Customers Loved the AI Features and the P&L Broke
Inference stopped behaving like a project cost and started behaving like cost of goods, which is why the gate that saves next year's guidance has to be installed at this quarter's ship reviews.
The fix that worked was not a discount
Canva's remediation is the most transferable fact here and the least discussed. Instead of throttling the features customers were consuming, the company rebuilt its delivery architecture, and it reports cost per task down nearly 90% since Canva AI 2.0 launched in April. That is a deeper reduction than any vendor price cut on the market. The conclusion is uncomfortable for anyone waiting out token deflation: margin recovery lives in your stack, not in your vendor's price list.
Look also at what actually broke. Melanie Perkins' explanation is that demand for the AI features significantly exceeded expectations. The forecast error sat on the demand side; the damage landed on the cost side. Average-case cost modeling cannot catch that shape of failure, which is why the gate that matters is cost per task at projected peak demand rather than at plan.
Where the evidence converges
Four independent readings converge on one point: compute became continuous cost of goods while most companies were still budgeting it as project capex. Roughly two-thirds of enterprises run AI in production, and many have no cost or utilization visibility by workload. In the same body of research, 95% delayed or canceled AI projects, with governance and compliance among the top blockers, and nearly three-quarters say AI made data governance harder rather than easier. Adoption is not the constraint. Instrumented adoption is.
| What's in the plan today | What it misses | What replaces it |
|---|---|---|
| Token or seat price by vendor | Retries, reasoning overhead, human rework | Cost per completed task |
| Average-demand cost model | Success cases where usage overshoots plan | Peak-demand unit economics at the ship gate |
| AI spend as a project line | Continuous cost-of-goods behavior | Workload cost reported beside gross margin |
The open question is which cost metric wins. Frontier labs are defending premium list prices with an efficiency argument rather than a capability one, and at least one industry analysis claims premium models finish work for less once retries and human rework are counted. That analysis does not disclose its sponsor and it happens to flatter two US labs, so treat it as a hypothesis. The hypothesis still hands you the right question before your vendors do.
The demand-side version of the same bill
Upwork cut guidance and blamed AI automation and deteriorating organic search performance in the same breath. That is the template for any business that monetizes human hours or acquires customers through search: the cost line and the revenue line move against you at once. If either describes part of your portfolio, the pricing question — outcomes or time — is immediate rather than planning-season.
The smart move
Install the gate before you fund more scope. A measured cost-per-task number, owned jointly by the CFO and the CTO, does three things no vendor negotiation can: it makes an AI feature refusable, it makes guidance defensible, and it converts inference efficiency from an engineering hobby into a named function with a target attached.
Customers loving your AI features is no longer evidence the strategy works.
What to do
Install a cost-per-task gate on every AI feature before the next ship approval, modeled at projected peak demand rather than average, with the CFO owning the gate and the CTO co-signing.
Commission a 30-day effective-cost benchmark across your top five production AI workloads, counting retries and human rework, and report it beside gross margin at the next board meeting.
Name an owner for inference efficiency this quarter with a target reduction on your three highest-volume workflows, covering routing, caching and distillation.