Two numbers came across my desk this week, and they belong in the same sentence.
Forty percent of organizations say they can’t trust AI because they can’t see how it works.
Only seventeen percent are doing anything to fix it.
Those numbers come from McKinsey, reported in the enterprise press two weeks ago. Not from me. Not from this archive. From the people who survey the companies actually buying and deploying AI.
Sit with the gap between those two figures for a moment.
Forty percent can name the disease. Seventeen percent are taking the medicine. The rest are standing in the waiting room, holding a diagnosis, hoping it treats itself.
It won’t.
The disease has a name in the industry. They call it the black box. An AI that gives you answers but no explanations. It performs the task, and you have no way to see how it got there or whether it should be trusted with the next one.
The article that carried these numbers made an argument I’ve been making in this archive for over a year.
It said we expect researchers to show their workings and cite their sources. We should expect the same of AI.
It said an AI system is like an employee. You hire it to do a job, and like an employee, you need to understand what it does in order to measure it.
It said trust doesn’t come from marketing promises. It comes from being able to see exactly what the AI did, why it did it, and what it used to get there.
And it said the productivity leap everyone wants from AI agents only arrives if a human stays in the loop for the decisions that matter.
I didn’t write any of those lines. An enterprise technology writer did.
But if you’ve been reading this archive, every one of them should sound familiar.
Compliance demonstrated, not declared. That’s been in the record here since April. It has a name: ATP-1. Any AI can state that it’s behaving. The standard requires the behavior to be shown and tested, not announced.
The human staying in the room for decisions that matter. That’s been in the record here too — and it’s about to get its date.
Two days from now, on July 4, a protocol called AGP-1 gets ratified into The Faust Baseline.
Here’s the part of that story worth telling.
AGP-1 started as something else. It was designed as a governance system for autonomous AI agents. Five gears. A severity-weighted governor. A machine that could watch the machine.
And then I stopped building it.
Because somewhere in the design, the obvious question landed. If the AI governs the AI with no human in the room, who exactly is being protected?
An autonomous governor inverts the entire point. The founding thesis of this framework is that a person stays at the wheel. Discipline is chosen and held by the operator, not delegated to the thing being disciplined.
So AGP-1 was rebuilt. Not as an autonomous protocol. As a gate standard. The human stays in the room.
That decision was made and published months ago. It sits in this archive with a date on it.
Now the enterprise press is telling forty percent of the market that agents are only trustworthy with observability and a human in the loop for key decisions.
I don’t claim they read my work. The evidence doesn’t support that, and this archive doesn’t make claims the evidence doesn’t support.
What I claim is the date.
This is called convergence, and it keeps happening. In June, a government settlement pushed the largest AI companies toward a shared safety gate before their models speak. This archive had that gate in writing before the event, under the name POVL-1. Now the enterprise trust conversation is arriving at demonstrated compliance and human-held governance. This archive had both, named and dated, before the articles ran.
One convergence is a coincidence. A pattern of them is a direction.
And the direction points at the twenty-three percent gap.
Forty percent of organizations see the wall. Seventeen percent are working on it. That leaves nearly a quarter of the entire market standing between the diagnosis and the treatment, waiting for something simple enough to actually use.
Here’s what I’d say to them, plainly.
The reason you haven’t treated the problem isn’t that you don’t care. It’s that everything offered to you is either a platform you have to buy, a consultant you have to hire, or a promise you have to take on faith.
A written standard is none of those things.
A written standard is a set of rules you hold, that the AI is tested against, in plain language, session by session. Compliance shown, not declared. Evidence before claims. A gate before the output. A human in the room for anything that can’t be undone.
That’s not a product pitch. That’s a description of what governance actually is, and it fits on paper.
The industry will keep converging on these ideas because the problem keeps forcing it there. Every survey will keep finding the same gap between naming the disease and treating it. And every month, the treatment sits published, dated, and plainly written, for anyone who goes looking.
Forty percent have the diagnosis.
Seventeen percent started treatment.
On July 4, this archive adds one more protocol to the record. Human in the room. Gate at the door. Date on the page.
The waiting room is optional.
Look no further than here. The Faust Baseline.
Post Library – Intelligent People Assume Nothing
The Faust Baseline™ — intelligent-people.org
Codex 3.5 | Twenty Protocols | Ratified and dated on the public record.
Contact: micvicfaust@gmail.com
Purchasing Page – Intelligent People Assume Nothing
© 2026 The Faust Baseline LLC | All Rights Reserved
