How to Calculate ROI on an AI Agent
Most AI ROI models are built to produce a number that justifies a decision already made. Here is one that survives contact with a finance team.
The four-step version
1. Baseline before you build
Count how long the work takes now, in hours per week, measured rather than remembered. If you skip this you will never be able to prove the improvement, and the project quietly becomes unfalsifiable.
2. Price those hours honestly
Use the loaded rate, not the salary. BLS ECEC puts benefits and payroll taxes at 1.43× base pay for private industry. Use the median wage for the role, not the mean — the mean is dragged up by high earners and inflates your case. On General and Operations Managers the gap between mean and median is $29,170, which is enough to make a bad project look good.
3. Discount for what does not get recovered
This is where most models cheat. An agent that handles 80% of cases does not save 80% of the time, because someone still reads the queue, handles exceptions, and checks the output while trust is being established. Assume you recover half of the theoretical saving in year one. If the case only works at 100% recovery, it does not work.
4. Subtract the real running cost
Platform fees, the reviewer’s time, and the maintenance to keep it working as your tools change.
The formula
Annual return = (hours per week × 52 × loaded hourly rate × recovery rate) − annual cost of the agent
Run it with your own numbers →
Three things that make a CFO stop reading
- Savings with no cost line. Every model showing only benefits gets discarded on sight.
- Headcount reduction you will not actually make. If nobody is being let go, it is not a payroll saving — it is capacity. Call it capacity. Capacity is a real benefit and it survives scrutiny; a fake redundancy does not.
- Mean instead of median. See above. Finance will catch it.
The one that is hard to quantify and matters most
Error and delay costs. The invoice paid twice. The lead that went cold over a weekend. The renewal that auto-renewed. These are usually larger than the labour saving and almost never modelled, because they are invisible until someone goes looking. If you can attach a number to even one of them, do — it will often be bigger than everything else in your model combined.