← All tools AI risk analysis

Prompt permit

In industry, no hot work starts without a hot-work permit. Same logic before handing a task to an AI: two questions, a criticality matrix, protection barriers — and a verdict. Nothing is sent or stored: everything runs in your browser.

No. PP-— Scope this task only

Step 1 — Information sensitivity

What is the most sensitive piece of information you would have to give the AI for it to do the job? The most sensitive piece always classifies the whole.

S1 — PUBLIC Already published or publishable as-is Vendor documentation, a standard, your website copy, a general trade question. S2 — INTERNAL Internal but mundane Low-stakes meeting notes, a document skeleton, a generic procedure without figures. S3 — CONFIDENTIAL Client or commercially confidential A priced offer, project schedule, client installation drawing, non-public production data. S4 — CRITICAL Secret, personal or regulated Proprietary know-how, non-anonymised personal data, export-controlled or strict-NDA information.

Step 2 — Impact of an undetected error

If the AI gets it wrong and nobody notices, what is the worst plausible consequence? Think HAZOP: reason on the deviation, not the nominal case.

I1 — NEGLIGIBLE Cosmetic, no consequence A clumsy sentence in a draft, an unusable brainstorming idea. I2 — ANNOYING Internal rework, lost time An internal document to redo, a wrong intermediate figure caught by team review. I3 — SERIOUS Client or contractual impact An error in a sent offer, a wrong spec passed to the client, a badly worded commitment. I4 — SEVERE Safety, legal or reputation A safety sizing, the tuning of a protection loop, legal advice followed unchecked.

Criticality matrix

Raw position from steps 1 and 2; residual position after crediting the barriers ticked in step 3.

Delegate With conditions Restricted Refused

Step 3 — Protection barriers

As in LOPA: each independent protection layer earns one notch on the axis it protects. A barrier only counts if it is actually in place, not merely planned.

Anonymisation before sending Names, clients, figures and references replaced by neutral placeholders: the AI works on the structure, not the data. SENSITIVITY −1 Company-approved tool A company account with contractual guarantees: no training on your data, controlled hosting, approved by IT security. SENSITIVITY −1 Review by a competent person The output will be reviewed before use by someone able to catch a substantive error — not just a proofread. IMPACT −1 Objectively verifiable output The output can be tested: a recheckable calculation, runnable code, a translation compared to the source. An error cannot slip through. IMPACT −1

Verdict — AI delegation permit

Complete steps 1 and 2 to get the verdict.

VALIDITY: this task, this scope, these barriers. Any change of data or stakes requires a new permit.

Going further

AI comparison: sovereignty, GDPR and pricing → to pick an "approved tool" with full knowledge.
Guide: choosing your AI assistant → the criteria that really matter.
GDPR checklist → if your prompts touch personal data.

An educational tool inspired by criticality matrices and the LOPA method. It replaces neither your company's information security policy nor your CISO or DPO — when torn between two levels, pick the higher one.