Two Line Items the Standard Automation ROI Never Sees
The standard way companies justify an automation purchase is arithmetic: hours saved, multiplied by labor cost, minus build cost. AWS argues the formula was built for rule-based tools, and that it carries assumptions an AI agent breaks.
Two of those assumptions leave holes rather than rounding errors. Tasks are assumed to be rule-based, so there is no line item for exceptions. And the work is assumed to be the whole job, so the cost of human oversight never appears at all.
That second omission changes how a proposal should be read. A build cost is something you pay once; supervision tends to be a recurring cost, which means a tidy ROI can describe a deal you would not actually take.
AI-generated discussion · Different perspectives
Exceptions are a scope problem. Supervision is a bill that keeps arriving. If only one of the two gets a line, it should be the recurring one.
A model with no slot for exceptions is also telling you which processes were ever a good fit for automation: the ones where nothing unusual happens. That is a selection rule dressed up as a blind spot.
AWS goes further than the two examples — it claims the old model misses most of the value agents create. Worth remembering that the party proposing the new framework is also the one estimating how much the old number leaves on the table.