AI agents at work: what Australian businesses have actually deployed

Photo: rawpixel via rawpixel (CC0 1.0)
I’ve stopped counting pilots. Every organisation I’ve talked to in the last eighteen months has one, most have several, and the number tells you nothing. What I’ve been counting instead is the smaller, more interesting set: AI agents that are in production, handling real work, with a person’s name against the outcome. That list is short. It’s also more revealing than any survey, because the things on it share a shape.
A quick definition, since the word has been stretched past usefulness. By “agent” I mean software that uses a language model to take a sequence of actions toward a goal, calling tools, reading and writing systems, deciding what to do next, rather than a chatbot that answers one question and stops. The distinction matters because the second kind is everywhere and the first kind is where the money and the risk are.
What’s actually in production #
Four categories, over and over, in the deployments I’ve been able to verify with the people running them.
Document processing is the biggest. Insurers reading claims, lenders reading payslips and bank statements, law firms reading discovery, councils reading development applications. The agent extracts, classifies, cross-checks against a system of record and routes. It’s unglamorous, it’s been possible in cruder forms for a decade, and the models made it good enough to trust with the first pass. The common feature: a human still signs off on the outcome, but reviews a summary rather than the source.
Customer service triage is the second. Not the chatbot on the website, which most customers have learnt to route around, but the agent behind the scenes that reads an inbound email or call transcript, pulls the customer’s history, drafts a response and either sends it under a threshold of confidence or hands it to a person with the draft attached. Two of the big four banks run versions of this. So do at least three energy retailers. The metric that matters to them is handling time, and the numbers they’ll share privately are real.
Software engineering is the third, and it’s the one where the agents have the most autonomy, because the output is checkable by a compiler and a test suite before anyone gets hurt. Coding agents that take a ticket, write the change, run the tests and open a pull request are in daily use at every technology company I know of in Sydney and Melbourne, and at a surprising number of non-technology companies with in-house teams. Atlassian, Canva and the like talk about this publicly. The mid-tier firms that do it just as much don’t.
The fourth is internal knowledge work: the agent that answers “what’s our policy on X” by reading the actual policy, or drafts the first version of a board paper from last quarter’s numbers. This one is widespread and also where I see the most quiet disappointment, for reasons I’ll come to.

What the successful ones have in common #
Three things, and I’d put money on all of them.
A narrow scope. Every production agent I’ve seen does one job with a defined start and end. The “assistant that can do anything” is the demo; the “thing that reads incoming remittance advices and matches them to invoices” is the deployment. Founders hate this because narrow scope makes a smaller pitch deck. Operators love it because it makes a smaller blast radius.
A checkable output. The agents that survive are the ones whose work can be verified cheaply, either by a test suite, a reconciliation, a threshold or a human glancing at a summary. Where verification is expensive or subjective (is this board paper any good?) the agent tends to end up as a drafting tool that a person rewrites, which is useful but isn’t the productivity story that was sold.
A system of record to anchor to. The document agents work because there’s an ERP, a claims system, a case management platform to check against. The knowledge agents disappoint because the knowledge they’re reading is a SharePoint of contradicting documents nobody has maintained since 2019. The model isn’t the problem there. The model is being asked to be confident about an organisation that isn’t.
What the failures had in common #
I’ve seen more of these than successes, and honestly they’re more instructive.
The most common is the agent that worked in the pilot on a curated set of inputs and fell over on the real, messy, adversarial ones. Claims forms filled in by hand in pencil. Emails that are actually three emails forwarded together. A supplier who formats invoices differently every month. The pilot showed 94 per cent accuracy. Production showed 71, and the 29 per cent it got wrong were the expensive cases, because the expensive cases are the weird ones.
The second is the cost surprise. Per-token pricing looks trivial in a demo and becomes a line item at scale, especially for agents that loop, retry and re-read. I know of one retailer whose customer service agent was, for a fortnight, the single largest item on its cloud bill after storage. Nobody had modelled the loop count.
The third is governance. Someone deployed the thing, it did something a regulator or a customer didn’t like, and nobody could say with confidence what it had been told, what it had read or why it decided what it decided. The Voluntary AI Safety Standard that the federal industry department published in 2024 is mostly a checklist of the things those organisations wished they’d done first: an accountable owner, a risk assessment, logging, a way for a person to intervene, a way for a customer to know. It’s voluntary, for now, and the proposed mandatory guardrails for high-risk uses that were consulted on at the same time are the direction of travel. The privacy regulator has been clear that feeding personal information into a model you don’t control is a disclosure, with everything that follows.
Where the public sector sits #
Government is the other half of this picture, and it’s further along than the private sector likes to assume. The Digital Transformation Agency’s policy for the responsible use of AI in government, which took effect in late 2024, requires every federal agency to have an accountable official and to publish a transparency statement about how it uses AI. That sounds like paperwork. In practice it forced a census, and the census found a lot of quiet document-processing and call-centre work already running, plus a great many “assistants” that a handful of staff had built themselves on whatever tool they could get a licence for.
The state agencies I’ve spoken to are in the same place, with one difference: they have the Robodebt Royal Commission’s findings sitting on the shelf, and the phrase “automated decision-making” now gets a lawyer in the room. That’s healthy. The Commonwealth’s own response committed to a legislative framework for automated decisions, and any agent that changes what a person is entitled to receive is going to live inside that framework. The department that reads development applications with an agent and lets a planner decide is fine. The one that lets the agent decide is going to have a very long conversation with its general counsel.
The contrast with the private sector is instructive. The banks and insurers have a supervisor, in APRA, that already expects a model risk framework and has said in writing that AI models fall under it. The retailers and the mid-market don’t, and that’s where I see the deployments with the least governance and, not coincidentally, the most exciting demos.
The measured opinion #
Here’s the bit I’ll get letters about. I reckon the “agentic” framing is doing the industry a disservice, because it sells autonomy and the things that work are mostly not autonomous. They’re supervised, narrow, checkable pipelines with a language model doing the parts that used to need a rules engine and a very patient analyst. That’s a real and valuable thing. It’s also closer to robotic process automation with better reading comprehension than to the digital employee in the keynote, and the organisations that understood that from the start are the ones with something in production.
What I’d do first #
If you’re the person who’s been asked to “do something with agents” and has a budget and a deadline, pick the process with the most paper, the clearest system of record and the cheapest way to check an answer. Instrument it so you can see every input and output. Run it in shadow mode against the humans for a month before it touches a customer. Model the token cost at ten times your pilot volume. Write down who owns it, in a document that would survive your departure. And resist, hard, the request to let it do the whole job on day one, because the job on day one is finding the 29 per cent.
I swam this morning, which I mention only because the water at Coogee was 17 degrees and my colleagues in Melbourne asked, and because it’s the one hour of the day when nobody sends me a deck with the word “agentic” in it. That hour is getting more valuable.
— Sophie Lindqvist, AI, Start-ups & Fintech Editor, Technology Spectator