Article chapter 07 of 08
Use observed economics to make decisions
From Budgeting and governing an AI workflow after the prototype
After controlled use begins, replace assumptions with measured values. Keep the original budget so differences remain visible. Review cost per completed unit, exception rate, active review time, source maintenance and correction work together.
Segment the results enough to support a decision. One input type or business unit may account for most exceptions. A model change may reduce inference cost while increasing review because answers become less consistent. Aggregate cost per case can hide both effects.
Optimisation should address a measured cost or failure. Options include improving input validation, removing an unnecessary model call, caching stable context, using a smaller model for a bounded classification, changing retrieval, clarifying the review interface or declining unsupported cases earlier. Test one material change at a time against a fixed evaluation set and operational sample.
Watch for cost shifting. Shorter prompts may increase retries. Lower review coverage may increase support and correction. A cheaper provider may need new integration and governance work. Record the full workflow effect before claiming a saving.
Set decision points in advance. Examples include pausing intake when exceptions exceed operational capacity, returning to manual handling when source quality falls below the accepted condition, or requiring a new approval when monthly spend crosses the funded range. The exact thresholds need local evidence and accountable sign-off.
Some workflows will remain expensive and still be worth operating because they reduce delay, improve consistency or make previously unmanageable work possible. State the intended benefit and how it will be observed. Do not convert every benefit into a speculative dollar figure.