Free tool · No sign-up

AI Policy Score.

Most organisations discover they do not have an AI policy at the moment a customer sends a security questionnaire. Thirteen questions later you will know which kind of company you are — and which of the five domains is holding the score down.

Progress

0 of 13 answered

01 · Governance
A named owner is accountable for AI use and can stop a deployment.
02 · Governance
Use cases are risk-tiered before they are approved to build.
03 · Governance
Model and prompt changes go through a reviewed release process.
04 · Data
Data classes that may never enter a model context are defined.
05 · Data
Personal or regulated data is redacted before it reaches a vendor.
06 · Data
Retention and deletion for prompts and outputs are specified.
07 · Access & Tools
Agents get scoped credentials, never shared or human ones.
08 · Access & Tools
Tool calls are allow-listed and deny-by-default.
09 · Access & Tools
Irreversible actions require human approval above a threshold.
10 · Incident Response
There is a runbook for an AI-specific incident, including model rollback.
11 · Incident Response
Every agent action is logged to append-only storage with an actor.
12 · Supply Chain
Model, plugin, and MCP vendors are reviewed for data use and training rights.
13 · Supply Chain
Retrieved and third-party content is treated as untrusted input.

Answer all 13 questions to unlock your score.

Method

How the score is calculated.

Each answer is weighted equally: documented and owned scores full credit, informal practice scores half, and absence scores nothing. The result is normalised across the questions you answered and expressed as a percentage, then broken out by domain so the weakest area is visible rather than averaged away.

The five domains — governance, data, access and tools, incident response, and supply chain — mirror the sections a vendor security questionnaire actually uses, which is why the output maps cleanly onto the evidence a buyer will ask for.

All free tools · Trust policy

AI Policy Score.

A named owner is accountable for AI use and can stop a deployment.

Use cases are risk-tiered before they are approved to build.

Model and prompt changes go through a reviewed release process.

Data classes that may never enter a model context are defined.

Personal or regulated data is redacted before it reaches a vendor.

Retention and deletion for prompts and outputs are specified.

Agents get scoped credentials, never shared or human ones.

Tool calls are allow-listed and deny-by-default.

Irreversible actions require human approval above a threshold.

There is a runbook for an AI-specific incident, including model rollback.

Every agent action is logged to append-only storage with an actor.

Model, plugin, and MCP vendors are reviewed for data use and training rights.

Retrieved and third-party content is treated as untrusted input.

How the score is calculated.