Most Australian businesses running AI in production find out what it cost them once a month, when the invoice arrives. By then the spend has already happened and there's nothing left to do but react. Cost observability is about closing that gap -- seeing spend building up in near real time, before it becomes a surprise line item on a P&L nobody was expecting to move.
Why the invoice is the wrong place to find out
A 30-person Perth marketing agency discovered a runaway document-summarisation workflow only when their monthly Anthropic invoice came in at $2,100, nearly triple the usual $750. The cause was a looping error in a scheduled task that had been silently re-processing the same batch of files for eleven days. With basic usage tracking in place, that would have shown up as an unusual daily spike within 24 hours, not a month-end shock discovered by the bookkeeper.
What a basic observability setup actually needs
Daily token usage logged per workflow or project, not just a single account-wide total
A simple alert threshold (email or Slack) that fires when a day's usage exceeds a set multiple of the normal average
Tagging requests by originating workflow so a spike can be traced to its source in minutes, not hours of log-diving
A weekly five-minute glance at the trend line, not just a monthly total, to catch drift before it becomes a pattern
You don't need enterprise tooling to start
This doesn't require a dedicated observability platform. For most Australian SMBs, a simple script that logs token counts per API call to a spreadsheet or lightweight database, checked against a daily threshold, catches the majority of runaway-cost incidents. The build cost for a basic version typically sits under $1,500, against the kind of surprise the Perth agency had, which cost them an unbudgeted $1,350 above their normal month for a fault that ran undetected for over a week.
The everyday value, not just the disaster case
The bigger payoff isn't catching disasters, though. It's the ordinary visibility that lets you notice a workflow's cost creeping up gradually as usage grows, and decide deliberately whether that growth is worth it, rather than discovering it after the fact in an invoice you can't do anything about. A gradual 15% month-on-month increase is easy to miss on a single total, and obvious the moment you're looking at a per-workflow trend line each week.
If you're running AI workflows with no visibility into daily spend, that's usually a half-day build to fix properly. Reach out through /contact and we'll help you set up tracking that actually catches problems before the invoice does.
Building the alert without over-engineering it
A working alert doesn't need machine learning or anomaly detection to be useful. A rule as simple as 'flag any day where token usage exceeds 2.5 times the trailing seven-day average' catches the overwhelming majority of runaway-cost incidents, because most cost blowouts are step-changes (a loop, a misconfigured retry, a workflow accidentally triggered twice) rather than gradual drift. Gradual drift is better caught by the weekly trend-line review than by a daily alert, which is why both layers matter rather than just one.
The other detail worth getting right early is where the alert goes. An alert that lands in an inbox nobody checks until Monday defeats the purpose. For most Australian SMBs, routing it to a phone via SMS or a Slack channel someone actually watches during business hours is the difference between catching a Perth-agency-style incident on day one instead of day eleven.
Finally, treat the first month of tracking as a calibration period rather than a source of alarms. Normal usage has its own natural variation -- a busy week for client work will look like a spike if you set the threshold too tight from day one. Give the system a few weeks to establish what a genuinely normal range looks like for your business before trusting the alert to fire only on real anomalies, otherwise the first few false alarms will train everyone to ignore it exactly when a real one shows up.
This same logging also pays off outside the alarm scenario, at tax and budget-review time. Being able to show an accountant or a board a clean monthly trend by workflow, rather than a single opaque total, turns an AI line item from an unexplained cost into a well-understood, defensible part of the operating budget, which matters more than it sounds once spend crosses a few thousand dollars a month and someone starts asking questions about it.
Set it up once, review the trend weekly, and it becomes background noise that occasionally saves you a genuinely expensive surprise.
