Build vs buy: when an Australian SMB should build custom AI, and when not to

That is not a knock on custom builds. We build them. It is that the default has to be buy, because the off-the-shelf tools are extraordinarily good and cheap relative to what it costs to build and own something yourself. Make custom prove its case.

Default to buy, and here is what “buy” means

For almost everything an SMB wants AI to do, the answer is a paid business tier of Claude, ChatGPT or Gemini, configured properly, with an automation layer wired over the top.

“Configured well” is the part people underestimate. It is not just handing out logins. It is set-up projects and custom instructions for your common tasks, your own documents and SOPs connected so the tool answers from your business, and clear data rules. A well-configured deployment handles drafting, research, summarising, analysis, customer comms, internal Q&A and most of what an SMB needs - without a line of bespoke code.

The automation layer connects that intelligence to your other systems. Tools like n8n - whose community edition is free to self-host - or Zapier let you trigger AI steps inside real workflows: an email arrives, the AI drafts a reply, it lands in a queue for a human. That is “building” in the sense most SMBs actually need, and it sits on top of bought models rather than replacing them.

Build custom only when one of four things is true

Custom earns its keep when at least one of these is genuinely true. Not “would be nice”. True.

• It’s a genuine differentiator. The workflow is part of how your business wins - the thing customers come to you for - not generic back-office work every business has. Build what makes you different; buy what everyone needs.
• No off-the-shelf tool can fit it. You’ve genuinely tried to configure the shared tools and they cannot be made to do the job. This is rarer than people think, so be honest about whether you’ve actually tried.
• The volume justifies it. You have enough seats or automated runs that subscription pricing would cost more than building on usage-based API pricing and running it yourself. This is a real number you can calculate, not a feeling.
• Data or control demands it. Sensitivity, residency or control requirements mean the workflow cannot live inside a shared product, even a paid no-training business tier.

One trigger can be enough. Zero triggers means you are buying. “It would be cool to build” is not on the list.

The true cost of building is everything after version one

Here is the trap that sinks most SMB build projects: the first version is the cheap, fun part. Owning it is where the cost lives, and it lives there forever.

• Maintenance. Things break. Integrations change. Edge cases appear that no one imagined. Someone has to fix all of it, indefinitely.
• Security and data handling. Doing this properly - access control, keeping client information safe - is real, ongoing work, not a one-time checkbox.
• Model upgrades, the big one in AI. Frontier models update every few months. A build wired to a specific model needs testing and adjustment each time to stay current and cheap. This cost does not exist when you buy, because the vendor absorbs it. When you build, it is yours.

There is also the raw usage cost underneath. API pricing is usage-based - Anthropic’s Sonnet-class models, for example, are priced per million tokens of input and output - so a high-volume build runs a real monthly bill on top of the engineering.

The hybrid reality: buy the model, build the thin layer

When the answer is “build”, it almost never means build a model. No SMB should train its own - that is a job for organisations spending hundreds of millions. What “build” actually means in 2026 is the hybrid: buy the hard, expensive part, and build a thin layer on top.

• Buy: the model itself, through an API like Anthropic’s or a business subscription - the world-class intelligence you could never build yourself.
• Build: the small custom layer that does your specific job - the prompts, the connection to your data and tools, and the workflow logic that makes it yours.

You are renting frontier intelligence and wrapping a small, owned layer around it. That layer is cheap relative to its value, and it is where your differentiator lives. Almost every SMB custom AI project worth doing is this shape.

How to actually decide

Run the candidate workflow through three questions in order.

1. Can a well-configured off-the-shelf tool do this?Genuinely try, with proper configuration and your data connected. If yes, you’re done - buy it.
2. If not, is one of the four triggers true - differentiator, no tool fits, volume, or data and control? If no trigger is clearly true, the answer is still buy, and you keep looking for a configuration that works.
3. If a trigger is true, build the thin layer over a bought model - and price in the maintenance, security and upgrade burden before you commit.

The bias should always be towards buy, then hybrid, and only rarely towards a full custom build. The businesses that get the most from AI are usually not the ones with the most bespoke software. They are the ones who bought well, configured properly, and got their whole team using it.