Building an AI system and running one are different jobs. A Brisbane logistics operator we spoke with had a working Claude-based dispatch assistant within three weeks of starting -- and then nobody owned it. Six months later the prompts were stale, nobody had checked the API bill since launch, and two staff had quietly gone back to doing it manually because a recent update had started producing odd output.
What actually needs looking after
An AI system that touches customers or money needs the same ongoing attention as any other piece of business infrastructure, just in smaller doses. That means someone checking output quality on a rolling basis, someone watching the monthly spend line, and someone who knows what to do when a model update or connector change breaks something that worked yesterday.
None of this requires a dedicated engineering team. For a 10 to 30-person Australian business, LLMOps in practice is usually one person spending two to four hours a fortnight: skimming a sample of recent outputs, checking the API dashboard for spend spikes, and keeping a short changelog of what prompts or workflows changed and why.
The three things that actually break
Prompt drift -- a prompt tuned for one product line quietly stops fitting as the business adds new products or changes terminology.
Silent cost creep -- a workflow that ran occasionally becomes a daily habit and the bill triples without anyone noticing until the invoice lands.
Upstream changes -- a connected tool changes its data format or a model version updates its behaviour, and output quality shifts without an error being thrown.
A minimum-viable ops routine
Set a recurring fortnightly slot, even just 30 minutes, to review a handful of recent outputs against a simple pass/fail bar. Pair it with a monthly look at total AI spend against a budget line -- for a business spending $400 to $1,500 a month on AI tools and API usage, catching a cost spike within a month rather than a quarter is the difference between a minor correction and an awkward finance conversation.
Write down who owns this. Not IT, not 'whoever set it up' -- a named person, even if AI oversight is 5% of their role. Ownership without a name is the most common reason a working system quietly degrades in an Australian SMB.
Building the routine into someone's actual job
The businesses that keep this working long-term don't treat it as a side project bolted onto someone's week. They put a recurring calendar slot in place before the AI system launches, not after it starts drifting, and they name the fortnightly review in the same document that describes who owns invoicing or who owns the website. A Gold Coast trades business we spoke with folded their AI check into the same Friday slot where the office manager already reconciled the week's job sheets -- fifteen extra minutes, same day, same person, no new habit to build from scratch.
Worth budgeting for separately: the occasional half-day when something genuinely breaks, a connected tool changes its API, or a model update shifts behaviour enough that a workflow needs re-testing. This happens a handful of times a year, not monthly, but it's real cost that a lot of Australian SMBs forget to plan for when they budget an AI project as a one-off build rather than an ongoing system with a small maintenance tail.
Signs the routine has already lapsed
A few warning signs are easy to spot if you know to look: staff quietly going back to the old manual process for part of a workflow, an API bill that's crept up without anyone remembering why, or nobody being able to say when the prompts were last reviewed. Any one of these on its own is a minor flag. All three together usually means the ownership gap has been open for months, and it's cheaper to fix it now with a short audit than to wait until a customer-facing mistake forces the conversation.
None of this needs to be formal. A shared note with three headings -- what changed, what it costs, who's watching it -- kept up to date fortnightly does more for the health of an AI workflow than an expensive monitoring platform nobody reads. Start there before buying tooling.
If your AI workflows have been running unattended for a while and you're not sure what state they're actually in, we run a short health check that covers cost, output quality and ownership gaps. Reach out through /contact and we'll tell you honestly whether it needs a rebuild or just a Tuesday afternoon of housekeeping.
