How to Choose an AI Agency in Australia: What to Ask First

Choosing an AI agency in Australia comes down to seven questions, and the answers tell you more than any portfolio: what exactly will be built and in which of your systems, what proof exists that it worked on a comparable job, where your data goes and who decided, how the price is structured and what the quote must contain, who owns the agent after launch, what happens when it breaks, and whether the agency will tell you when an agent is the wrong tool. An agency that answers all seven in plain terms, with numbers, is a safe hire. One that answers with a demo is not there yet.

I run an AI agency, so read this as the questions I would want a client to put to me, and the answers I think they should expect. The Australian Bureau of Statistics found that just over one in ten businesses, 12 per cent, used AI in 2024–25, up from 1 per cent two years earlier (ABS). That is a lot of first purchases happening at once, most of them by people who have never bought this kind of work before, and the market selling to them is young. Most of what is written about choosing an agency is written by agencies, this post included. So I have kept it to things you can check.

What does an AI agency actually sell, and what are you buying?

An AI agency sells a working process, not a model: it scopes one job in your business, builds the software that reads the input, decides and acts inside the systems you already run, tests it against real cases, and keeps it running after launch. The model underneath, whether it comes from Anthropic, Google Cloud or a server you host yourself, is a component the agency chooses and can swap. What you are paying for is the scoping, the integration and the maintenance around it.

Amazon Web Services puts the definition this way: “An artificial intelligence (AI) agent is a software program that can interact with its environment, collect data, and use that data to perform self-directed tasks that meet predetermined goals.” (AWS). In a small Australian business the environment is Xero, MYOB or HubSpot, an inbox and whatever runs the jobs. The plug between the agent and those systems that I now see on most new builds is the Model Context Protocol, whose own documentation says to “Think of MCP like a USB-C port for AI applications.” (Model Context Protocol). An agency that can tell you which connectors it will use, and which it will have to write, has already done the scoping work.

The word “agency” covers very different businesses. Some are marketing agencies adding AI tools to campaigns. Some are automation shops building rule-based workflows in n8n, Zapier or Microsoft Copilot Studio. Some implement a vendor platform such as Salesforce Agentforce, and some write custom agents against a model API. None of these is the wrong kind; they are different purchases. Ask which one you are talking to, then ask what they would build if your job turned out to be a fixed rule, because the honest answer is often a workflow rather than an agent, and the three ways to automate a business process sets out when each one wins.

What proof should an AI agency show you before you sign?

The proof worth asking for is a measured result on a comparable job: the hours the task took before, the hours and error rate after, how long it has been running, and a client who will take your call. A demo proves the agency can make something run on a screen. A baseline and an after-figure prove it ran in someone’s business for long enough to measure, which is the only thing you are buying.

The industry’s own numbers are why I would insist on it. MIT’s NANDA initiative studied enterprise generative AI deployments in 2025 and, as Fortune reported, found that 95% of generative AI pilots at companies are failing (Fortune). The same report found that purchasing AI tools from specialised vendors and building partnerships succeeded about 67 per cent of the time, while internal builds succeeded only one-third as often. That is the case for hiring an agency at all. It is also a warning: a two-in-three success rate on the better path means one paid engagement in three still did not land, and the proof question is how you avoid being that one.

Capability is moving quickly, which is exactly why a dated result matters more than a general claim. Stanford’s 2026 AI Index reports that organisational adoption of AI reached 88 per cent and that agents jumped from 12 per cent to roughly 66 per cent task success on the OSWorld benchmark (Stanford HAI). In TheAgentCompany benchmark, which simulates a small software firm with real tasks across its tools, the most competitive agent could complete 30 per cent of tasks autonomously (arXiv). Both are true at once. Agents are far better than they were and still finish only part of a realistic job on their own, so the proof you want includes how the agency handles the part the agent did not finish.

So, what I would ask for: one named job, its baseline, its after-figure, how long it has run, and what broke in the first month. An agency that has done the work answers in a minute. If you get logos instead of numbers, ask again.

