RD. Contact

AI Integration·Insights

"APRA just named 'shadow AI' as a failure. Here's what it means for a Melbourne fintech."

A few months ago APRA wrote to every bank, insurer and super trustee and said, plainly, that AI is moving faster than the controls around it. The phrase they used for staff quietly running unapproved tools was "shadow AI" — and they called it a governance failure, not a training gap. If you run a small regulated business in Melbourne, that letter is now your problem too.

A few months ago APRA wrote to every bank, insurer and super trustee and said, plainly, that AI is moving faster than the controls around it. The phrase they used for staff quietly running unapproved tools was "shadow AI" — and they called it a governance failure, not a training gap. If you run a small regulated business in Melbourne, that letter is now your problem too.

Here's the thing most firms miss. They read that letter, nod, and write another acceptable-use policy. That's the wrong reflex. APRA didn't ask for a PDF. It named enforceable controls as the gap.

The part most firms get wrong

The mistake is treating this as a people problem. "Tell staff not to paste customer data into ChatGPT." Sure. But shadow AI exists because your people already have a real job that AI makes easier, and you gave them no sanctioned alternative. So they found one. Banning the tool and hoping doesn't close the gap APRA is pointing at. It pushes the behaviour underground, where you can't see it or log it.

APRA's letter lists four failure areas: information security (prompt injection, data leakage, agent manipulation), governance maturity, supplier risk, and assurance. Look at the first one. Data leakage from an unapproved model isn't a hypothetical. It's the example they put first.

Why "just block it" falls over

I've watched this play out. A 40-person lender in town blocks OpenAI at the firewall, feels virtuous for a week, then finds three teams quietly using personal accounts on locked-down devices. The risk didn't go away. It stopped showing up in any report a board could read.

The uncomfortable truth: the tools are useful, your staff know it, and a ban you can't enforce is worse than no ban, because it lets leadership pretend the problem is solved.

The fix I'd actually build

Give your people a tool that does the job and keeps the data home. That's the whole move.

For most small regulated firms, that's a local-first AI box: open-weight models running on a single GPU in your office, retrieval wired to your own documents, drafting and summarising and searching with nothing leaving the building. I run one myself, a machine that cost less than a decent laptop. It handles the everyday work your staff are already sneaking onto cloud tools, and the data never touches a third party's training corpus.

On top of that box you layer the boring controls APRA actually wants: an inventory of every AI tool in use, a use-case register, a block on unapproved endpoints, continuous logging of what went where, and a kill-switch on any agent. Done right, that's board-ready reporting, not a vibe.

This isn't theory. It's the same local-first approach we build for health and legal clients who can't let a record leave Australia. APRA just made it relevant to a lot more balance sheets. And it answers the fear I keep hearing from executives: Writer's 2026 data found 35% of them couldn't immediately "pull the plug" on a rogue agent. Local deployment plus a real kill-switch is how you stop being in that 35%.

What it costs a Melbourne SME

The number that surprises people: a readiness assessment and tooling inventory runs A$15–40K. That's the floor of the work, and it's the part that closes the "we have no idea what our staff are using" gap. Build the controlled local stack on top and you're still under what a Big 4 retainer charges before anyone writes a line of code. And you own the box afterwards. No per-token bill following you forever.

Versus the alternative: wait for a supervisory letter, then pay a panic premium to prove you were compliant all along.

The question to ask this week

If APRA knocked on your door tomorrow and asked for a list of every AI tool your staff touched this month, plus proof the data stayed in Australia, could you produce it? If the honest answer is "I'd have to guess," that's the gap. Not a model problem. A control problem.

We do a fixed-fee AI-readiness pass that tells you exactly where you stand against CPS 234 and the April letter. Two weeks, a register you can hand a board, and a plan. Happy to walk a Melbourne founder through it. No deck, just the gaps.

Sources:

Ravijeet Dang — AI-first business leader. This piece also appears on Substack and LinkedIn.

← All insights