We've now run enough AI cost reviews across Australian small and mid-sized businesses to see the same five waste patterns show up again and again, in roughly the same order of frequency. None of them are exotic -- they're the AI-era version of paying for a gym membership nobody uses, just harder to notice because the bill arrives as one line instead of several.
Leak one: unused or forgotten seats
The single most common finding. A team subscribes five or ten seats to an AI tool during initial rollout enthusiasm, two or three people actually use it regularly, and nobody revisits the seat count for a year. A 22-person Sydney firm we reviewed was paying for 18 seats on a per-user AI tool; usage logs showed 7 active users. That's roughly $600 a month in dead seats on a single tool alone.
Leak two: the same context sent fresh every time
A workflow that re-sends a full policy document, product catalogue or knowledge base with every single query, instead of using caching to remember it between calls, pays full price for the same content repeatedly. This one is invisible on the invoice -- it just looks like normal usage -- and only shows up once someone actually checks how the workflow is technically configured, not just what it costs.
Leak three: wrong model tier for the task
Simple, repetitive tasks (formatting, basic classification, short extraction) routed to the most expensive, most capable model by default, usually because that's what the first working prototype used and nobody revisited the choice once it scaled to production volume. This is often the single biggest line item once found, because it multiplies across every call a high-volume workflow makes.
Leak four: real-time pricing for non-urgent work
Overnight reports, bulk summarisation and batch data processing running at real-time API rates because nobody set up batch processing for work that never actually needed a live response. Non-urgent, high-volume workloads are the clearest candidates for this fix, and the ones most often left on the expensive path simply because it was the default when the workflow was first built.
Leak five: overlapping tools doing the same job
Two departments independently subscribing to functionally similar AI tools without knowing the other exists.
A trial subscription from six months ago that nobody cancelled once the trial period rolled into a paid plan.
A tool kept 'just in case' after the workflow it supported was discontinued or replaced.
What a typical audit finds in total
Across the reviews we've run, the five leaks combined typically account for 25-45% of total AI spend being addressable without any drop in output quality -- not a rebuild, just a cleanup. A Melbourne business spending $2,400 a month found $780 a month addressable across all five leaks once someone actually went looking, none of it requiring new tooling, just configuration changes and a seat-count review.
The pattern holds because none of these five leaks are anyone's fault specifically -- they're the natural residue of a business adopting AI quickly, which is usually the right call, without building in a review cadence to catch drift afterwards. If your AI spend has grown past $1,000 a month and nobody's specifically checked for these five things, there's a good chance at least two are quietly present.
Running a light version yourself before booking a full audit
You don't need an external review to catch the two biggest leaks. Pull up your AI billing dashboards this week and check two things specifically: actual per-seat usage against seats paid for on any per-user tool, and whether your highest-volume workflow is using the cheapest model tier that still does the job well. These two checks alone, done in under an hour, typically surface the largest single leak in a business's AI spend without needing anyone else involved.
The other three leaks -- repeated context, batch-eligible work on real-time pricing, and overlapping tools -- are harder to spot without technical access to how workflows are actually configured, which is usually where a proper review earns its cost. But starting with the two self-serve checks first means you're not paying someone to find something you could have found yourself in an afternoon.
None of these five leaks require a rebuild to fix once found. A seat count gets trimmed in an afternoon, a model tier gets changed with a configuration update, and an overlapping tool gets cancelled with an email. The hard part was never the fix -- it was nobody having looked in the first place.
If you want an actual audit rather than a checklist to run yourself, get in touch through /contact and we'll tell you honestly what we find before recommending anything.
