What is shadow AI, and what do you do about it?

Somebody in your business is already using AI for work. The question is whether you know which tools, on what information, and whether they would tell you if you asked.

The short answer

Shadow AI is when people use AI tools for their work outside of an official policy or approved approach. That is the National AI Centre’s definition, and it is the one to use. The tools are usually ordinary: a personal ChatGPT account, a browser extension that summarises pages, a meeting recorder someone installed for free, or an AI feature that quietly switched itself on inside software the business already pays for.

The word “shadow” makes it sound like misconduct. It is nearly always the opposite: a person with too much to do, finding a faster way, and not mentioning it because nobody said they could. The National AI Centre makes the same point. Shadow AI, in its words, is often a signal of unmet need, time pressure, or curiosity from early adopters.

How common it is

Common enough to assume. Microsoft’s 2024 Work Trend Index found that 75 percent of knowledge workers worldwide were using generative AI at work, and that 78 percent of those users were bringing their own AI tools rather than using ones their employer provided. At small and medium-sized companies the figure was 80 percent, and it held across every age group, not just the youngest staff.

Set that against what Australian businesses say about themselves. The Australian Bureau of Statistics reported that around 11 percent of small and micro businesses used AI in 2024-25, against 22 percent of medium and roughly 35 percent of large businesses. Those figures count AI the business knows about and chose to report. The gap between “11 percent of small businesses use AI” and “most people who use AI brought their own” is shadow AI, and most of it sits inside businesses that would say they have not started.

The National AI Centre’s own guidance opens from the same assumption: whether you have an approved policy or not, it is likely that some of your staff are already using AI at work without you knowing.

Why people hide it

If you want staff to stop hiding AI use, it helps to know why they do. The National AI Centre lists the reasons it hears: people may consider AI use cheating, or not want to be seen as lazy or less capable at their job. Microsoft’s survey put numbers on the same feelings. Of people using AI at work, 52 percent were reluctant to admit using it for their most important tasks, and 53 percent worried that using it on important work made them look replaceable.

That matters for the fix. Shadow AI is not mainly a technology problem or a discipline problem. It is a trust problem, and policies written as threats make it worse.

What the actual risk is

Two risks are real. The rest is noise.

Information goes where it should not. Generative AI tools invite you to paste things in, and the things that get pasted are customer emails, supplier contracts, staff details and financial figures. The Australian Cyber Security Centre’s small business guidance notes that some AI providers may use customer-submitted data to train or refine their models, depending on the configuration settings or the type of subscription, so information entered could be reused or disclosed in unexpected contexts later. The privacy regulator, the Office of the Australian Information Commissioner, recommends as best practice that organisations do not enter personal information, and particularly sensitive information, into publicly available AI chatbots and other publicly available generative AI tools. Its guidance walks through exactly the shadow scenario as a worked example: employees at an insurer experimenting with a public chatbot by entering a customer’s claim details, including health information.

The OAIC’s position is that the Privacy Act applies to all uses of AI involving personal information. Businesses turning over $3 million or less usually sit outside the Act, with exceptions including health service providers and businesses that trade in personal information. Being under the threshold changes what is mandatory. It does not change what a customer expects to have happened to their details, or what a client will ask when they do their own due diligence on you.

Unchecked output goes out under your name. The second risk has nothing to do with data. It is a quote, a contract clause, a product claim or a piece of advice produced by an AI tool and sent without a person verifying it. Shadow use makes this worse because there is no rule saying who checks, and the person using the tool has a reason not to mention that AI was involved.

What is not on the list: the AI itself. The tools staff reach for are, in most cases, the same tools you would approve if asked. The problem is the missing decision, not the software.

What not to do

The instinctive response is a ban. The National AI Centre’s advice runs the other way: turning shadow AI into productive AI means creating clear expectations, safe alternatives and open conversations, so people do not feel they need to choose between getting work done and doing the right thing. A ban with no approved alternative is precisely that choice. The work still needs doing, the tool still exists on everyone’s phone, and the only thing a ban changes is that nobody tells you.

The second mistake is the opposite one: deciding that because the tools are useful, the question can wait. It cannot, for the data reason above. One pasted spreadsheet is enough.

