§ SanctumShield · Audit walkthrough · November 19, 2026 · about 8 minutes

How the AI governance audit works, step by step.

A plain-language tour of what you will be asked to know, do and decide, and who needs to be in the room.

Companion to the sample outputs. That page shows what you receive. This page shows how you get there.

By Lindsay Hiebert · Founder · CISSP

Dark card: SanctumShield, Audit walkthrough. Headline: See the method before you start. How the AI governance audit works, step by step: what you will be asked to know, do and decide. A grid of twelve cards, one marked in red.

Most organizations already use more AI than anyone has written down. A chat assistant here, a feature switched on inside the accounting software there, an agent a team set up last quarter. The AI governance audit is a guided review that turns that picture into a clear record: what you have, what you decided about it, and who is responsible.

You do not need to be technical to take it. The review asks questions in plain English, explains each one, and builds your documents from your answers. Here is the path.

The method in four moves

Every part of the review follows the same discipline. You will see it repeat for each tool and each agent.

Figure 1 · The method

  1. Move 1KnowList the AI your organization uses, including the AI nobody formally approved.
  2. Move 2DecideFor each one, record a clear decision.
  3. Move 3OwnName the person or role responsible for every AI system and every agent.
  4. Move 4ProveReceive four dated documents you can hand to a board, an insurer or an auditor.
Know, decide, own, prove. The first three are your work. The fourth is what SanctumShield produces from it.

First, choose your level

You pick how much time and depth you want to commit. Each level is complete in its own right. You can go deeper at any time, and every answer comes with you.

Figure 2 · Three levels, one result

Light

For an owner or manager without an IT lead. 27 plain questions about the AI your people use day to day, plus up to 4 follow-ups when an earlier answer calls for them.

Standard

Up to 53 questions: everything in Light, plus 23 more for organizations with someone who looks after technology or security.

Complete

A full working session for IT and security teams, with a record for each AI tool and each agent.

Same outcome at every level: all three cover the same twelve areas and produce the same four documents. Only the depth of the questions changes.

Twelve areas, at every level

Whichever level you choose, the review covers the same twelve areas of AI governance. Each one is named as a plain statement your organization should be able to make. The questions show how close you are to making it.

Figure 3 · The twelve areas

  1. 01We know which AI we useWhich tools are in use, including the ones nobody formally approved.
  2. 02Work accounts, not personal onesHow people sign in to AI, and what happens when someone leaves.
  3. 03Agreed tools on the right termsWhich tools the company chose on purpose, and on what terms.
  4. 04We know what information goes inWhat kinds of information people put into AI.
  5. 05Clear rules everyone has agreed toWhat is written down, and who has seen it.
  6. 06Clear ownership and decisionsWho is in charge of AI, and how decisions get made and recorded.
  7. 07AI that acts on its own is controlledWhat agents may do, how they sign in, and who answers for them.
  8. 08Fair decisions about peopleWhere AI helps decide anything about a person.
  9. 09Open and checked with customersWhen people are told AI is involved, and who checks what goes out.
  10. 10People know how to use AI safelyWhat training people have had for the AI they use.
  11. 11Work continues if AI stops or changesWhat you would do if a tool you rely on went away or changed.
  12. 12Ready if something goes wrongWho can shut a tool off, and what happens next.
For each area you will be asked what you know, what is in place, and who decided. “Not sure” is always an accepted answer.

How a session runs

You answer one area at a time, on screen or in a spreadsheet workbook. Where an earlier answer makes a follow-up relevant, it appears. Where it does not, you never see it. At the Complete level you also keep a short record for each AI tool and each agent. At the end you read your answers back once, then create your documents.

The layers of AI the review looks at

AI enters an organization in layers, and each layer raises a different question. The review works through them from the most visible to the least.

