How to choose an AI agency in Australia: the map-before-money test

Most AI agencies look alike online: the same promises, the same logos, the same polished demo. What separates them is how they start, what they hand over, and what happens when the system breaks at 4pm on a Friday. This guide gives you a process for choosing an AI agency in Australia, one step at a time, with a brief to send, the questions to ask and a scorecard to fill in.

Disclosure: I’m Jackson Dong, the author of this guide and the founder of AppliedAgents, an AI implementation agency. No firm paid to be mentioned. Where I refer to published prices, they come from each firm’s own website, checked 1 to 3 November 2026. For named firms, see my comparisons of AI implementation agencies in Australia and AI consulting firms in Australia.

What does an AI agency actually do?

An AI agency helps a business put AI to work, but the four kinds of agency hand over very different things: a plan, a package, a working system inside your own tools, or a product. The label on the website rarely tells you which one you are dealing with. The useful question is what you hold at the end of the first engagement. Answer that and the type of help picks itself.

You don’t need an agency at all when the fix is a setting in software you already pay for, a simple rule your own team can set up in your existing tools, or a task that happens too rarely to be worth automating. A good agency says so in the first meeting, and the process below helps make sure yours does.

How do you choose an AI agency? Eight steps

Here is how to choose an AI agency: work on the problem before you work the market. Define it, pick the type of help, shortlist with a strict test, then compare three firms on the same brief, the same questions and the same contract checks.

1. Pick one problem and put a number on it

Start with one workflow that costs you time every week, and measure it before you talk to anyone. Hours a week, days of delay, jobs lost, enquiries nobody answered. Pick something frequent and painful rather than impressive: enquiry handling, quote follow-up, document collection. The number does three jobs. It tells you whether the problem is worth paying to fix. It gives every agency the same target. And it becomes the baseline you judge the build against later, which is why it has to be taken before any work begins.

If all you know is that the team is stretched, write that down too: you need someone to find the bottleneck first, which changes the type of help you need.

2. Decide which type of help fits

Match the type of agency to the problem, using the table above. A plan across the whole business suits an AI consultant. A common, well-defined problem may suit a package. A fix inside your existing tools is work for an AI automation agency. Something your customers log into is custom software. Keep the shortlist to one type, or one firm prices a chatbot, another prices a full rebuild, and comparing them teaches you nothing.

3. Shortlist three using the map-before-money test

The map-before-money test asks whether an AI agency maps how your business actually runs, with the people who do the work, before any money leaves your account. It is my own test, and I apply it before prices or portfolios. An agency passes only if it does all five.

1. It looks at the work, not the org chart. Someone sits with the people who do the job and watches a real enquiry, quote or invoice move through. In my experience, the owner’s description of a process and the process itself often differ.
2. It names the bottleneck that costs the most, in your numbers. Hours a week, days of delay, jobs lost. The first proposal targets that, not the easiest step to demo.
3. It tells you where AI is the wrong answer. Some steps should stay as fixed rules, some need a setting in software you already pay for, and some need a person.
4. It writes the findings down, and they are yours. You can take the map to another agency, or do nothing with it, without penalty.
5. No money leaves your account before you have the findings and a fixed quote. Not for an audit, not for a roadmap, not as a deposit. If you have to pay to find out what is worth building, the agency is asking you to carry the risk of its own discovery.

This matters most for small and mid-sized businesses. A large company can absorb a wrong first project. A twelve-person business usually cannot, and its problems are more particular: the way a plumbing office triages after-hours calls has grown up around specific people and tools. If a firm sells packages or roadmaps, ask how it makes sure the first build hits the right problem.

You can usually run the test on a first call: ask how the agency starts with a new client, and listen for who it talks to and when the first invoice arrives. Three firms that pass is enough.

4. Send all three the same brief

Agencies are only comparable when they answer the same brief, so write one page and send it to all three. Describe the problem, not a solution, and ask for the same six things back.

BRIEF FOR AI AGENCIES

Business: [what you do, where, how many staff]
The task: [one task, start to finish, e.g. “website enquiries are copied by hand into our job system and assigned”]
Who does it today: [roles] for about [hours] a week
Tools involved: [CRM, inbox, job system, accounting, phone]
Where it goes wrong: [what gets lost, delayed or duplicated, and how often]
A good result looks like: [e.g. “complete requests assigned once, missing details flagged”]
Constraints: [data that must stay in Australia, systems that cannot change, budget range]

Please respond with: the first problem you would solve and what you would leave out; which steps stay as rules and where AI makes a judgement; how we would test it on our real examples; what we pay before, during and after; what we own if we part ways; and a time to show us a similar system handling a case that goes wrong.

Write “unanswered” beside anything a proposal skips, and don’t fill the gap with something said in a sales call.

5. Ask the same questions

Ask every agency the same eight questions and judge each answer on whether it is specific to your process, because a weak answer would fit any business. These are the questions to ask an AI agency before you sign anything.

