Every AI vendor claims transformative ROI. Few show their arithmetic. This post shares the actual math from 340 agent deployments across our client base — including where returns disappointed, because that's where the useful lessons live.
Headline numbers: median first-year ROI of 3.1×, mean of 3.8× (pulled up by a few exceptional automation projects), with the middle 80% of deployments landing between 1.6× and 7.2×. ROI here means measured value (labor hours at fully-loaded cost, recovered revenue, avoided hires) divided by total cost (implementation plus subscription).
What separates the top quartile
High-ROI deployments share three traits. First, they automate high-frequency work: an agent handling 3,000 monthly conversations amortizes its cost far better than one handling 100. Second, they had a clean baseline — clients who knew their current handle times and conversion rates could target the right process and prove the delta. Third, they gave the agent real system access; agents restricted to answering without acting deliver roughly a third of the value.
The bottom decile is instructive too. Weak returns almost always traced to one of two causes: automating a low-volume process that didn't cost much to begin with, or an organization that never routed real volume to the agent — the AI equivalent of hiring someone and giving them nothing to do.
How to forecast your own number
The estimate that holds up in practice is simple: (hours currently spent on the process per month) × (fully-loaded hourly cost) × (realistic automation rate, usually 60–75%) plus any revenue effect (recovered calls, faster lead response), divided by total first-year cost. Run it honestly and it will tell you both whether to automate and what to automate first.
Our ROI calculator implements exactly this arithmetic, and a scoping call pressure-tests your inputs against comparable deployments. Skepticism about vendor ROI math is healthy — insist on the formula, not the testimonial.