It’s a Tuesday in July and it hasn’t stopped raining since before the first tram. Maria stands behind the counter of her cafe with five people rostered on and a room that’s stayed empty since the early rush died at nine.

She built this roster last Wednesday, from memory, the way she always does. Tuesdays are usually steady, so she put the numbers on. Now the rain has done what rain does, the regulars have stayed home, and she’s paying five casuals to wipe down tables that are already clean. Every one of them is on at least $31.19 an hour, the casual hospitality minimum since July 2025, and the till isn’t covering the floor. She sends two of them home early, which helps her wage line and stings theirs, and she drives home feeling like she’s just set fire to a few hundred dollars.

The next day the sun comes out.

The other guess

Wednesday is glorious, and Maria has rostered light because Wednesdays are quiet. By half past twelve there’s a line out the door, two people flat out on the coffee machine, and a group of four who wait, look at the queue, and leave for the place across the road. She doesn’t see them go. She’s too busy to. What she loses that lunchtime never shows up on any report, because covers you didn’t serve don’t leave a receipt. They just don’t happen.

Same cafe, same owner, same skill. Two days, two guesses, and both of them cost her: once in wages burned on a dead shift, once in covers she couldn’t serve. This is the quiet arithmetic of running a venue by gut. You’re never really wrong on average. You’re wrong in both directions, all week, and the averaging-out is exactly what hides the bleed.

Why the roster is always a little wrong

The roster is a forecast. That’s the thing nobody says out loud. When Maria sets next week’s shifts from memory on a Wednesday afternoon, she’s predicting the weather, the school holidays, the game at the ground down the road, the long weekend, and the mood of a few hundred strangers. She’s doing it in her head, between the morning rush and the supplier call, in about the four hours a week that rostering quietly eats.

Labour is roughly 40% of what it costs to run a hospitality venue, an indicative industry benchmark rather than a law of nature, and wages have kept rising. Trying to underpay your way back is genuinely dangerous ground now: since January 2025, intentional underpayment is a criminal offence in Australia, with company fines up to $8.25 million. So the roster is where the money is, it’s built on a guess, and the guess is made by the busiest person in the building. That isn’t a failing of Maria’s. It’s the job asking a human to do arithmetic humans are bad at.

What actually changes

Here is the honest version, with none of the hype the word “AI” usually drags along.

The signals that would have warned Maria about her rainy Tuesday and her sunny Wednesday already exist. They’re in her point-of-sale history, the same Tuesday last month and last year, wet and dry. They’re in her bookings. They’re in the weather forecast and the events calendar for her suburb. What’s never existed is the time to pull all of that together into next week’s shifts by hand.

An AI roster assistant does exactly that and nothing more. It reads the demand signals and it proposes a roster: fewer hands on the wet Tuesday, an extra barista and an extra floor person on the sunny Wednesday lunch. Then it stops. Maria opens it, sees the reasoning, and changes it. She knows one of her regulars is having a wake on the Thursday. She knows Jess can’t do Fridays anymore. She knows a hundred things no dataset holds. She adjusts it, and she publishes it. The machine never sends a shift to a single staff member. It hands Maria a better first draft than memory could, and she owns the roster the way she always has.

That’s the whole shift hiding in the word “decides”. The AI doesn’t decide who works. It kills the blank page. Vendor-reported figures put the gain directionally: each one per cent of forecast accuracy you claw back is worth about half a per cent off your labour cost. A Sydney newsagency-cafe that moved its rostering onto this kind of footing with Deputy cut weekly admin from around eighteen hours to a few. That’s a published customer result, not ours, but it’s the shape of the thing.

The catch nobody puts on the brochure

Whether any of this actually works for Maria isn’t decided by how clever the AI is. It’s decided by one unglamorous fact: whether her point-of-sale system will let an agent read her own data.

This is the part that gets skipped, so I’ll be blunt. If Maria runs Square, Shopify or Lightspeed, and rosters on Deputy or Tanda, she’s in luck, because those expose clean, self-serve APIs an agent can genuinely reach. If she runs one of the Australian hospitality tills, an Impos or an OrderMate or an H&L, she reaches her data through Doshii, the aggregator that fronts those partner-gated systems and acts as the one-to-many shortcut into that layer. And if someone has sold her rostering on RosterElf, she needs to know it only exports a CSV. It has no open door an agent can walk through, so the live roster data has to come from Deputy or the POS instead.

Then there’s the line every vendor uses: “our app already has AI built in.” Square’s does. Shopify’s Sidekick does. Now Book It ships a booking agent called “Sadie”. All true, and it doesn’t help you the way you’d think. That AI lives inside the product. It can’t be reached from outside your account, it can’t read across your roster and your prep and your reviews at once, and it can’t act on your rules. The reasoning layer is bought separately and pointed at your data through the systems that will actually open. Name the system before you buy the AI. That one sentence is most of the game.

The by-products that pay for it

Once an agent can read the demand signal, the roster is only the first thing it fixes.

The same forecast that sizes the shift also sizes the prep and the order. Over-prepping and over-ordering is quiet money: an avoidable kilo or two of protein binned each day runs a single venue somewhere around $5,000 to $10,000 a year, an industry estimate and labelled as indicative, part of the 250,000-plus tonnes of food Australian hospitality throws out annually.

Then there’s the phone. On a slammed Wednesday, Maria’s phone rings out, and a ringing-out phone is a booking walking to the venue next door. One Sydney operator had an AI phone agent field 859 overflow calls a month, turning about two hundred of them into bookings and handing staff back twenty to forty hours a week. A Tasmanian pub, the Beach Hotel in Burnie, had a booking agent answer 439 calls in a month and confirm 213 bookings worth over $43,000 in secured revenue. Both are vendors’ published numbers, not RMAI’s, and I’ll flag it plainly: these are other companies’ results, shown because they’re the pattern we build toward, not a promise about your venue.

The honest ledger

Two things I won’t dress up. First, every headline figure above belongs to someone else. RMAI is new to this sector, and we would rather show you a real third-party result with its name attached than invent a marquee number of our own. Second, the tool alone doesn’t pay you back. The roster assistant plus a manager who actually uses the better draft pays you back. Buy the AI, keep rostering from memory anyway, and you’ve bought nothing.

With that said: a focused build usually ships in three to six weeks in the $10k to $60k range, and it starts with a free 30-minute discovery call that walks one real path through your POS and roster data and names the one workflow costing you the most; any build is quoted fixed-scope after that, on your numbers. You test the pattern on one site before you commit to anything larger. And because around 85% of Australians worry about how their data gets used, it’s worth saying out loud: your data stays in your own tenancy, and it’s never used to train anyone else’s model.

Where to start

You don’t need a transformation programme to find out whether this is real. You need one honest look at which of your systems an agent can actually reach.

Two ways in, both low-risk:

• Read the map first. We’ve written a plain-English brief on the retail and hospitality software landscape: which of your systems (POS, bookings, roster) an AI agent can genuinely reach, by what mechanism, and what “partner-gated” really costs you. → https://realmindsai.com.au/guides/retail-hospo/
• Book a free 30-minute discovery call. We’ll take one real path, next week’s roster or the calls you’re missing, name the system underneath it, and show you the highest-value place to start, before anyone quotes you a build. → https://outlook.office.com/book/[email protected]/?ismsaljsauthenabled

It’s Wednesday lunchtime somewhere in your business right now, and a table of four is looking at the queue, deciding whether to stay. The only question worth asking is whether next week’s roster is still going to be a guess.