Economics
Cost per decision
Definition
Cost per decision is the fully loaded cost of getting one operational decision made — model calls, retries, tooling, and the human review time that remains — expressed per claim, submission, or file.
It is the only AI cost metric that can be compared to anything. It sits next to the pre-automation cost of the same decision, and it exposes the review burden that pilot economics usually hide from the initial business case. Programs that look cheap per token frequently look expensive per decision once escalation time is counted honestly, which is why this metric should be the one that governs a go or no-go decision rather than the token rate.
Review time is part of the cost
If a system produces a fast answer that a senior person must spend eight minutes verifying, the eight minutes belong in the unit cost, priced at that person's actual loaded rate rather than treated as a free byproduct of the automation. Ignoring them is how automation gets funded and then quietly fails to save anything, because the labour cost simply moved from doing the work to checking the work, without the total ever shrinking the way the original business case assumed.
It makes the routing decision obvious
Once you have cost per decision by segment, the question stops being whether to automate and becomes which slice to automate first — usually the high-volume, low-ambiguity one, where the cost savings are largest and the review burden is smallest. Segmenting the metric this way turns a single, contested number into a prioritized list that operations leadership can actually act on, rather than an abstract debate about whether AI is worth the investment in general.
Compare it against the pre-automation number honestly
The comparison only has value if the pre-automation baseline was calculated with the same rigor — fully loaded labour cost, including the supervisory and quality-review time the manual process also required, not just the front-line handler's time. A common mistake is comparing a fully loaded AI cost per decision against a stripped-down manual cost that excludes overhead the manual process actually carried, which produces a flattering comparison that does not survive scrutiny from a finance team asked to validate it.
In practice
- Compute the pre-automation baseline cost before launch, not after
- Count exception handling and escalation minutes inside the metric
- Break the number down by segment; averages hide the profitable slice
- Load both the manual and automated baselines with the same accounting rigor