What should you ask about your data, the Privacy Act and where the model runs?

Ask three things: what information the agent will read, where that information is processed and stored, and who on the agency’s side and the vendor’s side can see it. The Privacy Act covers organisations with an annual turnover above $3 million and private-sector health service providers at any size (OAIC), so a two-doctor clinic is inside it while a trades business under the threshold may not be. Either way, the agency should answer in writing before you sign.

Australia’s privacy regulator has already said what that due diligence looks like. The OAIC’s guidance says that when looking to adopt a commercially available product, organisations should conduct due diligence to ensure the product is suitable to its intended uses, including whether it has been tested for those uses, how human oversight is embedded, the privacy and security risks, and who will have access to the personal information it takes in or generates (OAIC). A good agency has those answers ready because it has been asked before. I would add one more: how are the agent’s credentials scoped? An agent with a standing login to your accounting system and your inbox has more access than most of your staff, and the agency should be able to show you exactly what it can and cannot touch.

Two numbers frame the risk. The ABS found that just over one in five Australian businesses, 21 per cent, experienced a cyber security incident in 2024–25 (ABS), and Roy Morgan found 13.6 million Australians, 58 per cent of people aged 14 and over, already using AI tools in an average four weeks in the March quarter of 2026 (Roy Morgan), an indicator we track in our index of Australian small-business AI data. Your staff are already pasting things into something. The question for the agency is whether the sanctioned version will be safer than the habit it replaces, and the options, from a public chatbot to a model that never leaves the building, are set out in the four levels of AI data security.

How do AI agencies charge, and what should a quote contain?

AI agencies charge in four main ways: a fixed price for a scoped build, a monthly retainer that covers building and running, a platform subscription with a setup fee, or time and materials by the hour. Each one puts the risk in a different place, and the structure tells you how the agency expects the work to go. Our own AI and automation work is quoted as a per-project flat fee from $1,990, and projects over $5,000 are billed 50 per cent upfront, 30 per cent at the midpoint and 20 per cent on completion (Compound Co pricing).

Pricing model	What you pay for	Where the risk sits	What to check
Fixed-price build	A defined scope, delivered	With the agency, if the scope is written tightly	The scope names the task, the systems and the acceptance test
Monthly retainer	Build plus running, monitoring and changes	Shared; you pay whether or not the agent is busy	What is included, the notice period, what happens to the agent if you stop
Platform subscription plus setup	A seat on a vendor’s platform, configured for you	With you, as the per-seat fee continues	Who owns the configuration, the exit path, whether there is an Australian data region
Time and materials	Hours, as they are used	With you	A cap, weekly reporting, a written definition of done

Whatever the model, I think a quote you can hold someone to contains six things: the job as a single sentence with an input and a weekly volume, the baseline it will be measured against, the acceptance test that decides when it is done, the running costs per month after launch including model usage, who owns what, and what support costs in year two. A quote that is one line and a number has left all six for later, and later is when they get expensive. We have covered what a good first agent engagement looks like and what it should cost separately; the short version is that the build should be sized against the hours it removes.

Who owns the agent after launch: the code, the prompts and the accounts?

You should own the agent you paid for: the code, the prompts, the workflow definitions, the documentation, and every account and API key it runs under, all in your name from the first day. An agency that holds the keys holds the agent, and the fee for leaving is then whatever it decides. This is the clause I would read before the price.

Ownership is practical as well as legal. The systems an agent connects to each publish a developer platform, and Xero, MYOB and HubSpot all have one (Xero, MYOB, HubSpot). A connection made through them is tied to an account, and if that account is the agency’s, your agent stops when the relationship does. The same goes for the model provider account, the n8n or Copilot Studio workspace, and the hosting. Ask for a handover document that another developer could pick up cold, and ask whether the agency has ever actually handed one over.

I would not expect a ten-person business to read the NIST AI Risk Management Framework (NIST), but the idea underneath it scales down to a folder: the code, the prompts, the list of credentials and what each one can reach, and a page that explains how to switch the agent off. If the agency cannot produce that folder for a past client, it has not finished a project in the way that matters to you.

