AI, kept inside the shield

01The data rules come first

Before any tool is switched on, the data rules are set — in writing:

• Credentials never go in. Passwords, keys, and recovery codes are never typed into an AI chat. Not by us, not by your staff, not as a shortcut.
• Sensitive work runs in controlled accounts — business tenants with named users, not personal accounts and not public demos.
• Client data stays the client’s. What your business feeds an AI under our setup is fenced to your work — never pooled, never training material for somebody else’s model of your business.
• Treat every AI like a new hire with a photographic memory: clearance before access, and only the access the job needs.

02One badge per engine

Every AI engine at Henschel works under its own badge: its own app registration on each system it touches, with permissions scoped to exactly that job — one badge for Claude, one for Grok, one for James, per system, and credentials are never shared between them.

Why it matters: if one engine misbehaves or one key leaks, we revoke that one badge and the rest of the company keeps working. A shared master key turns one small problem into a company-wide incident. Badges turn it into a Tuesday.

03The god door

Some doors only a human opens. At Henschel, three things always stop at a person — the god door:

• Money. No AI moves it, spends it, or commits it.
• Erasure. No AI deletes people, accounts, or records.
• Consent. No AI grants new access or permissions.

And one rule sits under the door: blanket approvals are void. A general “sounds good” authorizes nothing specific. Each consequential action gets its own human yes, on the record. The system proposes; the human disposes.

04Everything on the record

AI work at Henschel is logged as it runs: what the system did, when it did it, and what it touched. That record is what lets us run the buyer’s test from the Work Floor — counted, logged, checked, metered — and it is what makes an audit a short conversation instead of an excavation.

The rule of the house: if it isn’t written down, it didn’t happen. An AI that works off the record is not a tool; it’s a rumor.

05Vetted tools only — no shadow AI

The biggest AI leak in most companies isn’t a hacker. It’s a helpful employee pasting company data into a free tool nobody approved. So the bench is named and known — the engines on our Technology Wall — and anything else needs a human decision before company data goes anywhere near it. When we set AI up for a client, the same discipline arrives with it: an approved-tools list, the data rules, and staff who know both.

06When something looks wrong

It stops, it surfaces, and it gets fixed in the open. AI output is a draft until a human checks it — nothing an AI writes speaks for your business unreviewed, and no AI talks to your customers without a person on it. When the machine is wrong, the record shows that too. Exceptions are part of the record, shown, not hidden.

♦Why this room exists

The owner wrote an essay about the deeper danger — tools that only amplify, with no ethical check and no correction mechanism. It is the reason this room was built before any of the fun rooms:

Read “Amplification Is Not Intelligence” →