Ask most Australian business owners what their AI stack costs and they will quote a per-seat number: $30 a head for Claude, $20 a head for a scheduling tool, and so on. Ask what it delivers and the answer gets vague fast. Per-seat cost is the wrong unit for judging whether AI spend is working, because it measures access, not output.
Why cost per seat misleads
A licence sitting unused costs the same as one running fifty tasks a day. Per-seat pricing makes AI spend look like a fixed cost, similar to office software, when the actual value varies enormously by how much real work runs through it. Two staff on the same $30 seat can produce wildly different returns, and a per-seat report will never show you which is which.
Cost per outcome flips the question. Instead of 'what do we pay per person', it asks 'what does it cost us to produce one finished thing', a drafted proposal, a processed invoice, a resolved support ticket. That number is comparable across tools, comparable over time, and directly tied to a business result an owner actually cares about.
Building the number
The calculation itself is not complicated. Take total AI-related spend for a defined period, subscriptions, API usage, and a reasonable estimate of staff setup and review time, and divide by the count of a specific completed outcome over that same period.
Pick one outcome that matters: invoices processed, proposals sent, support tickets closed, job quotes issued.
Total the AI cost tied to that workflow specifically, not your whole AI bill.
Divide cost by outcome count for the period to get a per-unit figure.
Compare that figure to what the outcome cost before AI, using staff time at a loaded hourly rate.
A worked Sydney example
A Sydney property management firm processes maintenance requests through a Claude Cowork workflow: intake email, triage by urgency, drafted response to the tenant, and a work order sent to a tradesperson. Direct AI cost for the month, API usage plus a share of the Cowork subscription, came to $340. The workflow processed 610 maintenance requests. Cost per outcome: roughly 56 cents per request. Before the workflow, a property manager spent an average of eleven minutes per request at a loaded cost of about $38 an hour, or $6.97 per request. The AI workflow is not free, but at 56 cents against $6.97, the comparison is not close.
That per-outcome figure is also what makes a renewal decision easy. A per-seat number invites a vague debate about whether $30 a month feels worth it. A per-outcome number of 56 cents against a previous cost of $6.97 is a business case, not an opinion.
Where this gets harder, and how to handle it
Cost per outcome breaks down when a workflow produces outcomes of wildly different sizes, a one-line customer reply and a twelve-page proposal are not the same unit of work. The fix is to bucket outcomes by rough size or complexity before averaging, rather than lumping everything into one number that hides the variance. It is more setup work up front, but it protects the figure from being gamed by a run of easy tasks.
It is also worth separating the cost of building a workflow from the cost of running it. The Sydney property firm's 56-cent figure covers ongoing API usage only. The initial setup, roughly fifteen hours of a Cowork skill build at a loaded rate, was closer to $1,100 as a one-off. Amortised over the first year at 610 requests a month, that setup cost adds about 15 cents to the true per-outcome figure, still a fraction of the $6.97 manual baseline, but worth stating honestly rather than pretending a workflow was free to build.
Using the number for vendor decisions
Once a business has a genuine cost-per-outcome figure for its two or three highest-frequency workflows, vendor conversations change shape. Instead of negotiating a per-seat discount, an Australian business owner can ask a vendor to quote against the actual unit economics: what would this tool need to cost per invoice, per ticket, per proposal, to beat what we are already running. Most vendors have never been asked that question and it is a far harder one for them to dodge than a seat-count discount request.
Start with one workflow, not your whole business. Pick the AI-assisted task your team runs most often, work out its true cost per outcome this month, and use that single number as the benchmark for every renewal and every new tool pitch from here. It will tell you more in five minutes than a stack of per-seat invoices ever will.
