How do you calculate the ROI of AI agents?

Multiply tasks per month by minutes saved per task, loaded cost per minute and the share of time you actually redeploy. Add the value of fewer errors. Subtract tokens, hosting, maintenance and review. Divide net monthly benefit into build cost for payback. Replace every input with measured values from a pilot, not vendor claims.

How do I measure AI productivity in my team?

Time the same real tasks with and without AI, record error rates, and compare on a frozen evaluation set. Do not rely on self-reports. In METR's 2025 trial, developers expected to be 24% faster yet took 19% longer. Measure output, quality and review time, then check whether saved time was redeployed.

How long does an AI automation take to pay back?

It depends mostly on review time and how much saved time you redeploy. In our invoice example with a €4,000 build, payback is 2.8 months when review takes 2 minutes and all saved time counts, 6.5 months at half realisation, and 18.7 months with 4-minute reviews. All inputs are stated assumptions, not measurements.

How big should an AI evaluation set be?

For a first pilot, 100 to 200 real, anonymised cases including hard ones, labelled by two people and frozen. Size sets your precision: 95 correct out of 100 gives a 95% Wilson interval of about 89% to 98%, while 190 of 200 narrows it to about 91% to 97%. Add cases for failure types you find.

When is AI automation not worth it?

When volume is too low to repay the build, when mistakes are costly and cannot be checked by a person, when the process or data keeps changing, when a rule or script does the job, or when review takes nearly as long as the task. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027.

AI Agents ROI: A Practical Formula for Your Business

AI agents ROI comes down to time saved times loaded cost, minus tokens, review and maintenance. A formula, a worked example, a pilot method and when not to automate.

By Alexandru Secureanu4 October 20267 min read

AI agents ROI is the net monthly benefit of an automated workflow divided by what it costs to build and run. The benefit is time saved multiplied by loaded labour cost, scaled by the share of that time you really redeploy, plus the change in error cost, minus tokens, maintenance and human review. Build the case per workflow from measured inputs, and expect review time to decide the result. Figures below were checked on 4 October 2026.

How do you calculate the ROI of an AI agent?

Use one formula per workflow and fill it with measured numbers.

Monthly benefit  = V × (t0 − t1) × c × r  +  V × (e0 × k0 − e1 × k1)
Monthly run cost = V × m + F + M + O
Net monthly      = Monthly benefit − Monthly run cost
Payback (months) = Build cost ÷ Net monthly
12-month ROI     = (12 × Net monthly − Build cost) ÷ Build cost

V is tasks per month. t0 is minutes per task today, and t1 is human minutes per task with the agent, review included.
c is loaded cost per minute: wages, employer contributions and overhead. r is the share of saved time that becomes cash or valuable output, from 0 to 1.
e is the error rate that reaches the customer or the books, before (0) and after (1), and k is the cost of one error.
m is model cost per task (see our guide to AI inference costs), F is hosting and tools, M is maintenance and O is oversight such as sampling outputs.

Our earlier article on automation ROI for small teams covers rule-based automation. Agents add two lines that rules do not have, token cost and review of outputs that can be wrong.

What does research say about AI productivity effects?

Published effects are real but uneven, and they depend on the task. Four studies set the range.

In a study of 5,179 customer support agents, access to a generative AI assistant raised issues resolved per hour by 14% on average, 34% for novices and little for experienced workers (published in the Quarterly Journal of Economics, 2025).
In an experiment with 758 consultants, AI users did 12.2% more tasks, 25.1% faster, on tasks inside the AI frontier, but on a task outside it correct answers fell from 84.5% to between 60% and 70.6%.
METR's 2025 randomised trial of 16 experienced open-source developers found tasks took 19% longer with AI, although they expected to be 24% faster. A February 2026 follow-up with 57 developers estimated changes in time of −18% and −4% in two groups, with wide intervals, and METR says selection effects make the data weak evidence and that speedups are likely larger now.
In Denmark, the typical chatbot user reported 2.8% time savings, and the authors found precise null effects on earnings and recorded hours, ruling out effects larger than 2%.

In the same Danish data, workers report that only 3–7% of their time savings translate into higher earnings. Our reading is that self-reports are unreliable, effects vary by task, and time saved is not money saved unless you redeploy it, so estimate r conservatively. That is why the formula has r. For context, Eurostat reports that 5.2% of Romanian enterprises with 10 or more employees used AI in 2025, against 20.0% in the EU. Our article on AI adoption in Romanian SMEs goes deeper.

Worked example: supplier invoice processing

Every input below is an assumption to replace with your own measurements. A distributor handles 1,500 supplier invoices a month. A clerk spends 6 minutes per invoice. Loaded cost is €16 per hour, which is €0.267 per minute, so the baseline is €2,400 a month. Today 3% of invoices carry an error that costs 15 minutes of rework (€4.00) to fix. With the agent, a person reviews each proposed entry, and 2% of invoices still carry an error.

