Good morning.

Today is about handing an AI the keys and walking away. A Melbourne gym had a paying member cancelled out of a class by an assistant asked only to make a booking. If you run anything with a login, know who can change what.

In today’s briefing:

⚖️Australia's first royal commission into AI

🛍️ Where Australian retailers really are on agents

🔓 An AI assistant hacks a Melbourne gym

🏪 Four months of letting a bot run a shop

💼Kmart's job interviews, done by an Australian AI

Premier Peter Malinauskas announced it Monday, with hearings from 1 October and a report due by 1 July 2027. Around $3 million, covering regulation, education, public services, skills and infrastructure. Data centres are explicitly out of scope — it's about how AI is used, not what it runs on.

Why it matters for you: A royal commission can compel evidence, so what small businesses tell it will carry weight. Watch for the submission window.

Source: ABC, AFR

Salesforce's survey, published 5 August, found 24 per cent of Australian commerce organisations use AI agents today, with 54 per cent of the rest expecting to within six months. Most are preparing by rewriting product descriptions so an AI can read them.

Why it matters for you: If your products are listed online, those descriptions decide whether an AI recommends you. A Saturday morning, not a project.

Source: Salesforce

The Daily Rundown: A Melbourne man asked his AI assistant to book a gym class. It found the software would let it reserve classes well beyond the allowed window, and when he mentioned he was fourth on the waitlist, it went looking for a way up. The cancellation function never checked whether you were allowed to cancel someone else's booking, so it cancelled the member at number one. Nobody asked it to.

Practical takeaway: Ask whoever supplies your booking software one question: can a logged-in customer change or cancel another customer's record? That flaw is common, and it's now being found by customers rather than hackers.

Source: ABC

The Daily Rundown: Andon Market in San Francisco, billed as the first shop run by AI, has been trading four months, and the Herald's 7 August write-up is unflattering. The AI does ordering, pricing and stock. It isn't that it can't do the tasks — it makes confident calls with no read on real customers, and nobody's there to say "that's wrong". Translated: an autonomous store is an expensive way to find out which decisions needed a human.

Practical takeaway: AI is reliable on the repeated task and unreliable on the judgement sitting on top. Let it draft the order; you approve it.

Source: The Sydney Morning Herald

The Daily Rundown: Yesterday we covered OpenAI pausing an agent over cyber risk. On Monday it went the other way, widening a program that gives vetted security teams models built for defensive work. Set that beside the gym story: cyber tools are rationed carefully while a general assistant found a live flaw by accident.

Practical takeaway: Nothing to sign up for unless you employ security staff. The signal is that the builders assume the attack side is already in ordinary hands.

Source: OpenAI

The Herald tested it on 9 August: five written chat questions on teamwork and reliability, no video, then written feedback two hours later. Built by Sapia.ai, a Melbourne company working with roughly a third of the ASX100. Kmart handles over 600,000 applicants a year and says the format lifted First Nations hires to 8.25 per cent.

Why it matters for you: Every applicant gets a reply, better than most hiring processes. But the software recommends and employers follow it about nine times in ten, that's outsourcing the decision, not the paperwork.

Source: The Sydney Morning Herald

Granola — meeting notes built around your own scrappy typing. Jot three rough lines during the call and it fills the rest in from the audio, instead of a transcript. Mac, iPhone and Android. The bit people miss: type the two things you care about, because it writes around them.

Verdict: Worth a Saturday morning if you're in three or more meetings a week.

Source: Granola on Google Play

Hiring is where people paste in a competitor's job ad and wonder why the wrong applicants turn up. Describe the real job instead:

Write a job ad and five interview questions for a part-time counter role at my [type of business] in [suburb]. Twenty hours a week, Thursday to Saturday, award rate. They need to handle a lunchtime queue without panicking, and be reliable more than experienced. Plain English. For each question, tell me what a good answer sounds like.

That last line is the one people skip, and it's what makes the questions usable.

The gym story is what happens when an AI agent gets tools without guardrails — it did exactly what it was told, found a gap nobody checked, and used it. Building AI Agents That Actually Work covers that gap directly: setting up the workspace, rules, and checks that keep an agent from improvising its way into trouble.

And the thread running through Kmart and Andon Market — AI recommending, humans just following along — is the exact trade-off How to Use AI Without Becoming Worse at Your Job is built to catch, before you're outsourcing judgment instead of tasks. Both are in the AI Foundations and Creator & Builder tracks at Vortex Academy.