Guide · AI software · Melbourne, October 2026
You’ve seen the demo. Someone in a hoodie built an AI assistant in eleven minutes, and it answered every question perfectly. Meanwhile your software vendor has emailed to say the product you already pay for now “has AI”, for a small extra fee per seat, per month, forever. So which is it? This guide is the honest version of build vs buy AI for Australian businesses: when to buy, when to build, and why the best answer is usually the unglamorous bit in the middle.
Fixed price A$1,950 + GST. We’ll tell you when buying is the right call. It often is.
Short answer: For most Australian businesses, the build vs buy AI decision comes out as “buy” for generic jobs like writing help, meeting notes and search, and “blend” for anything tied to your own workflow: buy the AI platform, then build the thin layer that connects it to your systems, checks its work and hands decisions to people. Build fully custom AI only when the workflow is how you win work and nobody sells it.
Side by side
There are three options, not two. The middle one is the one nobody puts in the keynote.
| Buy | Blend | Build | |
|---|---|---|---|
| What it is | Off-the-shelf AI software, or AI features inside a product you already use | Buy the AI platform, build the connections, checks and hand-offs around it | Custom AI written for your workflow, usually on a model provider’s API |
| Examples | Microsoft 365 Copilot, the AI features in your accounting or CRM software, an industry-specific AI tool | A Copilot Studio agent over a curated library, an AI step inside a workflow with rules and a review queue | A document-reading service that knows your forms, a retrieval app over your project history |
| Time to first use | Days | Weeks | Weeks to months |
| You pay for | Licences, usually per user per month | Licences plus a smaller build | The build, then hosting, model usage and upkeep |
| Fits your process? | You change to fit the tool | The tool is fitted to you, within limits | Made to measure |
| Who fixes it | The vendor, on their roadmap | The vendor for the platform, you or a partner for the glue | You, or whoever you pay to |
| Best for | Generic jobs every business has | Your workflow on someone else’s engine | Work that sets you apart, with data nobody else has |
The options
Before the scorecard, a confession. We genuinely enjoy flat-pack assembly. Yes, even the wardrobe with the 212 parts and the diagram where the little cartoon man looks suspiciously calm. It’s a useful hobby for this topic, because buying, blending and building AI map neatly onto furniture.
Option 1 · Buy
The assembled sofa from the showroom. Comfortable today, same as everyone else’s.
Option 2 · Blend
The off-the-rack suit, taken in by a good tailor.
Option 3 · Build
The hand-made dining table. Beautiful, exactly the right size, and somebody has to oil it.
The AI model itself is the part you should almost never build. Training your own model is a job for companies with research teams and very large electricity bills. What makes AI useful in a 40-person business is everything around the model: which documents it can see, what it does with the answer, who checks it, and where the result goes next.
That surrounding layer is where your process lives. It’s also the part no vendor can sell you, because they’ve never seen how your business actually works. So the sensible split is: buy the engine, and build (or configure) the fit.
If you’re weighing up who should do that fitting, our guide to what an AI consultant costs in Melbourne covers the pricing models. And if you’re not sure you need outside help at all, read AI consultant vs software.
What changed
Three years ago “build vs buy” was a question for IT departments with a whiteboard budget. Now it lands on the desk of every operations manager, for two reasons.
First, AI arrived inside everything you already buy. Your accounting software, your CRM, your document storage and your email all grew an AI button sometime in the last two years, usually with a price attached. You’re already buying AI software whether you decided to or not.
Second, building got cheaper to start. Model APIs mean a developer can stand up a convincing prototype in a day. Convincing prototypes are the most dangerous kind, because they hide most of the work.
And plenty of businesses are already in it. The ABS reported that around 12% of Australian businesses used AI in 2024-25, rising to 22% of medium businesses, while the National AI Centre’s adoption tracker put SME adoption at 44% in February 2026 (the two surveys measure different things, hence the gap). The same tracker found around 65% of businesses not using AI cited distrust in AI decision-making or a strong preference to keep human control. Fair enough. So do we.
