INSIGHT: SHADOW AIShadow AI: when staff paste company data into AI tools you never approved

What is shadow AI?

Shadow AI is the AI version of shadow IT: any AI tool, feature or agent used for work that the organisation has not approved, cannot see, or has not set up with the right controls. In practice it takes four forms:

• Personal accounts on public AI tools. Someone pastes a contract into a free chatbot from their own login.
• Browser extensions and plug-ins. Writing and summarising helpers that read whatever is on screen.
• AI features inside approved software. A transcription or drafting feature that arrived in a product update and was switched on without review.
• AI agents connected to work systems. An assistant given access to a mailbox, calendar, files or CRM. A chatbot only sees what you paste in; an agent can read and act across everything it is connected to.

The National AI Centre’s guidance makes the same point from the other side: your AI register should include AI “embedded in other systems, like human resources and customer engagement tools”, not just the tools you bought as AI. [8]

How common is shadow AI?

Very. Three recent studies, each measuring something slightly different:

• Okta, fielded March 2026. 52% of knowledge workers had used unsanctioned AI tools at work, and 24% did so regularly, while 90% of executives were confident in their visibility of AI tools. In Australia, 94% of leaders were confident and nearly 60% of workers used shadow tools. The survey covered 292 executives and 492 knowledge workers who had used AI in the previous three months, across seven countries. [1]
• University of Melbourne and KPMG, surveyed November 2024 to January 2025. Of more than 48,000 people in 47 countries, 58% of employees intentionally used AI at work. Almost half admitted to using it in ways that contravene company policies, including uploading sensitive company information into free public AI tools. 57% said they hide their use of AI and present AI-generated work as their own. [2]
• IBM Cost of a Data Breach 2025. One in five of the 600 organisations studied reported a breach due to shadow AI, and only 37% had policies to manage AI or detect shadow AI. [3]

The numbers differ because the questions differ. The direction does not: in most organisations, more AI is in use than leaders think.

Why do staff use AI tools that were never approved?

Okta asked the people who use unapproved tools why. Only 6% said they did not know approval was needed. The rest were choosing speed and normality over process. [1]

Inside shadow AI

Why they use unapproved AIWhat they share with it

• Easier to use their own accounts80%
• The team already uses it; it is normal78%
• Approval is too slow or difficult57%
• Approved tools do not meet their needs49%
• Did not know approval was needed6%

quantana.com.au

Bar chart from Okta's AI Agents at Work 2026 survey. Why knowledge workers use unapproved AI tools, and what those using unapproved tools shared with AI in the past six months.

Item	Share of respondents
Why: Easier to use their own accounts	80%
Why: The team already uses it; it is normal	78%
Why: Approval is too slow or difficult	57%
Why: Approved tools do not meet their needs	49%
Why: Did not know approval was needed	6%
Shared: Internal messages and emails	54%
Shared: HR-related information	45%
Shared: Confidential documents, including financials and contracts	39%
Shared: Banking and payment information	28%

Two more findings explain the gap. Only 40% of employees in the University of Melbourne and KPMG study said their workplace had a policy or guidance on generative AI use, and only 47% had received AI training. [2] In the Okta survey, 65% of executives said their AI usage policies were very clear; only 43% of knowledge workers agreed. [1]

So shadow AI is mostly a design problem. If the approved path is slower, weaker or unknown, people take the other one.

What are the real risks of shadow AI?

1. Data leakage. Information leaves your control without a decision. Among Okta respondents using unapproved tools, 54% had shared internal messages and emails with AI, 45% HR-related information and 39% confidential documents such as financials and contracts; over 20% had shared login credentials and passwords. [1]
2. Intellectual property and confidentiality. Client documents and your own know-how go to services whose terms you have not reviewed, which can breach a client contract or a confidentiality agreement. IBM found shadow AI incidents compromised intellectual property in 40% of cases, against 33% across all breaches studied. [3]
3. Privacy. The Privacy Act applies to personal information put into AI tools and to output that contains it. The OAIC recommends organisations “do not enter personal information, and particularly sensitive information, into publicly available generative AI tools”. [6] IBM found shadow AI incidents compromised personal information in 65% of cases, against 53% overall. [3] If a leak is likely to cause serious harm, the Notifiable Data Breaches scheme requires you to notify the people affected and the OAIC. [7]
4. Inaccurate output. 66% of employees in the University of Melbourne and KPMG study relied on AI output without evaluating its accuracy, and 56% had made mistakes in their work because of AI. [2] When the use is hidden, nobody checks.
5. Cost when it goes wrong. Organisations with high levels of shadow AI saw an average of US$670,000 more in breach costs than those with low or none, in IBM’s 2025 study. [3] In its 2026 study, more than 20% of organisations reported a breach targeting AI models or applications, most often through compromised APIs, applications or plug-ins. [4]

For APRA-regulated banks, insurers and super funds there is one more: supervision. APRA’s April 2026 letter reported entities managing unapproved AI by relying “primarily on policy direction” rather than “enforceable technical restrictions”. [5] See our guide to AI under APRA.

General information, not legal advice. Your obligations depend on your industry and how you use AI.

How do you find shadow AI in your organisation?

Start by asking, then check the evidence. Each method sees a different part of the picture.

Put what you find into an AI register: tool, owner, purpose, the highest class of information it touches, and a decision (approve, replace or retire). The National AI Centre publishes an AI register template. [8]

How do you fix shadow AI without banning AI?

The tempting choice: block every AI site and send a stern email. The better choice: make the safe path the easy path.

A ban pushes use onto personal phones, where you see nothing. It also tells your best people that the organisation is not serious about AI. ChatGPT isn’t your AI strategy, but neither is pretending nobody uses it.

1. Approve good tools, per class of information. An enterprise AI assistant with suitable contract terms for internal work, and a clear rule for what may go where. Okta’s respondents went around IT mainly because their own accounts were easier; give them a work account that is just as easy, with single sign-on. [1]
2. Offer private AI for sensitive work. For personal, client or confidential information, run models inside your own environment, so the data never leaves it. SecGPT Vault is our private AI that runs on-premises with no internet connection. Our private AI guide compares hosted, private cloud and on-premises options.
3. Write a short, clear policy. Which tools, which information, who reviews the output, how to report a mistake. Start with our AI policy template.
4. Back the policy with technical controls. Restrict unapproved AI sites and extensions on managed devices, review OAuth grants, and add data loss prevention rules for uploads. This is the gap APRA called out. [5]
5. Make approval fast. 57% of Okta’s respondents said approval was too slow or difficult. [1] Publish a turnaround, such as ten business days, and a short list of what gets checked.
6. Train with real examples. Show each team what “confidential” looks like in their own work, how to check output, and what prompt injection and AI-assisted scams look like.
7. Treat agents as identities. Any agent connected to email, files or core systems needs approval, a named owner, its own login with least access, a log, and a way to switch it off.

Shadow AI checklist

Ten things to have in place by the end of the quarter:

• A named senior owner for AI use.
• An anonymous staff survey on AI use, run in the last six months.
• An AI register, including AI features inside existing software.
• At least one approved AI assistant, with work accounts and single sign-on.
• A private option for personal, client and confidential information.
• A published AI policy with information classes and a review date.
• Technical controls on unapproved AI sites, extensions and app grants.
• A tool request process with a published turnaround.
• AI training completed by everyone who uses AI at work.
• An incident route for AI mistakes and leaks, with no blame for early reporting.