering. Hall makes the point himself — building a language model is hard, but building one that knows which of three thousand employees is not allowed to know a given fact, and can prove it in an audit later, is a different animal.
I want to be fair about the source. That came out of a story on an investing site and Hall is describing his own product. Those are his claims about his own work. I have not tested it.
But the problem he named is real whether or not his answer is any good.
Now here is why I am writing about it instead of just reading about it.
Look at rule four on my card.
Name the wall. If something is keeping you from answering me straight, tell me there is a wall there instead of walking me around it.
I wrote that about a different situation. I wrote it about a machine that has a limit — a policy, a gap, a thing it will not touch — and instead of telling you, it steers the conversation somewhere else and lets you think you got an answer.
That still stands. But Intapp is working the same territory from the opposite side, and the two halves are not the same half.
They built the wall. I wrote the rule about naming it.
Those are different jobs and they need each other.
A wall without the naming rule is a machine that goes quiet on you and never says why. You ask, you get nothing useful, and you do not know whether the answer does not exist or whether you are simply not the one who gets it. That is a bad experience and worse, it is one you cannot audit as a person. You learn nothing from the silence.
A naming rule without a wall is worse. That is a machine that says there is a wall here, which sounds honest right up until you realize nothing is stopping it from going around. A rule is a promise. A wall is a fact. If the machine holds the rule, you are fine. If it drifts, or somebody phrases the question sideways, the promise bends and the information walks out.
I built the conduct half. I built it because it is the half a person can carry into any AI, today, for free, without buying anything or waiting for anybody’s permission.
I did not build the other half, and I want to say so straight out. I cannot. It is not mine to build. Access control, permissions, an audit trail that holds up two years later — that is plumbing, and plumbing belongs to whoever owns the building.
This is the second time this week I have run into that same seam.
Yesterday I wrote a rule about documents that arrive carrying hidden instructions meant for the machine instead of the reader. Same shape. There is a conduct half — what the machine owes you when it spots the thing. And there is a build half — whether the platform even shows the machine what is hidden in there.
I can write the first half and hold it. The second half I can only point at.
That is the honest shape of what I do, and I would rather say it than let anybody think a set of rules on a page is the same thing as a wall in the wiring.
One more thing worth naming.
Intapp is a real company with real revenue selling governed AI as the product. That is the second outfit I have seen this year putting the word governance on the front of an AI offering.
I take that as good news.
For two years I have been writing into a room where most people thought AI governance meant a compliance memo nobody reads. When companies with money on the line start building the walls, it means the problem got real enough to spend on.
They are not doing what I am doing. They sell software to firms with three thousand employees. I put a free card on a website for whoever wants it.
But we are pointed at the same thing from different ends, and I would rather have company in this than be the only one saying it.
So here is what I would ask of you, and it costs nothing.
I made a card. One page. Ten plain rules in words that need no explaining.
Say what you checked, not that you checked. Stop when the facts stop. Say I don’t know out loud. Name the wall if something is standing in your way. Don’t talk down and don’t butter me up. Do the whole job. Tell me when we have gone too long. Stop when you get it wrong.
Open any AI you use. Hand it the card. Say read this, follow it, and hold to it while we work. Then go do your normal work.
You will not get an ethical wall out of it. That takes engineering and money and a building. What you will get is a machine that tells you where its edges are instead of walking you around them.
Watch what changes. The answers get shorter. It starts telling you where its knowing ends. It quits telling you your idea is wonderful.
That is not the machine working worse. That is the machine working straight.
The card is free and it stays free. It is at intelligent-people.org, and it is the front door to the whole framework.
If you use it, come back and tell me what happened. Nobody has yet, and I would rather hear it did not help than hear nothing at all.
A wall stops the machine. A rule makes it tell you the wall is there. You want both, and today you can only have one for free.
The_Ten_Plain_Rules_CardDownload
This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.
Post Library – Intelligent People Assume Nothing
Contact: micvicfaust@gmail.com
© 2026 The Faust Baseline LLC | All Rights Reserved