Then ask the demo to handle an exception. A happy-path demo shows the easy case, so ask to see the same workflow with a missing field, an ambiguous request and a destination that is temporarily down. For an enquiry handover, check that it keeps the customer’s own wording, tells a requested appointment from a confirmed booking, and leaves a clear next action for a person.

6. Read the proposal and the contract

A good proposal is one you can hold the finished build against, so read it for seven things before you look at the price.

• Fixed scope. The first problem it solves and, just as clearly, what is left out.
• Acceptance checks. How the build will be tested on your real examples, written down before the work starts, and who signs off.
• Ownership. Who owns the workflows, prompts, code and data, and which accounts the system runs in.
• Minimum term. Any lock-in, notice period or exit fee, stated in the contract rather than explained on a call.
• Exit. What you keep, which access ends, and what help you get to move the system or hand it to someone else.
• Support terms. What counts as a defect under warranty and what is a paid change, the response time, and whether an API update or a model retirement is covered.
• Ongoing costs. Licences, AI usage at your volume and support, each listed separately rather than folded into one monthly figure.

Check that the proposal and the agency’s standard terms say the same thing. Where they differ, the standard terms usually win, so get the difference fixed in writing first.

7. Check what happens after launch

Judge an agency on its failure plan, because any automated system will eventually meet a case it cannot handle. How will you find out: an alert, or a customer complaint? What does the system do while it waits: stop safely and flag a person, or keep going? Who fixes it, within what time, and is that in the support agreement or a favour? Then ask about reporting. You should be able to see what the system did this month, in a view you can open or a report you receive, and have one named person to call when something needs to change.

Data access belongs in the same conversation. Before you share live customer records, know which fields leave your systems, where each AI step runs and whether any of it trains a model. Depending on the provider, product, region and settings, the prompts sent to an AI model may be processed outside Australia. That matters for businesses covered by the Privacy Act: under Australian Privacy Principle 8, an organisation that discloses personal information to an overseas recipient is generally accountable for how that recipient handles it, as the OAIC’s APP 8 guidelines explain. Ask the agency to name the provider and region for each AI step, use de-identified examples in mapping and demos, and confirm in writing that its access to your systems ends with the engagement. Whether the Privacy Act covers your business needs its own check.

8. Start with one workflow and judge it against the acceptance checks

Start with the one workflow from step 1, agree in writing what it must prove, and expand only once it has proved it. One workflow is small enough to map properly, build in weeks and measure against the baseline you took before the work began. When the agency says it is done, run the acceptance checks on your own real examples and compare the result with the number from step 1. A build that passes earns the next workflow. To agree expected behaviour and record the result, use the editable implementation acceptance checklist.

What should you look for in an AI agency?

Look for an AI agency that behaves like a careful tradesperson: it looks before it quotes, writes things down and plans for the day something breaks. When I think about what to look for in an AI agency, these are the green flags I weight most:

• It asks to watch the work before it talks about tools. Mapping comes first, and the people who do the job are in the room.
• It tells you where AI is the wrong answer, and points to a rule or a setting you already pay for instead.
• It can show a live system handling a case that goes wrong, not just a demo of the easy path.
• It runs AI automation in its own business and can tell you what it cost and what it gets wrong.
• Ownership, exit and support are in the contract, and the proposal matches the standard terms.
• Costs come apart: the first build, licences, AI usage and support are listed separately.

Red flags when hiring an AI agency

The biggest warning sign when hiring an AI agency is confidence about your business from someone who has not looked at it. Others:

• A fixed price for your job before anyone has seen your process. A published starting price is fine; a quote for your specific job before any mapping is a guess.
• “AI for everything.” No step left as a simple rule, and no mention of a feature you already pay for.
• No failure plan. If nobody can say what happens when the model is unsure or a connection is down, assume a customer finds out first.
• No ownership terms in writing. A verbal “you own it” is not enough.
• Lock-in. Systems that only run in the agency’s accounts, long minimum terms or exit fees.
• Results with no baseline. “Saved 80% of admin time” means little without the before figure and who measured it.
• Vagueness about data. If “secure” is the whole answer to where your customers’ details go, keep asking.

Who should own what the agency builds?

You should own what the contract says you own, so make sure it says you own the system, or at least hold everything needed to run it without the agency: the accounts it runs in, the workflows, prompts and code assigned to you, exportable data, and documentation written for your team. Don’t assume ownership by default. IP Australia’s general guidance is that “IP created by a contractor is the property of the contractor unless otherwise stated in the contract”. Read the clause, check that the proposal and standard terms agree, ask about licensed base code that will not transfer, and get legal advice if the terms are unclear.

What does it cost to hire an AI agency?

It depends on the shape of the engagement, and there are five common shapes in Australia. Knowing them makes quotes comparable.