The most quoted number in this debate comes from MIT NANDA’s 2025 report on generative AI in business: tools built with external partners reached deployment about 67% of the time, against about 33% for tools companies built on their own, as Fortune reported. [Puts on reading glasses.] Two caveats the headlines skip. It’s a preliminary study of 52 organisations, and the authors say the link doesn’t prove cause. And “buy” in that study means customised tools built with a partner, not software straight out of the box. Which, conveniently, is what we call Blend.
The scorecard
Pick one workflow, not “AI in general”. Answer honestly. The pin moves as you go.
How it scores: each answer is worth 0, 1 or 2 points towards building. Answers that point to generic work, a product that already exists, little maintenance capacity or a short deadline push towards buying. Unusual data, many systems, high volume, strict control and a long-term owner push towards building. 0 to 5 is Buy, 6 to 10 is Blend and 11 to 16 is Build. It’s a conversation starter, not a verdict: one “2” on control can outweigh everything else.
The bit after the demo
Both options have a price tag and a second, quieter price tag. Here are both.
Building custom AI
The build quote
Tax invoice · monthly · forever
Thank you for your business. And your business. And your business.
The iceberg isn’t our invention. Back in 2015, Google engineers published a well-known paper, Hidden Technical Debt in Machine Learning Systems, arguing that the machine learning code is a small fraction of a real system, and the surrounding plumbing is where the long-term cost sits. Swap “machine learning” for “your AI assistant” and nothing has changed.
The practical rule we use: a custom AI build should pay for itself within about a year from the hours and errors it removes, after you’ve counted the running costs. If the numbers only work when you leave out maintenance, buy. Our guide to the ROI of AI automation walks through that maths with real hourly rates.
And for buying: count seats honestly. AI licences priced per user are cheap for five people and a serious line item for fifty, especially if half of them open it twice and go back to doing things the old way.
Worked example
An illustrative example, not a client story. Picture a 40-person building surveying firm. Staff lose time every week hunting for the current version of a procedure, a clause from a past report, or “that thing we did for the warehouse job in Dandenong”. Someone suggests an AI assistant. Here’s how the same problem looks under each option.
If the firm already lives in Microsoft 365, Copilot can search the files, emails and Teams chats each person already has access to. Microsoft lists the Copilot Business add-on at AU$31.40 per user per month plus GST, paid yearly (a promotional AU$26.91 runs to 31 December 2026), on top of a qualifying Microsoft 365 plan. For 40 staff at list price, that’s about AU$15,000 a year plus GST.
The catch: Copilot respects your existing permissions, which is great until you discover what your existing permissions are. The salary spreadsheet shared with “everyone” in 2019 is now very easy to find.
Fits when: files are reasonably tidy and permissions have been cleaned up first.
Use a platform such as Copilot Studio to build an assistant that only reads the approved library: current procedures, standard clauses, finished reports. It answers with a link to the source and says “I’m not sure, ask the technical lead” when it isn’t.
Most of the work is choosing and tidying what it can see, which is work the firm needed to do anyway.
Fits when: the answers must come from approved documents, not whatever’s lying around.
Custom AI on a model API, hosted in an Australian cloud region, connected to the project system and document store, with access rules per project, citations on every answer, a test set of 50 real questions and logs of what it said.
It’s the best of the three and the most work to keep running.
Fits when: the knowledge is what clients pay for, or answers go outside the firm.
For this firm we’d start with Blend. The documents are the firm’s own, so plain Copilot answers from too much, and a full custom build is more machinery than a search problem deserves. If the library later becomes the basis of a client-facing service, that’s the moment to look at Build. The same logic applies to reading documents rather than searching them; our document AI vs OCR guide and the document intelligence page go deeper.
The plot twist
Here’s the part vendors leave out of the keynote: a lot of “AI projects” don’t need AI. If the job is moving data from one system to another, applying the same rule every time, or sending a report every Monday, that’s ordinary automation. It’s cheaper, faster and never makes things up.
