Your staff are already using AI. You just don't know which bits.
Somewhere in your business this week, someone pasted something into ChatGPT that you’d rather they hadn’t. Not out of malice. They were behind, the thing needed to sound better than they had time to make it sound, and there was a tool on their phone that would do it in nine seconds.
You probably won’t find out. That’s the part worth sitting with.
Last week I wrote about what should never go into a general AI tool. That post assumed you were the one doing the typing. In a business with staff, you usually aren’t.
The numbers are worse than most owners expect
PagerDuty published a survey of 1,250 office staff in June. Non-technical roles only, and run deliberately at large companies — minimum $500 million in revenue, the kind with IT departments, acceptable use policies and mandatory training modules. Two hundred and fifty of the respondents were Australian.
Two thirds had used AI tools at work they believed weren’t permitted. Eighty-eight per cent had put work information into a public tool. A third had entered customer data.
Those are the organisations with the guardrails.
IBM’s 2026 breach report shows where that ends up. Security incidents involving shadow AI — tools nobody had sanctioned — more than doubled this year, to 43 per cent from 20 per cent. The breaches that involved them cost more than the ones that didn’t.
Don’t ask this in a meeting
The obvious move is to get everyone in a room and ask what they’ve been using. Don’t. Nobody is going to tell a room with their manager in it that they pasted a client’s scope of works into a free ChatGPT account on their phone. You’ll get a tidied-up list, you’ll conclude the problem is small, and you’ll be wrong.
Split it in two instead. The awkward questions go out anonymously beforehand. The interesting one gets asked live.
First, an anonymous poll. Four questions, no names, nothing that ties an answer back to a person:
What have you used AI for at work in the last month? Anything at all.
Which account were you signed into — work or personal?
What did you actually paste in?
How long would that task have taken you without it?
In the same message as the poll, say plainly that nobody is in trouble for anything they put in it and that none of it will surface in a performance conversation. Then mean it. If one person gets pulled up over an answer, you’ve burnt the only mechanism you had, and the next time you ask you’ll get nothing back.
One caveat for smaller teams: in a team of six, an anonymous poll isn’t really anonymous and everybody knows it. If that’s you, don’t pretend otherwise. Say out loud that the amnesty is the point rather than the anonymity, and keep the questions answerable without anyone having to describe a job only they touch.
Question three tells you where your risk actually is. Question four is the one that pays for the whole exercise.
Then run a workshop, not a briefing. You walk in already holding the answers, which changes the room completely. Rather than prising information out of people, you’re showing them what the team said collectively, and asking the one question that’s safe to ask out loud:
What do you wish you could do with AI that you’re currently doing by hand?
That’s the question people actually want to answer. It’s aspirational instead of confessional, nobody has to own up to anything, and it’s where your build list comes from. It’s also where the guardrails get written, and rules a team helped set survive contact with a deadline far better than rules that arrive as a policy from above.
What comes out of it is one page
Not a policy document. Nobody reads those, and yours will be out of date by Christmas. One page, three lists, in the language your team actually uses — and written in the room, off the back of what the poll turned up.
Green — go ahead, with an approved use. A use you’ve already approved and recorded in your tools register: the tool, the account and the type of information. A good starting point is drafting, rewriting and brainstorming with public or made-up information in approved tools. Once a use is approved, nobody needs to ask again each time.
Amber — ask first. A new tool or account, a different type of information, switching on a connection, or letting AI take an action. Internal documents, commercial details and meeting recordings need assessing before they’re approved. Describe the task and wait for the answer. In the meantime, use a public or made-up example, or do it without AI.
Red — do not proceed. Work information in unapproved tools or accounts. Passwords, access codes or other secret credentials in an AI prompt. Personal or confidential information unless that specific use has been assessed and approved. A paid subscription or personal account alone isn’t permission. And AI making the final call on someone’s employment, health, finances or safety. As a rule of thumb: anything you’d have to ring someone up and apologise for if it leaked.
Then, underneath, the part most policies leave out — who to ask when it isn’t obvious. Name a person. Without a name, everything ambiguous quietly defaults to green.
Which account matters more than which tool
Most of the worry I hear is about picking the right tool. The bigger lever is usually which account it’s being used from.
The consumer version of a tool and the business version of the same tool can be handled very differently — what’s retained, for how long, whether it’s used for training, who at the vendor can see it. Same logo, different terms. So it’s worth reading the terms of the actual plan you’re on rather than the ones you assume you’re on, and worth making sure the work happens in an account your business controls and can switch off when someone leaves.
That last part catches people out. If the only record of six months of work sits in a personal account, it walks out the door with the person.
What you’ll find when you do this
You’ll get a policy out of it, and that’s the smaller half.
The bigger half is the list from question four. Expect a version of this: a report someone rebuilds by hand every Monday from three different systems. Quotes bottlenecked on one estimator. Invoices being re-keyed from one system into another. Debtors that slip because chasing them is nobody’s actual job.
That’s not a policy problem. That’s an operations problem your team has been quietly patching with whatever tool they could find on their phone. Which, honestly, is initiative — it’s just initiative without a seatbelt.
A poll, a workshop and one page. It’s the cheapest piece of AI work you’ll do all year, and it has to come before the rest of it.
If you’d rather not run this yourself — and there’s a fair argument that the poll lands better coming from outside the business — it’s how an AI Readiness Day starts. The poll goes out beforehand, the workshop runs on the day, and you finish with the one page, a costed list of what’s worth automating properly, and a team that’s been trained rather than told. $3,500 + GST, with a credit towards a larger project if you commission one within 30 days.
Sources: PagerDuty Shadow AI Survey, June 2026 (n=1,250, conducted by Wakefield Research across the US, UK, Australia and Japan); IBM Cost of a Data Breach Report 2026.
