Verification ROI

What does it cost to have a person check every agent action?

Pre-action verification checks each agent action against your rules before it runs and produces a receipt anyone can verify. The review burden is real. A&O Shearman reports that AI saves its lawyers two to three hours a week and cuts contract review time by 30 percent, yet Stanford's RegLab found the leading legal research tools hallucinate 17 to 33 percent of the time, so a person still checks every citation before it goes out. In-house, GC AI's customer study reports 14 hours saved per lawyer per week, with the same verification step on every output. Actions that reduce to a checkable rule no longer need that person. This calculator estimates the review hours that frees, the cost behind them, and how many more actions your team can afford to delegate. Edit any field; every figure updates.

Agent activity
%
min
$
What verification covers
%
Capacity
$
Work stuck in review
$
One bad action
$
Read how verification works

How the estimate works. Review hours today = actions × share reviewed × minutes ÷ 60. Hours freed = review hours today × share of actions that are checkable. Delegation capacity = the actions your current review hours could cover if only judgment calls needed a person.

These are estimates from your inputs, not measured results. Verification does not remove the lawyer's responsibility for the work; it changes what the lawyer has to look at.

Where the figures come from. A&O Shearman time savings and 30 percent contract review reduction: Harvey customer page and Microsoft UK (vendor and customer published). Hallucination rates of 17 to 33 percent: Stanford RegLab, Magesh et al. (independent, peer reviewed). Hourly rates: Clio Legal Trends 2025 (independent). In-house 14 hours per week: GC AI, December 2025 (vendor self-reported, over 100 customers). Compliance officer cost: Robert Half Salary Guide midpoint of $102,750, divided by 1,880 working hours (independent salary guide). No published source measures minutes per review of a single AI action or the share of AI output actually reviewed; those two fields are yours to set.