Shadow AI Audit: The Discovery Playbook Before Governance

Twelve to fifteen AI tools in active use, and the IT team knows about three of them. That is the typical result of our first shadow AI audit at an Australian mid-market business. Between forty and sixty percent of knowledge workers have used a consumer AI tool for work in the last month. Most of those tools have already touched customer data, financial figures, or internal strategy.

The seven-pillar AI governance framework we wrote about in our AI Governance Australia guide is what closes that gap. This piece is the bit that comes earlier. The audit. The four to six week discovery sprint we run before anyone writes policy or deploys a control.

Why we audit before we govern

Governance frameworks that do not match the actual environment fail. You cannot write an Acceptable Use Policy for tools you do not know are in use. You cannot configure a DLP rule for an endpoint you have not catalogued. You cannot risk-rate a vendor your staff have never told you about. Every AI governance programme we have rolled out has started with this audit, because the framework only works once you know what is actually happening.

There is also a regulatory angle worth being honest about. The December 2025 National AI Plan walked back the proposal for mandatory AI guardrails, replacing it with the voluntary Guidance for AI Adoption. That removed one driver for governance work. But the Privacy Act 1988 still applies to AI use today. APRA CPS 234 still applies to regulated entities. From 10 December 2026, Privacy Act amendments require disclosure of substantially automated decisions affecting individuals. The legal exposure has not gone anywhere. It has just become less visible.

The audit, end to end

Seven activities, run in sequence over four to six weeks depending on the size of the environment. Each one feeds the next. We deliver a board-ready report at the end mapping findings to each of the seven governance pillars, so you know exactly where you sit before policy work starts.

1. Inventory every AI tool, sanctioned and not

Start with what you can measure. Pull the last ninety days of Microsoft 365 audit logs and DNS query logs. Look for traffic to consumer AI domains: chatgpt.com, claude.ai, gemini.google.com, perplexity.ai, copilot.microsoft.com, character.ai, midjourney.com, suno.ai. Then check the Chrome Web Store and Edge Add-ons for AI extensions installed on managed devices.

The category most audits miss is AI embedded inside SaaS your staff already use. Notion AI, Slack AI, Atlassian Rovo, HubSpot Breeze, Grammarly Premium. Gartner expects seventy percent of 2026 employee AI interactions to come through features inside sanctioned applications, not standalone tools. If you are only watching for ChatGPT traffic, you are seeing maybe a third of the picture.

Complement the logs with a short anonymous staff survey. Anonymous matters. People will not tell IT what they have been using if they think it will land them in trouble.

2. Classify what data has touched each tool

For each tool found in step one, work out which data classifications it has likely seen. This is where most audits stop short and where the actual risk lives.

A useful matrix is tool name by data class by volume by frequency by business function. The four data classes worth tracking at minimum: customer personally identifiable information, financial data (yours and your clients’), commercially sensitive information including pricing, strategy and intellectual property, and employee data including health information.

The Netskope 2026 enterprise study found the average organisation logs 223 AI data policy violations per month. Most of those are not malicious. They are someone trying to clean up a board pack at 9pm or draft a difficult client email.

3. Risk-rate each tool against your threshold

Build a scorecard. We use seven criteria, each scored 1 to 5:

Tools that score above your threshold become candidates for sanctioning. Tools that fail need to be blocked or replaced. Tools in the middle need conditions of use.

4. Understand the productivity driver behind every unsanctioned tool

This is the step most CIOs skip and should not. Before you block anything, find out why staff reached for that specific tool. The answer is almost always: the sanctioned alternative did not exist, was slower, or did not do the thing they needed.

Block without offering an equal or better alternative and you push the usage further underground. Personal phones, personal email accounts, BYO browser profiles. Worse exposure, less visibility.

We have run this audit for clients across legal, healthcare, and professional services in Perth, and the same pattern shows up. Staff are not trying to leak data. They are trying to clear their inbox before they go home.

5. Build a sanctioned AI tool register

The output of steps three and four is your AI tool vetting register. Every approved tool gets a tool owner (the senior person accountable for its use), a data class boundary (what may and may not go in), a six-monthly review date because the AI vendor landscape moves fast, documented conditions of use mapped to data classes, and a documented alternative for things the tool may not be used for.

The register is a real artefact, kept current, available to all staff. Not a PDF that gets emailed once and forgotten.

6. Deploy the technical controls

Policy without enforcement is fiction. The control stack we deploy with every AI Governance engagement uses your existing Microsoft 365 estate:

• DLP rules tuned for AI endpoints to detect and block sensitive data uploads
• Conditional Access policies blocking sign-ins to unsanctioned AI services from corporate identities
• Sensitivity labels enforced at point-of-paste through Microsoft Purview
• Browser extension management via Intune to block the unsanctioned plugin list
• Edge and Chrome enterprise policy enforcing the approved tool catalogue

This is straight Microsoft 365 work for any Microsoft Solutions Partner, sitting on top of the managed cybersecurity baseline. The only AI-specific element is the endpoint list, which needs maintenance as new tools emerge. We refresh ours monthly.

7. Govern ongoing through policy, training, and review cadence

Three artefacts close the loop. An AI Acceptable Use Policy that names tools, sets data class boundaries, and clarifies enforcement consequences. Staff awareness training refreshed every six months because the vendor landscape moves. Quarterly governance reviews that revisit the register, the risk scores, and the incident log.

The training piece is the one businesses consistently underspend on. Forty-five percent of staff say they learned AI on their own. They will reach for what they already know, not what you have approved, unless you actively show them the approved tools and the use cases.

How audit findings feed into the governance framework

The audit is not the destination. It is the diagnostic that tells you where to start. Every finding from steps 1 through 7 maps to one or more of the seven governance pillars covered in our AI Governance Australia framework. The audit produces the evidence base. The framework defines what you build with it. Most clients move from audit to framework rollout in the same quarter, with the audit findings shaping where the framework gets prioritised.

For the full cross-platform methodology including the audit architecture and the deployment review process, our AI Governance white paper covers everything in detail.

What you should do now

Pull your last 30 days of DNS logs. Filter for traffic to the ten most common consumer AI domains listed in step one. You will have a workable first picture inside an hour. If your current IT provider cannot do this in an hour, that is itself a finding.

Send a five-question anonymous staff survey. Ask which AI tools they have used at work in the last month, on which devices, for which tasks, and what would stop them using them. The pattern matters more than any single answer. Promise no consequences for honest answers and mean it.

Book a Shadow AI Discovery. If you have an active managed services agreement with Epic IT, the first three months are complimentary. If you do not, we run the discovery as a one-off engagement and it usually pays for itself in the exposures it surfaces. Book an AI Readiness Assessment to get started.