AI Audit Basics

The evidence to request when you audit an AI system, and how far to trust AI in your own audits

Every record arrived. The counts matched. Every value was wrong by a factor of 100.

The guide shows how that happens, and what to ask for so it doesn't get past you. You already know how to audit. Scoping, evidence, criteria and conclusions work the same way when the system is AI. What changes is what you ask for and how you test it, and AI is now showing up inside the tools you audit with, too.

AI Audit Basics is a short guide to both halves, with a one-page checklist you can take into your next engagement.

Get Your Free AI Audit Basics Guide


Name
Email
Company

By entering your contact info above, you are requesting updates about our upcoming training and access to our newsletters. You can opt out at any time. Your information stays private. destcert.com/privacy-policy

  • Seven evidence requests for an AI system, from the inventory to the vendor’s foundation model
  • The one question for each request that tells you whether the evidence holds up
  • How to trace a model version back to the exact dataset it was trained on
  • What to check before you rely on a full-population screen or an anomaly detection tool
  • What your workpaper should record when an AI tool did part of the work

Prompt Edits Count as Changes

Someone edits an AI assistant’s system prompt on a Friday so it refuses fewer requests. There’s no retraining and no change ticket, and nothing in the model registry moves. By Monday the system behaves differently, and the change process never saw it.

Model cards, AI bills of materials, data lineage records and prompt change logs are the new evidence requests. The guide covers what to ask for, what to compare it with, and when to be wary of what you’re handed.

What's Inside?

Illustration of a businesswoman holding a tablet and interacting with a glowing flowchart and shield icon inside a server room.

Part 1: Auditing AI Systems

Seven evidence requests, each opening with a situation you could meet on an engagement: AI bought on a corporate card that never made the inventory, a hiring tool used well past its model card, a drift alert nobody looked at, and a vendor that updated its foundation model over a weekend.

Illustration of two women reviewing a document while a humanoid robot points toward the future.

Part 2: Using AI in Your Audits

How to test with AI tools without claiming more than you tested. That covers reconciling the population, defining what counts as an exception, keeping targeted selection apart from a statistical sample, and keeping your continuous auditing separate from management’s continuous monitoring.

Owning the Output

What a reviewer needs to see in the workpaper, the confidentiality check before data goes into a tool, and a ten-minute exercise for checking generative output against its source.

One-Page Checklist

Every evidence request in one table: what to compare it with, when to be wary, and the question to ask.

For Auditors Now Asked to Audit AI

Audit teams are now expected to audit AI, and few people know how yet. If you’ve done audit work, you already have most of what it takes. The audit skills carry over, and the AI part can be learned. That’s the idea behind AAIA, ISACA’s Advanced in AI Audit certification, and behind this guide.

If your work is on the security management side rather than audit, AAISM is the AI credential built for that role.

John Berti, our co-founder, has helped create training materials for ISACA and taught courses for ISACA.

By downloading, you’ll also receive AAIA preparation tips and resources via email. You can unsubscribe anytime.

Copyright © 2026 Destination Certification Inc.

Victoria, BC, Canada