A four-week AI pilot with ten staff testing a new workflow almost always looks affordable. The API bill is small, the team is enthusiastic, and the business case writes itself. Then the workflow goes to all 200 staff and the monthly bill is eight times what the pilot maths predicted. This is not a pricing surprise. It is a modelling failure, and it happens in the same three or four places on almost every rollout we have seen across Australian businesses.
Concurrency, not just volume
Pilot maths usually multiplies cost per task by expected task count and stops there. Production adds concurrency: many staff hitting the same workflow at the same time, particularly around predictable peaks like Monday morning report generation or end-of-month reconciliation. Concurrent load can trigger retries, rate-limit backoffs, and duplicate calls that a low-volume pilot never surfaces, because ten people spread across a day rarely collide, while 200 people all opening the same Monday report at 8am reliably do.
A Brisbane retail group discovered this the hard way: their pilot for automated stock reorder suggestions cost $180 a month across three stores. Rolled out to 40 stores running the same batch job at store-opening time, the concurrent load pushed monthly cost to $3,100, not the $2,400 a simple 40 divided by 3 multiplication predicted, because the concurrency spike triggered additional retry calls the pilot never saw.
Context growth over the life of a workflow
A pilot workflow is usually tested with clean, short inputs. Production workflows accumulate context over time: a customer service thread that started as three messages grows to thirty, a document review process that references an increasingly long history of prior versions. Token usage per call creeps upward month over month even when task volume stays flat, and a cost model built on week-one token counts underestimates month-six spend.
Model cost against realistic concurrent load, not average daily volume spread evenly across a day.
Track token count per call monthly, not just call count, to catch context creep before the invoice does.
Include retry and error-handling calls in your projection, typically 5 to 15 percent overhead in production that a clean pilot never exercises.
Separate one-off setup cost from ongoing run cost so a healthy pilot doesn't hide a genuinely unaffordable run rate.
Logging, monitoring, and the invisible tail
Production workflows need observability a pilot skips: logging every call for audit purposes, monitoring for silent failures, and often a review layer where a human checks a sample of outputs. None of that shows up in the model API bill, but it is real cost, usually staff time rather than a line item, and it is routinely left out of the pilot-to-production business case entirely. A Melbourne financial services firm found the human review layer alone, checking 10 percent of AI-drafted client communications before send, cost more in staff hours each month than the underlying API usage.
How to model it properly before you scale
Before expanding any pilot past about 20 percent of its eventual user base, run a load test at something closer to real concurrency, even for a single day, and track token growth per call rather than assuming it stays flat. Build the retry and review overhead into the cost model from day one rather than discovering it in month two's invoice. A pilot that costs $200 a month can honestly become a $2,000 or $3,000 a month production workflow and still be a good investment, but only if that number was expected rather than a shock that triggers a panicked renegotiation with finance.
A simple staged rollout that catches the surprises early
Rather than jumping from a ten-person pilot straight to all-staff production, stage the rollout in three steps: the original pilot group, then roughly a quarter of the eventual user base for two to three weeks, then full rollout. The middle stage is where concurrency issues and context creep actually show up, at a scale small enough that a surprise costs hundreds of dollars to discover rather than thousands. An Adelaide manufacturing business used this staged approach for a supplier-communication workflow and caught a retry-loop bug at the quarter-scale stage that would have cost an estimated $1,800 extra a month if it had reached full production undetected.
The businesses that scale AI workflows smoothly are not the ones with the cheapest pilots. They are the ones who modelled production load honestly before signing off on the rollout, and budgeted for the version of the workflow that actually ships, not the clean version that ran for ten friendly testers over four quiet weeks. Treat the pilot number as a floor, never a forecast, and the production invoice stops being a surprise.