When should an AI agency tell you not to buy an agent?

An AI agency should tell you not to buy an agent when the job is a fixed rule, when the data is not reachable, or when nobody on your side will own the output, and its willingness to say so is the best single test of the agency. An agent costs more per task and takes longer per task than a rule. It earns its place only where a person currently reads something messy and makes a judgement.

Anthropic says the same to the people who build on its own models:

“When building applications with LLMs, we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.”

Anthropic, Building effective agents (Anthropic)

I couldn’t agree more, and I would put the same question to any agency: what would you build if this turned out not to need AI? If the answer is still an agent, be careful. If the answer is a workflow in n8n, Zapier or Copilot Studio at a fraction of the price, you have found an agency that quotes the job rather than the technology. If a business is not ready, a good agency names the gap instead of quoting around it; the five readiness checks are what I look for first, and the one most often missing is a measured baseline. If we are not the right fit for a job we say so, and I would rather lose a small project than build the wrong thing and have it switched off in six months.

How do you shortlist an AI agency in two weeks?

A two-week shortlist is enough: write the job down in week one, put the same seven questions to three agencies in week two, and compare the written answers side by side rather than the presentations. The written form matters because it is what you can hold them to, and because a numbered answer and a general one look very different on the same page.

Question	A strong answer sounds like	A weak answer sounds like
What exactly will you build, and in which systems?	”An agent that reads supplier invoices from your inbox and matches them to purchase orders in Xero, about 60 a week"	"An AI solution for your finance function”
What proof do you have on a comparable job?	A baseline, an after-figure, how long it has run, and one reference	A list of logos
Where does our data go, and who can see it?	A written answer to the OAIC’s due-diligence questions, with the hosting region named	”It’s all secure”
How do you charge, and what is in the quote?	The job sentence, the baseline, the acceptance test, running costs, ownership and year-two support, itemised	A single figure
Who owns the agent afterwards?	”You do: code, prompts, accounts and keys in your name, with a handover document"	"We host it for you”
What happens when it breaks?	A named person, a response time, a switch-off procedure and what each costs	”It won’t”
Would you ever tell us not to build this?	”Yes, and here is the rule-based version if that is what the job is”	A yes with no example

Then choose the agency whose answers you could hand to a colleague who was not in the room. That is how our own AI engagements begin, with a scoped task and a baseline, and it is often a smaller project than the client expected. If you would like to put the seven questions to us, start a conversation and we will answer them in writing, including the one about whether you need an agent at all.

Frequently asked questions

How do I choose an AI agency in Australia?

Put seven questions to each agency in writing: what exactly will be built and in which systems, what measured proof exists on a comparable job, where your data goes and who can see it, how they charge and what the quote contains, who owns the code, prompts and accounts afterwards, what happens when it breaks, and whether they would ever advise against building an agent. Compare the written answers side by side and choose the agency that answers with numbers.

What should an AI agency’s quote include?

A quote you can hold an agency to contains six things: the job as a single sentence with an input and a weekly volume, the baseline it will be measured against, the acceptance test that decides when it is done, the monthly running costs after launch including model usage, who owns what, and what support costs in the second year. A single figure with none of these leaves every expensive question for later.

Who should own the AI agent after the project?

The client should own the code, the prompts, the workflow definitions, the documentation and every account and API key the agent runs under, all in the client’s name from the first day. If the connections to Xero, MYOB, HubSpot or the model provider sit in the agency’s account, the agent stops when the relationship does. Ask for a handover document another developer could pick up cold.

Does the Privacy Act apply if I hire an AI agency to build an agent?

The Privacy Act covers organisations with an annual turnover above $3 million and private-sector health service providers at any size, so a small clinic is inside it. Whatever your size, the OAIC’s guidance says to conduct due diligence on an AI product: whether it has been tested for your use, how human oversight is embedded, the privacy and security risks, and who will have access to the personal information it takes in or generates. The agency should answer those questions in writing before you sign.