RKAZNA Wall Street Journal report this week put a number on something a lot of businesses are quietly experiencing: only 11% of nearly 400 companies surveyed can accurately project what they'll spend on AI. Not "estimate roughly" — accurately forecast. The other 89% are essentially flying blind on a line item that's growing fast.
This isn't a pricing problem, it's a budgeting category problem
Traditional software has a predictable cost shape: you pay per seat, per month, and the bill barely moves. Agentic AI doesn't work like that. Every task a model handles consumes tokens — and token consumption scales with how much work the system is actually doing, not with how many people have a login. That makes an AI system behave less like a software subscription and more like a human worker: the more it does, the more it costs, in real time, with no fixed ceiling unless you build one.
Most businesses still budget for AI the old way — a flat monthly number picked in advance — because that's how every other software line item has always worked. The forecasting problem isn't that AI pricing is opaque. It's that the budgeting model being applied to it is built for the wrong kind of expense.
The counterintuitive part: cheaper models aren't always cheaper
Here's the part of the study worth sitting with. Researchers from Stanford, CMU, Berkeley, and Microsoft ran over 6,800 math, coding, and science tasks across different models — and found that cheaper-per-token models actually cost more overall in 32% of scenarios. That sounds backwards until you think about why: a cheaper model that needs more retries, longer reasoning chains, or produces answers that need human correction can easily burn more total tokens — and more total time — than a pricier model that gets it right in one pass.
Cost-per-token is not the same as cost-per-outcome. Optimizing the wrong number is how budgets quietly blow up.
What this actually means if you're planning an automation project
- Model choice is a cost decision, not just a capability decision — the cheapest API price doesn't guarantee the cheapest outcome
- Forecasting needs to be based on volume and task complexity, not a flat guess — the same questions we use for automation ROI generally (how often does this run, how often does it need a retry or human correction)
- A system with no monitoring is a system with no budget — if you can't see per-task cost, you can't catch a cost leak until the bill arrives
- Pilot projects should measure cost-per-successful-outcome from day one, not cost-per-API-call
The same math we use for time applies to cost
This is really the same lesson from our piece on calculating automation ROI — the naive number (hours saved, or here, cost per token) is never the real number. The real number accounts for exceptions, retries, and the maintenance cost of keeping the thing accurate. If you're budgeting an AI project this quarter, run the volume-based numbers before you commit to a model or a vendor on sticker price alone — our free ROI calculator uses the same logic to estimate real cost, not list price.
This is also exactly why we build automation with monitoring and clear cost visibility baked in from the start, not bolted on after the first surprising invoice.