• Paid audit or roadmap, then a build. Published audit and roadmap prices run from A$1,999 to $7,500 ex GST, and some firms credit the fee to a build that follows. A paid audit fails check 5 of my test by design, because you pay before anyone knows what is worth building; that is a judgement about who carries the risk, not about quality, and a day of senior time can still be good value if you know what you get.
• Free mapping, then a fixed-price build. No cost until the quote; more of your time up front.
• Package prices. A set product at a set price, sometimes with a monthly fee.
• Fixed-cost software projects. Larger budgets and longer timelines, priced per project.
• Monthly care or retainer. A recurring fee to keep a build running and improve it.

Whatever the shape, ask for a proposal that separates the first piece of work from licences, AI usage, support and your team’s time, then compare the total cost of an agreed result. Named firms and their published prices are on the comparison of AI implementation agencies, and the AI implementation cost guide breaks down each part of a quote.

Where AppliedAgents fits

AppliedAgents, the agency I founded, is built around the map-before-money test, so weigh this paragraph accordingly. Its first step is a free business audit: a strategy call, a mapping session with the people who do the work (in person where possible), written findings, and a fixed quote for the first build. No money changes hands before you have seen the findings and the quote, and sometimes the finding is that AI is not the answer. Clients own the code, prompts, workflows and data built for them once the project is paid for (see the AI implementation service and how AppliedAgents handles data and AI). The honest trade-off: there is no published price list, because each system is tailored and costs vary with scope, and the mapping takes more of the owner’s time up front.

How do you compare the three proposals? Use a scorecard

Fill in one scorecard per agency from the proposal and the meeting, and treat anything unresolved as a follow-up question rather than inventing a score.

AI AGENCY SCORECARD

Agency: [name] Type: [consultant / productised / automation / custom dev]
Map before money: looked at the real work [y/n] named the costly bottleneck [y/n] said where AI is the wrong answer [y/n] written findings we keep [y/n] fee before findings [$ or none] credited to build [y/n]
First problem they would solve: [ ] Left out: [ ]
Rules vs AI per step: [explained / not explained]
Live example close to ours, with baseline: [y/n]
Exception handled in the demo: [y/n]
Acceptance test on our real examples: [described / vague]
Price: first build $[ ] software $[ ] usage est. $[ ] support $[ ] a month our time [hours]
Ownership clause seen [y/n] runs in our accounts [y/n]
Data: provider and region per AI step [ ] access ends at exit [y/n]
Failure plan: alerts [ ] safe stop [ ] fix time [ ]
Minimum term or exit fee: [ ]
Unanswered: [ ]

Common questions about choosing an AI agency

How do I pick an AI agent?

Pick an AI agent the same way you choose an AI agency: start from one workflow that costs you time, decide whether it needs a model at all or just a rule, then judge the agent on how it behaves when things go wrong. Ask to see it handle a missing field, and check who it flags while it waits. If you are not sure the task needs an agent at all, my plain-English guide to AI agents sorts ten common small-business tasks by the simplest tool that fits.

Which AI company is the best in Australia?

No AI company is the best for every Australian business. It depends on your size, how specific your process is, your budget, and what you want to hold at the end. My comparisons of the best AI implementation agencies and the best AI consulting firms in Australia name firms and say who each suits.

What is the 30% rule of AI?

The 30% rule of AI is an informal rule of thumb, not an official standard, and people use it in different ways. The most common version says AI should take on routine work while people keep roughly 30%, the part that needs judgement. When choosing an AI agency, the useful idea behind it is the same: ask which steps a person keeps, and why.

How much does it cost to hire an AI agent?

It depends on the engagement shape. Some Australian firms publish a price for a first audit or roadmap, some sell packages at a set price, and others, AppliedAgents included, quote after mapping the work. Whatever the shape, ask for the first build, licences, AI usage and support as separate lines. The AI implementation cost guide goes through each.

What questions should I ask an AI agency before hiring?

Ask what it is solving first and what is out, why the job needs AI, whether it can show a live system close to yours, whether it runs AI automation in its own business, what happens when the system is wrong, how you accept the build, what you pay after launch and what you control if you leave. Step 5 gives a good and a weak answer for each.

Should I choose a local AI agency?

Local matters for the mapping and the training, and much less for the build. Watching your team work and teaching them benefit from someone in the room; the build can be done from anywhere. Local or not, look for an Australian entity with an ABN you can check, a named person who answers in your business hours, and a clear answer on where your data is hosted and processed. In my view, an agency two suburbs away that cannot answer the data question is a worse choice than one interstate that can.

Sources

• Published audit and roadmap prices: each firm’s own website, checked 1 to 3 November 2026, as recorded with links in the comparison of AI implementation agencies.
• AppliedAgents: AI implementation service, how AppliedAgents handles data and AI, AI training and ongoing support.
• OAIC, APP 8: cross-border disclosure of personal information; IP Australia, Who owns IP.

Prices, terms and services change. Confirm them with each firm before relying on them. Next review: early December 2026.