The four-step fix

The National AI Centre’s guidance on unapproved use boils down to three moves: create a safe space to disclose, provide guidance on how and when AI may be used and which tools are acceptable, and invest in training and policies. A fourth, the register, comes from its related guidance. Together they fit in a fortnight for a business under about fifty people.

1. Make it safe to say so

Start with an amnesty, said out loud. Nobody is in trouble for what they have already used. Then ask, person by person or in a short survey: which AI tools have you used for work, for what, and what did you put into them? You will learn more from this conversation than from any audit, and the answers are your use-case list, already ranked by the people doing the work.

While you are asking, check the places shadow tools leave tracks: company card and expense statements for AI subscriptions, browser extensions on work machines, and the AI features now switched on by default inside accounting, email, document and customer service software you already pay for.

2. Write down what exists

Put everything you found in an AI register: the tool, who uses it, what for, and what information goes into it. The National AI Centre publishes free register templates as a Word document and a spreadsheet. Its reasoning is direct: a register builds understanding and awareness of AI across the organisation, helps staff see where and why AI is being used, and helps reduce fear, misinformation and shadow AI use. The register is a transparency document. It does not replace risk assessment, but at small business scale it is where risk assessment starts, because you cannot assess what you have not listed.

3. Approve something, and set it up properly

Shadow AI ends when there is an approved alternative that is at least as good as the one people chose for themselves. Pick the tools, usually one general assistant and whatever is already embedded in your core software, on a business plan rather than personal accounts. Then do what the Australian Cyber Security Centre recommends: review the configuration settings, terms and conditions and privacy policies of any AI platform the business uses, and check the data handling and privacy policies of the vendor. The setting that matters most is whether your inputs are used for training, and on business plans it can usually be switched off.

The approved tool comes with one data rule, written in a sentence staff can repeat: personal, confidential and financial information does not go into any tool that is not on the list, and personal details are removed or anonymised first wherever possible. That is the ACSC’s own recommendation, and it is the clause that does most of the work.

4. Put the rule in a policy, and train on it

The data rule, the approved list, who signs off on new tools, when a person must check AI output, and how to report a problem are the core of an AI policy. The National AI Centre’s free policy template covers all of it, and the guide linked there sets out a one-page version. The OAIC separately says organisations should establish policies and procedures for their use of AI systems, so the policy also answers the regulator.

Training is the half that gets skipped. The ACSC’s advice is to train and remind staff on responsible use, especially around sensitive and proprietary information. For most small teams that is an hour, with examples from your own business of what may go in and what may not, repeated when a new tool is approved.

A thirty-day version

• Week one. Announce the amnesty, run the conversation, check card statements and extensions.
• Week two. Fill in the register. Choose the approved tool and turn training off in its settings.
• Week three. Write the one-page policy from the National AI Centre template. Name an owner and a review date.
• Week four. Train the team on it, using the real examples you collected in week one.

At the end, the register, the policy, the named owner and the review date are the whole of AI governance at small business scale. Shadow AI, done this way, turns out to be the fastest route into it, because the use cases, the tool choices and the data rules all come from what staff were already doing.

How to tell it is working

Three checks, answered from records rather than impressions:

1. Does the register match what is actually in use, including anything adopted since it was written?
2. Has anyone asked to add a tool, or reported a problem, in the last three months? Silence after a policy usually means the policy is not being used, not that nothing is happening.
3. Can a new starter tell you the data rule in their first week without looking it up?

When outside help is worth it

Most businesses can do the four steps themselves, and should. Help earns its place when the conversation in step one turns up more useful AI work than anyone expected and the business wants to choose and build deliberately rather than tool by tool. That starts with ranking the candidates, which is what AI for small business covers, and it is the groundwork our AI foundations work is built around: team training, use case identification and a tool stack chosen on purpose, in accounts the business owns. The AI agency page shows how that fits with automation and ongoing operations as one practice.

If you would rather start with a view of where AI would pay in your business specifically, the free AI use-case audit ranks the candidates in plain English, with no obligation attached. Bring the list from step one. It is usually the best input we get.