Blocking ChatGPT on the office network feels like a decision.
Mostly, it is a way of not finding out.
I understand why firms did it. In 2023 nobody knew what these tools did with what you typed, and a block was quick, cheap and easy to explain to a risk committee. Three years on, plenty of firms are still relying on that block as their AI policy. The proper version went on the back burner, and the block has been quietly teaching staff to use AI where nobody can see it.
What shadow AI actually is
Shadow AI is staff using AI tools the firm has not approved, does not know about, or cannot see. It is the AI version of the old shadow IT problem, when people signed up for Dropbox on a personal email because the official file share was too slow and raising it with the help desk had become a project for tomorrow.
There is nothing sinister in it. People are trying to get their work done, which is exactly what makes it hard to manage.
And it's common. The University of Melbourne and KPMG's 2025 study of trust in AI found that 60% of Australian employees admit to having hidden their use of AI at work. In a twelve-person office, that's seven people.
How it usually unfolds
What follows is a composite. It is not a client, and the times are illustrative. Every step in it is ordinary, which is why I use it.
Monday, 8:40am. The firm's web filter blocks ChatGPT and a list of similar sites. It has been doing that since 2023, and everyone knows it.
Monday, 8:52am. A senior associate has a forty-page trust deed to summarise for a client meeting at eleven. She opens the deed on her laptop, takes photos of the key pages on her phone, and uploads them to the free AI app on her personal account. The phone is not on the office network, so the block does not apply.
Monday, 9:20am. The summary is good. Better than good. She tidies it up and pastes it into an email to the partner. Nobody asks how it was done.
Wednesday. She does it again, with a client's estate planning notes. Then a set of financial statements. Her personal account's history now holds three clients' names, their family structures and their assets. None of it sits in a system the firm controls, and all of it is covered by whatever terms she clicked through on a Sunday night.
Six months later. She is offered a role at another firm and leaves on good terms. Her laptop is wiped and her logins are switched off. The phone, the app and the history go with her, because they were never the firm's to switch off.
The following year. A client asks, as clients increasingly do, whether the firm uses AI with their information. The partner says no, and believes it.
Nothing in that timeline involves a hacker or a breach notice, and that is the point. Client information leaves the firm through the most ordinary door in the building, and afterwards nobody can say where it went. You would spot the dramatic version from a mile away. This one never raises its voice.
Beware: A network block only works on the network. The moment someone uses a phone on mobile data or a laptop at home, it is gone. All a block reliably does is move the activity somewhere you can't see it.
The car keys problem
If you have raised teenagers, you already know this one. (If you have not, ask someone who has. They will not need the explanation.)
Tell a sixteen-year-old they cannot have the car keys, and you have not stopped them driving. You have stopped knowing whose car they are in.
Your staff are the same, and I do not mean that unkindly. They are not hiding AI because they are sneaky. They're hiding it because the official answer was no, the deadline was Friday, and the tool on their phone did in twelve minutes what used to take an afternoon. The firm never gave them a safe way to do something they had already decided was worth doing, so they found their own way. Discipline was never the issue here.
What works instead
Three things, in this order.
Give people an approved tool. One that runs under the firm's own accounts, with business terms that keep your data out of model training and keep it inside your Microsoft 365 tenant. Enterprise data protection is the phrase to ask your provider about, and it either applies to a licence or it does not. While the licensing here changes often enough that I would check rather than assume, a firm already on Microsoft's business plans is usually either covered or one step away. Is anyone in your firm able to say which plan you are on without looking it up? If you want the longer version of what to look for, I've written about whether ChatGPT is safe for work use and the difference between the free app and the business tiers.
Find out what's actually in use. On company-managed devices, tools like Microsoft Intune and Microsoft's app discovery features can show which AI services are being reached, and can block the ones you've decided against. Don't forget the browser, because a lot of AI arrives as an extension rather than a website, and browser extensions are a risk most firms underestimate.
Write down the rule. Short, plain and specific. Which tools are approved. What can never go into any tool. What to do with a personal account that already has client information in it.
I have argued before that tracking every keystroke is the wrong way to manage people, and I have not changed my mind. This is different. Knowing which applications touch client data is about the firm's systems, not about watching people. You can do one without the other, and in your position I would.
Tip: Before you change anything, ask your team directly which AI tools they've found useful. You'll get a more accurate answer when it's clear nobody's in trouble. Then make the most popular one safe, rather than banning it again.
The number is usually higher than the partners guess, and the tools are duller than they fear. Mostly summarising and rewriting, almost never anything exotic.
What this won't fix
Here's the limit. Everything above works on devices and accounts the firm controls. It does not reach the personal phone in the associate's handbag, and nothing ethical will. Detection tells you what is happening on your own equipment. The phone is a matter of culture, a clear rule and an approved alternative good enough that people stop reaching for the other one.
Less satisfying than a block, I know. It is also the only version I have seen hold up. The car keys come back the same way: not by hiding them, but by being the parent who gets the phone call.
The good news is that this is a measurable problem. If you'd like to know what's actually in use across your firm's devices, we can run a discovery report and walk you through what it finds. No judgement attached, and it's a starting point rather than a verdict.
Where these numbers come from
- Trust, attitudes and use of artificial intelligence: A global study 2025, Australian findings, The University of Melbourne and KPMG
Yener
