How to measure whether an AI agent pays off for your business

The agent may finish a task in seconds and leave you with ten minutes of checks. If you only measure how long it takes to reply, you do not know what changed in the business's work.

To assess a trial, I would start with a task with a recognisable endpoint: a correct record, an approved reply or a document ready to use. I would compare the current and new processes through to that same endpoint, including cases that need redoing.

The time that actually becomes available

This example is hypothetical, with no client data. I assume 300 documents a month. The complete manual process takes six minutes per document; the new one requires two minutes of handling and review. In addition, 30 documents need five extra minutes of correction, and there are two hours of internal maintenance a month.

Human work per month	Calculation	Minutes
Complete manual process	300 × 6	1,800
Handling and review in the new process	300 × 2	600
Additional corrections	30 × 5	150
Internal maintenance	2 × 60	120

That leaves 1,800 − 600 − 150 − 120 = 930 minutes, equivalent to 15.5 hours a month. Correction is additional to review; maintenance has already been deducted. If a person has to wait in front of the screen, that wait also counts as human work.

I would not infer this time from the proportion of cases the agent resolves alone. One lengthy exception can consume more time than many simple cases. Record duration and outcome, as well as the number of interventions.

Put a euro value on time without turning it into cash

At an assumed labour cost of 24 €/hour, the freed capacity is valued at 15.5 × 24 = 372 €/month. If tools, usage and infrastructure cost 90 €/month, the estimated balance in capacity value is 282 €/month. I do not deduct the two internal maintenance hours again.

I assume an implementation cost of 1,800 €, including training and initial internal hours. Over twelve months, the estimated balance would be 12 × 282 − 1,800 = 1,584 €. It remains an economic valuation of available capacity, conditional on your being able to use it.

To discuss financial ROI, I would define, over the same horizon, (attributable monetary benefit − total cost) ÷ total cost × 100. Total cost would include implementation, operation and additional human work, counted only once. Monetary benefit would be demonstrated additional contribution margin or expenses you actually stop paying. Recovered time does not automatically count as income.

If pay remains the same and you do not know what activity will fill those hours, this example does not allow cash recovery to be calculated. It also excludes taxes, financing and damage from undetected errors.

Measure the full cost of each result

For an API, I would not count just the final message. Google documents input and output token consumption and usage records; the workflow may need several calls. Gemini token documentation. I would consult the provider's records and invoices, without transferring prices from one example to a different system.

Before approving the trial, establish which results are acceptable and when a person must intervene. A plausible reply that ends in an improper payment is not a resolved case. Errors with consequences need their own controls, separate from the minutes calculation.

There is no monthly amount that makes every task profitable. There is volume, implementation cost, review and a specific use for the capacity you recover. A trial deserves to continue when those data support the decision.