How we work
What we let AI do, and what we never let it do
The question is not whether it makes mistakes. It is whether your system notices before anybody else does.
Most AI governance is a document that was written, approved, and never opened again. This is the version that is actually enforced, in the work, every day. Two lists, and the reason each line is on the one it is on.
Start with the honest part
AI writes code here. Every day, in production work, on sites that are live.
It is also wrong often enough that a company which trusted it would already have broken something. Those two sentences are both true, and everything below exists because they are both true.
So the question we get asked is whether AI can be trusted, and it is the wrong question. It cannot, and neither can a new hire in their first week, and neither can a tired person at six on a Friday. The question that matters is different.
The question is not whether it makes mistakes. It is whether your system notices before anybody else does.
The word means steering, not stopping
Governance comes from the Greek kybernan, to steer a ship, by way of the Latin gubernare. The helmsman was the kybernetes, which is also where cybernetics comes from.
It never meant preventing the ship from moving. It meant somebody with their hand on the tiller while it does. Most of what is sold as AI governance is an anchor, and the companies that buy it stop moving and still have no idea where they are going.
What it is allowed to do, and is used for every day
None of this is delegation of judgement. It is delegation of volume.
- <b>Write the first version of anything.</b> Code, a draft, a translation, a test. First versions are cheap now, and that changes what is worth attempting at all.
- <b>Cross reference at a scale a person cannot.</b> Checking a new article against every previous one for repeated passages. Checking a list of dependencies against a list of known problems. Work that is simple, enormous and genuinely boring.
- <b>Explain a system back to you.</b> Reading unfamiliar code and saying what it does is one of the things it is best at, and it saves the most hours of anything on this list.
- <b>Find the thing a person has read past ten times.</b> A fresh reader with no memory of having written it has an advantage no author ever has.
What it is never allowed to do
These are refused structurally rather than by policy. A rule that depends on remembering it is not a rule.
- <b>Hold a credential.</b> No key, token or password is handed to a model, written into a file it can read, or pasted into a conversation. Keys go from the provider's screen into the console that needs them and nowhere else. A model that has never seen a secret cannot leak one.
- <b>Write to a repository.</b> It does not commit, push, reset or rebase. It proposes changes and a person sends them. One line of discipline that removes an entire category of accident.
- <b>Delete anything.</b> Files are moved aside into a folder a person empties. The cost of a wrong delete is unbounded. The cost of a wrong move is thirty seconds.
- <b>Decide what ships.</b> Every publish is approved by a person who has read it. Not reviewed in principle. Read.
- <b>Touch personal data.</b> Not in a prompt, not in a test fixture, not in a log, not in a repository. Invented data only.
- <b>Be believed about a fact.</b> Anything stated as true carries a source or it does not go out. This rule has caught more real errors than every other rule combined.
Four real mistakes, because a post without them is marketing
Every one of these happened in the last two weeks, and every one was caught by the structure rather than by somebody being clever.
It replaced a working file with an older copy of itself. Confidently, and the result looked plausible. It was caught by comparing a checksum against the version in source control, and restored from there. Caught by a check, not by a reading. A person reviewing that file would have seen working code and approved it.
It put a function inside the wrong function. Everything parsed. Everything looked right. The page would have failed for every visitor. It never reached anybody, because the publishing step builds the site before it sends anything and refuses to send a broken build. The gate did the work, and the gate does not get tired.
It stated a legal date that had been superseded. A European obligation was described using wording that had been replaced months earlier. It read as authoritative, because that is how a model writes everything. The rule that facts carry sources is the only reason it was checked at all.
It reported a fault on a page that was correct. An automated check read the wrong row of a table and declared an error. Acting on it would have broken a page that had been right all along. The lesson there is about us rather than about the tool: a report is a claim, and a claim gets verified before anybody edits anything.
What this costs, and what it buys
These rules are not free. Not letting it commit means somebody runs a command. Not letting it delete means a folder fills up. Making every fact carry a source slows writing down.
What they buy is that a mistake becomes an incident nobody outside ever hears about, which is the only definition of safety that means anything.
And the thing worth noticing is that none of it is sophisticated. There is no monitoring product and no compliance platform in that list. It is six refusals and one build gate, and it would fit on an index card.
If you are deciding this for your own company
The question is not which tool to allow. It is which of your actions are irreversible when they go wrong, because those are the ones a person keeps.
Everything else can be delegated, and should be, because that is where the hours are. How that work is done, and what the first two weeks give you.
- Anything involving a credential.
- Anything that deletes.
- Anything that reaches a customer.
- Anything that touches a real person's data.
AI governance is not a document that says be careful. It is a short list of actions a machine is structurally unable to take, and a gate that refuses to publish when something is broken. Write down which of your actions are irreversible, keep those, and hand over the rest.
MK!
AIPathTech Editorial
There is no tracking on this site. No analytics, no cookies, nothing recording where you went or how long you stayed.
But I would like to know if something here was worth your time. If it was, press the heart. It tells me nothing about you. It is a number going up, and it is what tells me to keep writing.
If you want more of this, I am on X. @Queendemetriak