Most Australian CFOs we talk to budget for AI the way they'd budget for a new software subscription: take last quarter's number, add a bit, done. That works fine when spend is flat and predictable. It breaks down for AI because usage tends to grow non-linearly as staff find new applications, and because per-token prices have been falling roughly in line with the wider industry trend -- two forces pulling the number in opposite directions at the same time.
A three-scenario model, not a single guess
Rather than one FY27 number, build three: a low case (usage stays roughly flat, current workflows only), a base case (usage grows 30-50% as one or two new workflows go live), and a high case (a significant new use case, like customer-facing chat or a large document-processing rollout, gets added mid-year). This isn't about false precision -- it's about giving the board a range they can plan around instead of a single number that's wrong the moment reality diverges from the assumption baked into it.
Layer in the price trend explicitly. Per-token API pricing for comparable model capability has fallen materially year over year across the industry, which means holding usage flat while assuming flat cost usually overstates the FY27 number. A CFO's model that doesn't account for this tends to over-budget by treating this year's per-unit price as a permanent constant.
Building the model line by line
Start with your trailing three months' actual API and subscription spend as the base rate, not a vendor's list price.
Apply a usage growth assumption per scenario, tied to specific planned initiatives rather than a flat percentage pulled from nowhere.
Apply a modest downward price adjustment (5-15% is a reasonable planning range) to reflect the ongoing industry price trend, reviewed each quarter rather than set once and forgotten.
Add a one-off line for any planned new workflow build cost, separate from the ongoing run-rate -- these are capital-like costs, not part of the recurring subscription line.
Set a quarterly review trigger, not an annual one -- AI usage and pricing both move faster than most FY27 budget cycles were designed to track.
A worked example
A 40-person Sydney professional services firm with $3,200/month in current AI spend built three FY27 scenarios: low case held at roughly $38,000 for the year, base case at $52,000 assuming a new client-reporting workflow launches in Q2, and high case at $78,000 if a planned client-facing chat pilot is approved and scales. Presenting all three to the board, rather than a single figure, meant the eventual overshoot into the base case in Q3 wasn't a surprise -- it was scenario two, already approved in principle months earlier.
This matters more than it sounds. An unexplained budget variance is what erodes trust in a finance function's numbers; a variance that lands inside a pre-approved scenario range is just the plan working as intended.
Presenting the range to a board that wants one number
Boards are used to single-figure budget lines, and a three-scenario AI forecast can initially read as hedging rather than rigour. Frame it the other way: a single number implies false precision for a cost category that's genuinely more variable than most software lines, while a range with named triggers for each scenario is the more honest and more useful artefact. Attach a one-line trigger to each scenario -- 'moves to base case if the Q2 reporting workflow is approved' -- so the board can see exactly what would move the number, rather than treating the range as an unexplained buffer.
Revisit the model every quarter, not annually. AI usage inside a growing team tends to shift faster than a typical FY27 budget cycle assumes, and a forecast that was accurate in July can be materially stale by November if a new workflow launched in between. A short quarterly ten-minute check-in against actuals is enough to catch drift before it becomes a year-end surprise line in the board pack.
None of this requires expensive forecasting software. A shared spreadsheet with the three scenarios, the trigger conditions and a quarterly actuals column does the job for the overwhelming majority of Australian SMBs and mid-market firms budgeting AI for the first time as its own line item rather than folding it into general software costs.
Start it this quarter rather than waiting for the FY27 planning cycle proper. A model with two quarters of actuals behind it by the time budget season arrives is far more credible in the boardroom than one built from a standing start in June.
If you're building an FY27 AI forecast and want a second set of eyes on the assumptions before it goes to the board, get in touch through /contact -- we'll sanity-check the model against what we're seeing across other Australian businesses at a similar scale.
