Almost every AI automation project starts with the same question: “will it pay for itself?”. The honest answer is “it depends”, but it depends on a handful of variables, and all of them can be measured before you invest. In this article I show the math I use with clients, the costs that usually get left out and a calculator to test your own case.
The automation ROI formula
The return on an automation comes from three sources: time freed up, errors avoided and faster response. The first is the easiest to measure and, on its own, usually justifies the project:
Monthly savings = hours saved per week × 4.33 × hourly cost
Net gain = monthly savings − monthly cost of the automation
Payback (months) = upfront investment ÷ monthly net gain
12-month ROI = (net gain × 12 − investment) ÷ investment“Hourly cost” is not just salary: include taxes, benefits and overhead. A fully loaded employee typically costs the company well above their gross pay.
Calculate your case
The costs that get left out
AI projects that “don’t pay off” usually got the cost wrong, not the benefit. Always include:
- Process mapping: understanding the current flow, the exceptions and who decides what. It is 20% to 30% of the effort.
- Integrations: connecting the agent to the CRM, ERP, email or spreadsheets. That is where most of the complexity lives.
- Model usage costs: APIs charge by volume of text processed. In high-volume processes, this becomes a meaningful fixed cost.
- Monitoring and tuning: for the first 2 to 3 months, set aside hours to review answers and fix rules.
- Change management: training the team and redesigning the role of whoever did the task manually.
Example: triaging customer service emails
A common scenario in service companies: one person spends about 15 hours a week reading emails, sorting them by topic, answering repeated questions and forwarding the rest.
| Item | Reference value |
|---|---|
| Hours saved per week (70% of the task) | 10.5 h |
| Fully loaded hourly cost | USD 30 |
| Monthly savings (10.5 × 4.33 × 30) | USD 1,364 |
| Monthly cost (API, hosting, monitoring) | USD 150 |
| Upfront investment | USD 3,000 |
| Payback | about 2.5 months |
Notice that the automation does not eliminate the person: they stop triaging and start handling the cases that require judgment. That is where the gains the spreadsheet does not capture show up, such as response time dropping from hours to minutes.
When AI automation is not worth it
- Rare process: if it happens a few times a month, the investment will not pay back.
- Unstable process: if the rules change every week, standardize first, then automate.
- Costly, hard-to-detect errors: in critical legal, medical or financial decisions, AI should suggest and a person should approve.
- Simple, fixed rules: if an “if this, then that” solves it, use traditional automation (Power Automate, Zapier, n8n). It is cheaper and more predictable.
How to measure after launch
Set the baseline before you start: how long the task takes today, how many items per week, error rate and response time. Then track it in a simple dashboard:
- volume handled by AI and volume sent back to a person;
- accuracy (a sample reviewed every week);
- cost per item processed;
- average response time to customers.
This tracking fits in a Power BI dashboard, and it is what supports the decision to extend automation to other processes.