The first question every CFO asks before deploying an AI agent is: how much does it cost and when will it pay back? The second is what the tech lead asks: how do we even calculate that ROI in a way that survives an audit?
Three ROI models
In our practice every deployment comes down to one of three models. Each has its own formula and its own gotchas.
Model 1: FTE reduction
The simplest. Count how many person-hours you free up, multiply by hourly cost, subtract the cost of the agent.
ts// Annual ROI from FTE reduction ROI = (saved_hours_per_year × hourly_cost) − (setup + 12 × monthly_cost)
Model 2: deflection rate
Fairer for support. Measure the percentage of tickets the agent closed without escalation, multiply by the cost of a manually-handled ticket.
Model 3: conversion lift
For sales and lead-gen. Conversion lift × LTV × lead volume − agent cost.
Numbers from 6 deployments
| Client | Model | Payback | 12-mo ROI |
|---|---|---|---|
| Nordic Retail | FTE | 3.1 mo | 412% |
| Kapital Bank | Deflection | 4.5 mo | 287% |
| Medline | FTE + Conv. | 2.8 mo | 521% |
| Unicum | Conversion | 5.2 mo | 198% |
| Delta Logistics | FTE | 6.1 mo | 164% |
| Smaczny Dom | Conversion | 4.0 mo | 311% |
What not to count in ROI
- —Soft savings like "the customer was happier" — can't defend
- —Hypothetical FTEs you didn't hire — that's not savings, that's an unspent budget
- —Brand value, NPS, employer branding — separate metrics, don't lump them in
- —Tech lead time to maintain the agent — actually add it to the cost
What next
If this is your first deployment — start with the model where data is cleanest. Usually FTE in support. Calculate conversion lift after the first 90 days, when you have a baseline.