Agent platform or AI consulting firm.

Not long ago, an organisation that wanted an AI agent had to build one. Now much of the business software an Australian organisation already pays for comes with an agent builder switched on, or one licence away. Microsoft Copilot Studio, Salesforce Agentforce and ServiceNow's AI agents are the best-known examples, and the vendors behind finance, rostering and care management software are heading the same way.

So the question has changed. It is no longer whether to use AI agents. It is whether to switch on the one in the software you already own, or pay an AI consulting firm to build something. The honest answer is that it depends on where the work lives, and most organisations end up doing some of both. This piece sets out how to tell which, and what each option leaves out.

Three things you could be buying.

• The agent inside software you already run. The vendor supplies the model, the connections to its own data and a builder your team configures. Fast to start, and the cheapest path when it fits.
• A standalone agent product. A specialist tool for one job, such as call handling, document intake or scheduling, sold as a subscription. Good when the job is common and your version of it is not unusual.
• A build on your own data. An AI consulting firm designs the workflow, joins the systems, builds the agent on a platform you choose and tests it against your rules. Slower to start, and the right call when the work does not fit inside one product.

These are layers more than rivals. A sensible programme often uses all three: the built-in agent for work inside one system, a product for a commodity job, and a build for the workflow that is specific to you. The logic we apply in build or buy your data platform holds here too: buy the commodity, build the part that is yours.

When the platform is enough.

Switching on the built-in agent is usually the right first move when four things are true.

• The work starts and finishes in one system. Summarising a service ticket, drafting a reply from a customer record, sorting a shared inbox.
• The data is already there, and already right. An agent is only as good as the records it reads. If they live in the platform and your team trusts them, the hard part is done.
• The vendor's rules are close enough to yours. Generic approval steps and permissions fit generic work. They fit less well when the rule is an Australian regulation the vendor has never had to encode.
• Someone on your side owns it. A named person who reviews what the agent does, changes it when the work changes and switches it off when it is wrong.

If all four hold, you do not need to hire anyone. Configure it, measure it for a month and decide from the evidence. We would rather tell a prospective client that than sell them a build they do not need.

When you need someone to build it.

The platform stops being enough at predictable points.

• The workflow crosses systems. A care provider checking a claim needs the roster, the service agreement, the participant's budget and the current price limits, which usually sit in four places. An agent inside any one of them sees a quarter of the picture. Our pieces on AI for NDIS providers and AI for home care providers show how common this is in care.
• The rules are Australian and specific. NDIS price limits, APRA's CPS 230 for financial services, and the Privacy Act change that, from 10 December 2026, requires an organisation to say in its privacy policy when a computer program makes decisions that could significantly affect a person. A global product will not encode those for you. Our piece on AI and the NDIA works through one example in detail.
• The data has to stay under your control. Prompts, logs and backups can leave Australia even when the main database does not. Our piece on where your AI data can live covers where it usually leaks.
• You need to know when it is wrong. A platform will tell you the agent ran. It will not tell you the answer was right. Testing an agent against your own cases, and keeping that test running as the rules change, is work somebody has to do.
• The data is not ready. This is where most agent projects stall. If your records disagree with each other, no licence fixes that. An AI readiness assessment tells you before you spend.

Some jobs are not ready for an agent at all, whoever builds it, and then the right answer is neither option. We have written about where AI agents do not work yet.

The costs a platform quote leaves out.

An agent licence looks cheap next to a consulting quote, and on the licence line it usually is. The comparison is only fair once you add the work the licence does not include, because someone does that work either way.

• Usage on top of the seat. Agents are increasingly priced per message, per conversation or per action as well as per user. Model a busy month, not an average one.
• Integration with everything else. Connections to the vendor's own products are included. Connections to your rostering system, your finance system or a government portal usually are not.
• Getting the data right. Cleaning, joining and governing the records the agent reads. In a real project this is often the largest line, and it sits outside every licence.
• Testing and upkeep. Checking outputs against your own cases, and updating the agent when a rule, a price or a policy changes.
• The owner's time. Someone on your side reviews, approves and answers for what the agent does.

Our guide to what AI consulting costs in Australia covers how to read the other side of the comparison.

What a good AI consulting firm should leave you with.

If you do bring someone in, the test of a good firm is how little you depend on it afterwards. Whatever it builds, you should own the configuration, the integration code, the tests that show the agent is right and the runbook that says what to do when it is not. It should build on the platforms you already pay for where they fit, and be able to explain in plain terms why it built anything the platform could have done. What AI consulting actually delivers sets out what an engagement should give you, whatever kind it is.

Five questions for the next vendor meeting.

Put these to the platform vendor and the consulting firm alike.

• Which of our systems will the agent read, and which will it write to? If the answer is one, a platform may be enough. If it is four, ask who builds the other three connections.
• Which Australian rules does it handle today? Ask for the specific rule and how it is checked, not a general statement about compliance.
• Where do prompts, logs and backups sit? "In Australia" should cover all three, not only the main database.
• How will we know when it is wrong? Ask to see the test cases and who keeps them up to date.
• What does a busy month cost? Ask for usage pricing at your peak volume, plus the integration and data work, as one number.

Why RUBIX.

RUBIX is an independent Australian data and AI consultancy, founded in 2011. That is 15 years, 450+ projects and 115+ customers, much of it for organisations where the data is sensitive and the rules are specific. Four things follow from that for this decision.

• We do not sell a platform of our own. We partner with Databricks, Microsoft, AWS, Snowflake, Qlik and Alation, so we can recommend switching on what you already own when it fits, and building only where it does not.
• Data first, then agents. Most of the cost and risk of an agent sits in the data it reads. We build the data foundation that joins your systems, then put AI agents to work on it.
• We have done the cross-system version in care. For a leading NDIS provider we joined six disconnected systems into one governed data foundation, a live role-based dashboard and an automation roadmap that quantified around $645K of annual impact, inside eight weeks. Read the NDIS provider case study.
• Senior people do the work. Our forward deployed engineers sit with your team, so the people who scope the work are the people who deliver it. Our AI consulting in Australia is built around that.

If you are weighing a licence against a build, start with a conversation about where the work actually lives. Sometimes the answer is the licence.

Frequently asked questions.

General information only, not legal, financial or regulatory advice. Current as at September 2026. Product names are used for illustration; check each vendor's current features and pricing.