How to calculate the ROI of an AI implementation

The ROI of an AI implementation starts before development. Without a baseline, the company knows how much it spent, but cannot prove what changed. The calculation must link the solution to an operational unit: ticket resolved, document processed, opportunity advanced or analysis time released.

Record the baseline

Measure a representative period of the current process. Record volume, time per unit, cost of people involved, rework, errors, deadline and business results. Avoid using only averages: an agent can perform well in simple cases and increase the cost of exceptions.

Calculate the benefit

For efficiency, use: volume × minutes saved × cost per minute. Treat freed-up hours as capacity, not automatic financial savings. Expense reduction only exists when a cost is avoided or removed.

For revenue, compare equivalent groups or periods. A commercial copilot can influence conversion, sales cycle or ticket, but campaigns, seasonality and price changes also affect the result. When there is no controlled experiment, state the limitations of the assignment.

Quality also has value. Calculate avoided rework, reopenings, fines, losses or repeated service. In sensitive processes, reducing risk can justify the project even without significant time savings.

Include the total cost

The cost is not just the model API. Add diagnostics, development, integration, licenses, infrastructure, data preparation, assessment, security, training, support and monthly operations. Add the time of experts validating answers. In agents, estimate cost per completed task, not per message.

The basic formula is: ROI = (net benefit − investment) ÷ investment × 100. The payback shows how many months the monthly benefit takes to recover the initial investment. Use conservative, probable and optimistic scenarios.

Protect the result with control metrics

Automation can seem cost-effective while making the experience worse. Track accuracy, escalation rate, human corrections, incidents and satisfaction along with financial metrics. NIST treats risk measurement and management as ongoing activities, not as a launch-only audit.

Simple example

Consider 4,000 requests per month, with six minutes of work on each one. If AI saves three minutes in 60% of cases, it frees up 120 hours per month. Multiply by the charged hourly cost, deduct operation and review and compare with the initial investment. Then, validate that response time, reopening and satisfaction remained within the target.

The best business case does not promise an absolute number. It shows assumptions, source of data, expected range and condition for continuing investing.