Ask “build vs buy AI?” only after you’ve asked “does this step need AI?” If it doesn’t, the choice is between a no-code tool and a custom workflow, and we’ve compared those in custom automation vs Zapier, Make and n8n. Our workflow automation page shows how we build that side.
Whichever you choose
Buying AI software doesn’t buy you out of responsibility. If it gets something wrong, your client calls you, not the vendor. So whatever the build vs buy decision, the same split applies.
That review step has a name: human-in-the-loop. Design it in from day one rather than bolting it on after the first embarrassing email.
[Switches to serious face.] These apply whether you build or buy, and they’re worth reading in the original.
The Privacy Act doesn’t cover every small business, so check whether it applies to you. If you handle client data for businesses that are covered, assume your clients will ask anyway.
Common mistakes
[Clears throat.] Number 3 is ours. We’ll let you guess which part.
Our take
We’re perfectly happy to tell you to buy. We start with how the work actually moves through your business, then pick the cheapest option that fixes it properly. Our rule of thumb fits on a sticky note: buy the commodity, blend the workflow, build the advantage.
Whatever the answer, we add the same things: written documentation, rules that check the AI’s output, a review step for the doubtful cases, and a handover so nothing depends on one person’s memory. Nobody should ever have to ask “who built this?”
See how we approach it on our solutions page, browse the automation use cases (the service desk ticket triage one is a good example of a blend), or read how we handle AI decision systems.
Who wrote this
FAQ
Build vs buy AI is the decision between buying off-the-shelf AI software, such as Microsoft 365 Copilot or the AI features in your existing tools, and building custom AI for your own workflow, usually on a model provider’s API. There is also a middle option, often the best one: buy the AI platform, then build the connections, checks and review steps that fit it to how your business works.
Most small and mid-sized businesses should buy first. Off-the-shelf AI handles generic jobs like drafting, summarising and search well, and it’s working in days. Build, or blend, when the workflow is specific to you: your forms, your approvals, your clients’ data. Fully custom AI makes sense when the work is part of how you win clients and no product does it.
Buying is almost always cheaper to start. Building can be cheaper over three years when a workflow runs at high volume, or when per-user licences would cover dozens of staff. The honest comparison counts running costs on both sides: seats and renewals for buying; model usage, hosting, testing and support for building. If the build only pays off when you leave out maintenance, buy.
Blend means buying the AI engine and building the fit. You use a bought platform or model, then add a thin custom layer: connections to your systems, rules that check the AI’s output, and a queue where a person reviews doubtful cases. You get vendor-maintained AI without bending your process around a product that was designed for everyone.
Custom AI makes sense when four things line up: the work sets you apart, no product does most of it, the data is unusual or must stay under your control, and someone will own the system after launch. High volume helps the case. If you can’t name who looks after it in year two, it isn’t ready to build.
For generic work inside Microsoft 365, often yes. As of October 2026 Microsoft lists the Copilot Business add-on at AU$31.40 per user per month plus GST, paid yearly (promotional AU$26.91 to 31 December 2026), plus a qualifying Microsoft 365 plan. It searches what each person can already access, so tidy permissions first. It won’t run your specific workflow, like checking a form against your rules, without some building around it.
No. If the Privacy Act covers your business, you stay responsible for personal information whether the AI is bought or built. The OAIC recommends not entering personal information into publicly available generative AI tools, APP 8 applies when data goes overseas, and from 10 December 2026 privacy policies must explain significant automated decisions. Ask vendors where data is stored and for proof of testing.
A 30-minute fit call is free. The Pain Point Audit is A$1,950 + GST for a half-day walkthrough of your processes, with a written report within 5 business days covering what to buy, what to build and what to leave alone, plus a fixed price for the first build. The fee is credited in full against a first build of A$5,000 + GST or more signed within 60 days.
Bring us the workflow, the vendor quote and the developer’s estimate. The Pain Point Audit is a half-day look at how work really moves through your business, with a written report in 5 business days: where the hours go, which jobs need AI and which need a simple rule, what to buy, what to build, what stays with your people, and a fixed price for the first fix. No hoodie required.
Fixed price A$1,950 + GST. Credited if you go ahead with a build.