Figure 4 · Five layers, five questions

  1. Everyday AI toolsThe assistants and apps your people chose on purpose. Which ones, and for what work?
  2. Shadow AITools in use that nobody formally approved. What is being used that we never decided on?
  3. AI inside your softwareFeatures switched on inside products you already pay for. What arrived without a purchase decision?
  4. AgentsAI that acts for you: sends, books, changes, buys. What can it do without asking a person?
  5. Access and sign-inThe accounts and keys AI uses to reach your systems. What can it reach, and who granted that?
The catalog behind the tool questions holds 131 entries: 82 AI tools and 49 AI features built into common software.

Pinning an owner to every agent

This is the heart of the review. An AI system that belongs to everyone belongs to no one, so the audit asks you to put a role next to each one. You do not need to know how the agent works inside. You need to know what it is for and who answers for it.

Figure 5 · What an agent record looks like

Invoice follow-up agentExample
What it is for
Sends payment reminders to customers
Information it touches
Customer names, invoice amounts, email
What it may do
Acts, reversible: drafts and sends reminder emails
Owner
Head of Finance
Who can stop it
Head of Finance, IT lead
Decision
Approved
An invented example. An agent with no owner named goes on Needs you, the short list of what only a person can decide.

Decisions, written down

Governance is mostly a matter of deciding and keeping a note of it. For each tool or agent, the review records the decision, who made it and when. “Not sure” is an accepted answer. It is recorded as unknown, which is more useful than a guess.

Figure 6 · The decisions you can record

  • Approved
  • Declined
  • Access widened
  • Access narrowed
  • Replaced an earlier approval
  • Changed
  • Stopped
  • Risk accepted
DateWhen it took effect
Tool or agentWhat the decision is about
DecisionOne of the choices above
Who decidedA name or a role your board would recognize
Each decision becomes one line in your Decisions log. Next time, the review shows what changed since.

Before you start

A little preparation makes the review faster. None of this is required, and you can begin with what you know.

Useful to have on hand

  • A rough list of the AI tools your teams use
  • The main software your organization runs on
  • Any AI rules or policy you already have
  • A list of agents or automations, if you keep one
  • Your last set of answers, if you have run the review before

People worth asking

  • Whoever looks after IT or security
  • Team leads, who know what is used day to day
  • Whoever handles contracts and legal questions
  • The person who will sign off on the result

Help along the way

Each question comes with a short explanation of what is being asked and why it matters. The AI Governance Coach answers questions by voice or text while you work. If you prefer a spreadsheet, you can download the whole review as one workbook, fill it in with colleagues, and bring it back.

What comes out

When you finish, SanctumShield builds four documents from your answers.

  • Executive Risk ReportYour findings and next steps, each with an owner and a date.
  • AI Acceptable Use PolicyThe rules for using AI in your organization, ready to adopt.
  • Board MemoOne page in plain language for the people who oversee the business.
  • AI Posture SummaryOne page for insurers and partners who ask where you stand.
Each document is dated and third-party-verifiable through a Verification URL that stays valid for six years. See a full sample of each.

Then run it again

AI changes quickly. Tools get updated, agents gain new abilities, and people try new things. The review is built to be repeated: each run can start from your last answers, and your documents show what changed since last time. Over time the dated documents form a record of steady, careful oversight.

Where the audit fits

SanctumShield works from what you tell it. It does not connect to your systems, watch your network or talk to your agents, so there is nothing to install. It sits alongside the monitoring and security tools you may already have. Those tools show what AI did. The audit records what AI was permitted to do, and who agreed.

This walkthrough shows the shape of the review. The questions themselves, with their explanations and examples, are inside the audit.

Three ways to take the next step

Try the free Shadow AI Check for a first read in 12 questions, look at the sample documents, or begin your own review. For the questions whose answers change the most, read The Right Questions.

Continue exploring on SanctumShield

The full library — 290 pages of CISO-grade content.

SanctumShield maintains the deepest publicly-available reference set for mid-market agentic-AI governance — primary-source cited, continuously refreshed. Related deep-dives on this article’s topic:

How the AI Governance Audit Works, Step by Step | SanctumShield