It’s two in the morning and the food-safety manager is standing at a whiteboard with a marker in one hand and her phone torch in the other. A supplier has just flagged a possible listeria problem in a batch of cream that went into three days of production, and she is trying to work out, from delivery dockets and a spreadsheet somebody else maintains, which finished pallets that cream ended up in, and which of those pallets have already left the loading dock.

The product is still out there. Somewhere tomorrow morning a stranger will open a fridge and reach for it. That’s the race she’s in, and right now she is losing it to a filing cabinet.

The race

A recall is a race between a contaminated batch and a stranger’s kitchen, and it is not a cheap one to lose. FSANZ has put the direct cost of an Australian food recall at around A$10 million before you count a cent of brand damage; analysts at Esko and Lumafield reckon the full bill, once you add lost production, air freight, retailer penalties and the scramble, runs three to five times that. In 2025 FSANZ coordinated 92 recalls, a touch above the ten-year average of 87. And 38% of them came down to undeclared allergens, the kind of failure that traces back to a packaging error, an accidental cross-contamination, or a supplier check that didn’t catch a changed ingredient.

Here’s what makes the 2am version of the race so brutal. Under Standard 3.2.2 of the Food Standards Code, an Australian operator has to be able to trace one step up and one step down, supplier in, customer out. On paper that sounds like a folder. In a real plant it’s dockets in one system, batch records in another, dispatch in a third, and a spreadsheet that lives on one person’s laptop. The US benchmark for how fast a regulator expects the records pulled is 24 hours, and while that’s an FDA figure rather than Australian law, it’s a fair mirror of the clock any buyer or auditor will hold you to. When the thread from ingredient to pallet lives in paper, you can’t pull it in 24 hours. So you do the safe thing and recall wide, pulling stock you didn’t need to pull, because you couldn’t prove which pallets were clean.

The bleed underneath

The 2am hunt is just the loud version of something that runs quietly every day.

Walk the same plant at eleven in the morning. A QA officer is reading a supplier’s certificate of analysis against the approved master spec, line by line, checking that the moisture and the micro counts and the allergen declarations all sit where they should. Ten to fifteen minutes a document, on operator and vendor benchmarks, and there are a lot of documents. Down the corridor the order desk is retyping orders that came in by email, PDF and voicemail into the system by hand, fifteen to thirty minutes each on the numbers Arnott’s reported from its own automation work. And underneath all of it, the lot records that would answer the 2am question, if anyone ever had the hours to knit them together, sit scattered across spreadsheets, PDFs and inboxes.

None of this is anyone slacking. It’s skilled people spending their day carrying information between systems that won’t talk to each other. And there are fewer of them every year: BDO’s 2025 survey found 75% of food and drink manufacturers reporting a skilled-labour shortage, up from half of them in 2023. The manual grind and the shrinking bench are the same problem seen from two angles, and the recall is the day that problem sends you the bill.

Why the button on the box won’t save her

If you’ve been sold anything about AI in food this year, it’s probably the button already built into software you own. Copilot in your ERP, Just Ask Xero in the ledger, an AI HACCP builder in your food-safety app. They demo well. They will not help her at 2am.

We spent June 2026 mapping which systems in a real food and beverage stack an outside AI agent can actually reach, and the finding was consistent enough to plan around: the reachable ground clusters at the ledger and the ordering edge, and the deeper you go into food-safety and process-manufacturing software, the more it closes up. The systems that open cleanly are a tight set. Your ledgers, Xero and MYOB, and the Microsoft 365 document layer. The B2B ordering hub Ordermentum. Inventory in Cin7 Core or Unleashed. Point of sale in Square. And one genuine food-safety standout, SafetyCulture, which lets you read and write inspections and templates and fire webhooks. That layer an agent can work with.

Almost everything else in the food-safety and QA tier, past SafetyCulture, is a closed door. The mid-market and enterprise ERPs, Business Central and Pronto Xi and NetSuite and SAP, have real APIs but they’re tenant-gated and integration-team heavy. Tools like FoodDocs, TraceGains and SafetyChain are closed SaaS or partner-gated, whatever their marketing says about an “open API”. Ordering tools like Fresho and Foodbomb are consumers, not doors; they push locked invoices into your ledger and expose nothing of their own, so you reach them only through the ledger they feed. And the vendor AI, all of it, lives inside its own product. None of it can read across your orders and your COAs and your lot records at once, or act on your rules, because none of it can be reached from outside the account it ships in.

