AI agent or automation: choosing the right tool.

Ask a vendor which of your processes an AI agent could run and the answer is usually all of them. Ask the people who will maintain it and the list gets shorter. Many jobs being scoped as AI agents in Australian organisations would be done better, more cheaply and more safely by ordinary automation: a workflow, an integration or a scheduled script that follows fixed rules.

This is not an argument against agents. We build them, and where they fit they take real work off a team. It is an argument for choosing the tool by the job. A good AI consulting engagement should tell you where an agent is the wrong answer as readily as where it is the right one.

Three tools that all get called AI.

Gartner has a name for the confusion: agent washing, the rebranding of assistants, chatbots and robotic process automation as agents. In June 2025 it estimated that only around 130 of the thousands of vendors claiming agentic capability were the real thing, and predicted that over 40% of agentic AI projects would be cancelled by the end of 2027. Telling the three tools below apart is the first step to not being one of those projects.

• Rules-based automation. A workflow tool, an integration or a script that does the same steps in the same order every time. It moves a record from one system to another, checks a value against a table, sends a reminder on a date. It does not read or judge. When it fails, it fails the same way every time, which makes it easy to test and audit.
• An AI step inside a workflow. A model does one bounded job in an otherwise fixed process: reading a scanned invoice into fields, sorting an email into a category, drafting a summary for a person to check. The workflow around it decides what happens next.
• An AI agent. A system that is given a goal, decides which steps to take, uses tools such as searching records or updating a system, and works through cases that do not follow a script. It handles variation that rules cannot. It also costs more to run, is harder to test, and needs a person to answer for what it does.

Much of the practical value sits in the middle tool, which gets the least attention.

Four questions that decide it.

Put each step of the process through these four questions. Ask them of each step, not of the process as a whole, because most processes contain all three kinds of work.

• Is the input structured? If the step reads fields from a system, it is a rules job. If it reads an email, a PDF whose layout changes by sender, a call transcript or a free-text note, it needs a model.
• Is there one right answer? Checking a line against a published price is arithmetic. Deciding what to say to a customer whose complaint touches three departments is not. Where a rule can check the answer, use the rule.
• Does the path change case by case? If the next step depends on what the last step found, and the possibilities cannot be listed in advance, that is the work an agent is for. If they can be listed, list them and automate them.
• What does a wrong answer cost, and who catches it? A rule gets the same case wrong every time, so testing finds it. A model can get one case in fifty wrong, in a different way each time. The more a wrong answer costs, the more checking and human review the step needs, and that cost belongs in the business case.

A step that is structured, has one answer and follows a fixed path is automation. A step that reads something unstructured but has a bounded job is an AI step. Only a step that is unstructured, open-ended and variable in path needs an agent.

Why the plainer tool often wins.

When a step could be done either way, the plainer tool usually wins on four counts.

• Running cost. A built workflow costs close to nothing per transaction. An agent pays for model usage on every case, and for the monitoring and review around it. At high volume the difference compounds.
• Predictability. A rule does what it says. An agent that is right 97 times in 100 still needs someone to find the other three.
• Audit. When a regulator, an auditor or a customer asks why something happened, a rule can be shown. An agent's reasoning has to be logged, kept and explained, and the 10 December privacy rule turns that into a disclosure obligation for decisions that could significantly affect a person.
• Upkeep. When a price, a policy or a field changes, a rule is updated in one place. An agent's prompts, tools and test cases all need checking.

None of this rules agents out. It means an agent has to earn those costs back through work a rule cannot do.

Most real processes are a mix.

Take supplier invoices, a process almost every organisation runs and few enjoy.

• An invoice arrives by email. Automation. A workflow picks it up and files it.
• The invoice is a PDF whose layout depends on the supplier. An AI step. A model reads it into fields, and anything it is unsure of goes to a person.
• The fields are matched to a purchase order and a receipt. Automation. A three-way match is a rule.
• The match fails. The reason could be a price change, a partial delivery, a wrong order number or a duplicate. This is the variable step, and the one for an agent: it can look up the order history, check the contract price, draft a query to the supplier and propose a resolution.
• Someone approves payment. A person. The decision and the accountability stay with a named approver.

One process, four kinds of work. The agent does one step of five, and it is the step where it saves the most time, because exceptions are where the hours go. A proposal that puts an agent across the whole process is paying agent prices for rule work.

The same pattern holds in care. Comparing a claim line with the current NDIS price limits is a rule. Working out why a line was rejected, and what to do about it, is where an agent helps. Our guide to AI consulting for NDIS and home care providers applies this to the agents a provider usually builds first.

The data decides more than the tool.

Whichever tool a step needs, it reads the same data. A rule that matches an invoice to a purchase order fails if the supplier is recorded three different ways. An agent fails on the same records, more expensively and less visibly. In most AI projects the deciding work is joining and governing the data the tools read, which is why we build the data foundation first. Our piece on where AI agents do not work yet covers the cases no tool will fix until the data and the process are fixed.

What to ask an AI consulting firm.

These four questions show quickly whether a firm is choosing the tool by the job or by what it sells.

• Which steps in this process would you not use AI for? A firm that cannot name any has not looked closely.
• What does each step cost per transaction at our peak volume? Ask for model usage, platform and review time, step by step.
• How will we test it? Ask for the test cases, the error rate the firm expects at each step, and who checks the cases a model gets wrong.
• What will we own and maintain? Rules, prompts, integrations and tests. Our piece on agent platform or AI consulting firm lists the costs a licence leaves out.

Why RUBIX.

RUBIX is an independent, Australian-owned data and AI consultancy, founded in Melbourne in 2011. For 15 years our work has been what automation and AI both stand on: integrations, pipelines, data platforms and governance for Australian organisations, and more recently the AI agents that run on top of them. We have delivered more than 450 projects for over 115 organisations.

• Process and data first. We map the steps and fix the data before choosing a tool. When a step is a rule, we build the rule.
• Agents where they earn it. When a step needs an agent, our agentic AI and automation team builds one with a named approver, a test set and a log.
• Fixed scope. Our engagements are fixed-scope, so you know what each step will cost before we begin. For a leading NDIS provider we joined six disconnected systems into one governed data foundation, with a live role-based dashboard, inside eight weeks. Read the NDIS provider case study.

Our about page sets out how RUBIX has grown since 2011.

Frequently asked questions.

Sources.

• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025.

General information only, not legal, financial or regulatory advice. Current as at October 2026.