For a decade, automation in most Australian businesses meant robotic process automation. You mapped a repetitive task, recorded the clicks, and a software bot repeated them. It worked, until a vendor moved a button or a supplier changed an invoice layout. In 2026 more buyers are asking a different question: instead of a bot that follows a script, what about an agent that understands the job? That shift, from RPA to agentic automation built on Claude, is changing what gets automated and how much it costs to keep running.
What RPA actually does, and where it breaks
RPA is rule-based. A developer records an exact sequence: open this system, copy that field, paste it there, click submit. The bot does precisely that, thousands of times, without tiring. For stable, high-volume, identical tasks it is fast and cheap to run once built.
The trouble is everything that is not identical. RPA bots are brittle. They read screen positions and fixed formats, so a redesigned portal, a new PDF template, or a supplier who writes Net 30 instead of 30 days can stop the whole process. Someone then has to notice the break, log it, and pay a developer to re-record the flow. Australian finance and operations teams know this tax well: the bot that saved twenty hours a month also generates a steady stream of maintenance tickets.
The three costs of RPA that rarely make the business case:
Build: recording, testing, and exception handling for each process, often $15,000 to $40,000 per workflow with a vendor.
Licences: per-bot annual fees that keep charging whether the bot runs once a day or once a minute.
Maintenance: developer time to fix flows every time an upstream system changes, which for a busy process can approach the original build cost each year.
What agentic automation changes
Agentic automation replaces the recorded script with a model that reads the task in context. Claude does not memorise button positions. It reads an email, a PDF, or a system record the way a person would, works out what needs to happen, and takes the next step. When the invoice layout changes, it still finds the total, because it is reading meaning, not pixel coordinates.
This matters most for work that has always resisted RPA: unstructured inputs, judgement calls, and exceptions. A bot can copy a purchase order into your accounting system only if every order looks the same. An agent can handle the order that arrives as a forwarded email with the quantities in the body and the price in an attachment, flag the one line that does not match the quote, and draft a reply to the supplier. That is the awkward 20 percent of cases that used to land back on a human desk.
The cost picture in AUD
The headline numbers can look similar, but the shape is different. A mid-sized Sydney firm might spend $30,000 standing up a handful of RPA bots, then $50,000 a year on licences and fixes. An equivalent set of Claude-based agents usually has a lower build cost, because there is no brittle click-path to engineer, and running costs that scale with volume rather than a fixed per-bot licence.
Where the money actually goes, compared:
RPA leans on upfront engineering and ongoing maintenance, so total cost stays roughly flat even when volume drops.
Agentic automation leans on clear instructions and good guardrails, so a well-scoped agent can run for a few hundred dollars a month and grows only with real usage.
The break-even favours agents wherever inputs vary or systems change often, which describes most Australian SMB back-office work.
Where RPA still wins
This is not a story where the robots lose everywhere. For genuinely high-volume, deterministic, stable tasks, such as moving thousands of identical records between two systems that never change, RPA remains the cheaper tool per transaction. If your process is truly identical every time and the underlying screens are frozen, a rule-based bot is hard to beat on unit cost. Plenty of businesses will run both: RPA for the stable core, agents for the messy edges.
How Australian businesses are deciding in 2026
The practical test we use with clients is simple. Ask how often the task involves reading something a human wrote, making a judgement, or handling an exception. The more of those, the stronger the case for an agent. Then ask how stable the systems are. The more they change, the more an RPA bot will cost you in quiet maintenance.
A quick way to sort your automation backlog:
Send to an agent: anything involving unstructured text, variable formats, or it-depends decisions, like triaging inbox requests, reviewing documents, or drafting tailored replies.
Keep on RPA: fixed, high-volume data movement between systems that will not change, where speed per transaction is the whole point.
Review the maintenance bill: if an existing bot breaks more than a couple of times a quarter, its true cost is higher than the licence line suggests, and an agent may be cheaper overall.
One consideration matters in Australia specifically. Whichever path you choose, the data handling has to hold up. Under the Privacy Act, and for regulated firms under ASIC and APRA expectations, you need to know where information goes and who can see it. Claude-based agents can be scoped to read only what they need and to keep a record of what they did, which is often easier to explain to an auditor than a bot quietly clicking through a system on a shared login.
The move from RPA to agentic automation is less about swapping one tool for another and more about matching the tool to the work. Stable and identical, keep the robot. Variable, judgement-heavy, or forever changing, an agent will usually cost less and break less. If you want a clear-eyed look at which of your processes fit which model, book a brainstorm with us and we will map your automation backlog against both.