That’s why “we integrate with standard APIs” is close to meaningless in this sector unless somebody names your systems out loud. The honest version is that the reasoning is bought once and pointed at whichever of your systems will actually let it in, and the first useful thing anyone can tell you is which of yours will.

What actually changes

So put the button aside and look at the thread from ingredient to pallet, which is the unglamorous thing nobody markets.

An assistant sitting on the reachable layer keeps that thread alive as it’s created, instead of leaving it to be reconstructed at a whiteboard after midnight. At the order desk, it reads the emailed and voicemail orders, pulls out the lines, checks them against the SKU master and flags the ambiguous ones for a person, rather than having them retyped. At goods-in, it reads each supplier COA against the approved master spec and flags what’s out of range, refusing to sign off when the source document is incomplete. Before a packaging run it checks the label against the master and the Plain English Allergen Labelling rules. And it keeps a live map of lots, suppliers, receiving, transformation and shipment, so that the traceability pack is assembled continuously and the missing records show up on a Tuesday, not at 2am.

Then, on the night the call comes, it drafts the recall picture. It runs the suspect lot through the map, builds the chronology, and shows what shipped where and which fields are incomplete. It does not decide anything. Nothing is auto-recalled, nothing auto-released, nothing auto-ordered. The food-safety manager still makes the call, still signs it, still owns it, because a recall is not a place for a machine to be the one who’s right. What changes is that she makes it from a picture that took minutes to assemble, with the pallets she doesn’t need to pull already ruled out, instead of guessing wide because the paper wouldn’t answer in time. The trace takes minutes, not a sleepless night. The deciding stays hers.

The honest ledger

Two things I won’t dress up. RMAI has not yet shipped this to a named food-sector client. We’re early here, and every headline number in this piece is somebody else’s published result, labelled as such, illustrating the pattern rather than proving our track record. When Arnott’s automated its order and invoice processing with UiPath, it reported clearing 75% of previously manual orders, saving up to about 60 hours a week, and reaching payback within roughly ten months; that’s Arnott’s result, not ours. When the fresh-produce wholesaler Fisher & Woods moved to Fresho’s digital ordering, it reported saving around 50 hours a week and cutting delivery errors from about ten a day to one or two, the hours corroborated by the Victorian government agency LaunchVic; theirs, not ours. Real proof that the pattern works, honestly attributed to the people who earned it.

The second thing: the tool alone doesn’t pay you back. A traceability assistant that no one has wired into a real receiving and dispatch process is just another dashboard. The payback comes from the tool plus the workflow around it, which is why we start small and on your own data. A free 30-minute discovery call walks one real path on your systems and names the highest-value fix; any build gets quoted fixed-scope after that, on your numbers, before you commit a cent. The build itself typically ships in three to six weeks in the A$10k to A$60k band. You own the code and the prompts; you can keep running it without us. You’d be an early food client, and this page says so plainly.

Where to start

You don’t need a transformation programme to find out whether this is real for you. You need one honest look at where the thread from ingredient to pallet actually lives in your plant, and which of your systems an agent can reach to keep it.

Two ways in, both low-risk:

• Read the map first. We’ve written a plain-English brief on the food and beverage software landscape, which of your own systems an AI agent can actually reach, by what mechanism, and what “closed” really costs you when the clock is running. → https://realmindsai.com.au/guides/food-bev/
• Book a free 30-minute discovery call. We’ll walk one real path, an order in or a lot traced, name the systems underneath it, and show you the highest-value place to close the gap, before anyone promises a build. → https://outlook.office.com/book/[email protected]/?ismsaljsauthenabled

There’s a batch moving through your plant right now whose paper trail you’d have to reconstruct by hand if the phone rang tonight. The only question worth asking is whether, at 2am, you’d be reading a picture that takes minutes, or standing at a whiteboard with a torch.