Shadow AI: Your Staff Are Already Using Tools You Never Approved

5 min read · 4 October 2026

This week: shadow AI. Your team is probably already using tools you never approved.

Hello,

The story that stuck with me this week is about shadow AI. A CIO report published on 2 October pulls together fresh research on small business AI use, and one figure stands out: an estimated 49 per cent of workers are using AI tools their employer has not approved. Not a handful of rogue staff. Roughly half.

Last week we looked at new data on how much AI scheduling and payroll can save. This week is the uncomfortable follow-up. The tools are working, people are using them, and in a lot of small businesses nobody has decided what those tools are allowed to see. That is the gap this edition is about.

Structure, not AI noise. Let's get into it.

The headline: AI is in your business whether you chose it or not

Bluevine's 2026 Small Business AI Trends Report surveyed 942 US small business owners in April. It found 74 per cent are using or testing AI. That number is not the interesting part anymore. Everyone has heard it.

The interesting part is what sits underneath. Only 22 per cent of owners feel completely confident that AI can handle routine tasks without human oversight. Data security is now the top barrier, cited by 33 per cent, up from 23 per cent the year before. And 48 per cent of the businesses using AI say they are getting back at least four hours a week.

Put those together and you get the shadow AI problem. Owners are cautious. Staff are not waiting. The CIO piece describes four stages of adoption: shadow AI (people using tools on their own), licensed but unstructured (you bought access, nobody mapped it to real work), structured enablement (specific tools for specific daily tasks), and governed and embedded (security controls and a paper trail). Most small businesses I talk to sit somewhere in the first two.

I would treat the 49 per cent as a signal rather than a precise count. It is a second-hand figure from a US-focused analysis, and I have not seen the underlying data. But the direction matches what operators tell me. If a staff member is pasting customer details into a free chatbot to write a reply faster, they are not being reckless. They are solving a time problem with the nearest tool. The failure is that nobody gave them an approved option or a clear rule.

This is not a tool problem. It's a systems problem.

Tools worth trying this week

AI Receptionist (BT Business)

BT launched this on 2 October as the first product in its new AI portfolio. It answers calls 24/7, handles routine enquiries, books appointments into Outlook or Google Calendar, and transcribes the call. It is a UK launch on UK phone systems, so it is not something you can buy in Wollongong today. I am including it because of the research BT quoted: 67 per cent of surveyed UK small businesses miss at least one potentially valuable call a day. If you run a trade, a clinic or a salon, that missed call is a missed booking, and Australian equivalents of this category are worth a look.

Dots (OpenAI)

Announced at OpenAI's DevDay, Dots are persistent agents inside ChatGPT that keep working on a task after you close the window. Details are still thin, and one write-up says they launched for Pro users and excluded the EU, UK and Switzerland at first. I am not certain what the Australian availability is, so check before you plan around it. The idea is worth understanding now: software that keeps going without you watching is exactly where an approval rule matters most.

ChatGPT Space (OpenAI)

A shared workspace where a team and its agents work from the same project materials. For a small business, the useful angle is not the agent part. It is having one approved place where staff use AI on shared files, instead of ten private chats you cannot see.

GPT-6.1 Sol (OpenAI)

A cheaper model OpenAI says approaches its top-tier GPT-6 Astra on several evaluations at around one-fifth of the token price. That is a developer and API story, so most owners will feel it indirectly as cheaper tools built on top. Worth knowing, not worth chasing.

Decisions API (OpenAI)

In limited preview. It is built for narrow, repeatable choices with fixed outcomes: approve, reject, escalate, ask for more information. If you have a job like triaging enquiries or checking a claim against a rule, this is the shape of automation that tends to hold up. It is for developers right now, so treat it as a preview of where tools you already use are heading.

If you are still working out where any of this fits, the pillar guide to implementing AI in a small business walks through the order of operations. Workflow before automation. Always.

Quick win: the 30-minute AI amnesty

You do not need a policy document to start closing the shadow AI gap. You need one honest conversation and one page.

1. Ask the team what they already use. Frame it as an amnesty, not an audit. No one is in trouble. You want the real list.
2. Sort the tools into three buckets. Approved, approved with limits, and not approved. Be specific about why.
3. Write a one-page data rule. What can never go into an AI tool: customer personal details, bank details, staff records, anything under an NDA.
4. Give people an approved option. A rule with no alternative just pushes the habit underground.
5. Put a review date in the calendar. Tools change monthly. Check the list again in 90 days.

Watch out for

Business Victoria's guidance on AI scams draws on Scamwatch data. It notes that false billing was the most frequently reported scam against small businesses, and investment scams caused the most financial loss. AI is making all of them more polished. Their framework is worth pinning up: Stop, Check, Protect.

1. The invoice that looks exactly right

AI tools can now produce polished fake invoices with the right branding, the right tone and a plausible supplier name. Do this: match every invoice to a purchase order or an agreed job before paying, and have a second person approve anything above a set dollar amount.

2. The supplier who "changed" bank details

Payment redirection is on the Scamwatch list for a reason. A convincing email says the bank account has changed and asks you to pay the next invoice there. Do this: confirm any change of bank details by phone, using a number you already had on file, before a cent moves. Make it a written standing rule so no one has to make the call under pressure.

3. The cloned voice of someone you know

Voice cloning is being used to impersonate managers and suppliers asking for urgent transfers or confidential information. Do this: agree an internal check for urgent payment requests, such as a call-back on a known number or a pre-agreed question only the real person would know. If you are caught out, IDCARE offers free support for cyber incident recovery.

Big picture: the gap has moved from adoption to structure

A year ago the question was whether small businesses would use AI at all. That question is settled. Three-quarters are in. The CIO article makes the point well: the gap is no longer how fast businesses want to move, it is how safely they are set up to move.

That Claude figure comes from Bluevine's own US customer base, so do not read it as a market-wide ranking. What it does show is how fast the tool landscape is shifting under your feet. Your team may be using three different assistants this quarter and a fourth next quarter. A rule tied to one product will be out of date by Christmas. A rule tied to the type of data, and the type of job, will not.

That is why the businesses getting real value are not the ones with the most subscriptions. They are the ones who decided, in plain English, what AI is for in their business and what it never touches. If you want a second set of eyes on where to draw those lines, the AI automation for Wollongong businesses page explains how we approach it with Illawarra operators.

Workflow before automation. Always.

Book Your Free Strategy Call

Run the amnesty this week. Thirty minutes now beats finding out the hard way. See you next week.
Ashley