Model cost: 12,000 input tokens and 1,500 output tokens per invoice on Claude Sonnet 5.5, listed at $2 and $10 per million tokens on Anthropic's pricing page. That is 12,000 × $2 ÷ 1,000,000 + 1,500 × $10 ÷ 1,000,000 = $0.039 per invoice, or $58.50 a month, treating $1 as €1. Add €20 hosting, €120 maintenance (4 hours at €30) and €48 sampling (3 hours at €16). Run cost is €246.50 a month. Build cost is €4,000.

Scenario	Review minutes t1	Realised share r	Monthly benefit	Net monthly	Payback	12-month ROI
A, optimistic	2	100%	€1,660	€1,414	2.8 months	+324%
B, half the time redeployed	2	50%	€860	€614	6.5 months	+84%
C, heavy review	4	50%	€460	€214	18.7 months	−36%

Scenario A: labour benefit is 1,500 × (6 − 2) × €0.267 × 1 = €1,600, plus €60 from fewer errors (1,500 × 1% × €4), so €1,660. Net is €1,660 − €246.50 = €1,413.50, payback is €4,000 ÷ €1,413.50 = 2.8 months, and 12-month ROI is (12 × 1,413.50 − 4,000) ÷ 4,000 = 324%. Scenarios B and C change only t1 and r. Each review minute costs €400 a month at full realisation, so the break-even review time is 5.5 minutes at r = 100% and 5.1 minutes at r = 50%. The same agent gives +324% or −36% depending on two inputs, and neither is token price.

Which risks erase AI agent ROI?

Hallucination cost. Expected loss is V × probability of an undetected error × k. In 2024 a Canadian tribunal ordered Air Canada to pay about CA$812 after its chatbot gave wrong fare advice, finding the airline did not take reasonable care to ensure the chatbot was accurate. In 2025 Deloitte Australia agreed to partially refund a 440,000 Australian dollar government report that contained apparent AI-generated errors.
Review time. It is the swing factor above. If reviewers distrust the output they redo the work, and t1 drifts back toward t0.
Vendor price changes. Anthropic's pricing page records a Sonnet 5 rise planned for 1 September 2026 that was cancelled, and Google's page lists Gemini 3.8 Flash at $0.75/$3.75 through 31 December 2026 and $1.50/$7.50 from 1 January 2027. In the example, model cost is €58.50 of €246.50, so doubling it moves net by about €58.
Project failure. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of costs, unclear value or weak risk controls.

How do you run a small AI pilot with an evaluation set?

A one-page business case template needs these fields: workflow and owner, V, t0, c, t1, r, e0, e1 and k, m, F, M and O, build cost, the payback you require, the criteria that stop the project and a review date. Log every call's tokens, latency and outcome from day one, because those logs feed m, t1 and e1.

Pick one workflow with steady volume and a named owner.
Time the baseline: 30 to 50 real tasks with a stopwatch, plus the current error rate.
Build an evaluation set of 100 to 200 real, anonymised cases, including hard ones, labelled by two people and then frozen.
Run the agent in shadow mode for 2 to 4 weeks and score it. Report accuracy with an interval: 95 correct out of 100 gives a 95% Wilson interval of about 89% to 98%, and 190 of 200 narrows it to 91% to 97%.
Switch to assisted mode, with a human reviewing every output, and measure t1, the share of errors reviewers catch and token cost per task from logs.
Decide against thresholds agreed in advance: accuracy lower bound, t1 and net monthly benefit. Stop if they are not met.

Anthropic's guidance on building agents says to find the simplest solution possible, add complexity only when needed, and that agentic systems often trade latency and cost for performance.

When should you not automate?

Volume is too low for 12 × net monthly to beat the build cost.
A wrong output is costly or irreversible and you cannot put a human check in front of it.
The process changes every month, or the data is missing or messy.
A rule, a script or a template does the job. Gartner notes that many use cases labelled agentic do not need an agentic implementation.
Reviewing takes nearly as long as doing the task, so t1 is close to t0.

How Danube Labs helps

Danube Labs scopes AI work as a short feasibility study with an evaluation set, which our service pages price as an indicative band of €1,200–2,500, before any production build. See automation and IT and automation consulting, or book a free intro call and bring one workflow with its volume and timing.

Frequently asked questions

How do you calculate the ROI of AI agents?: Multiply tasks per month by minutes saved per task, loaded cost per minute and the share of time you actually redeploy. Add the value of fewer errors. Subtract tokens, hosting, maintenance and review. Divide net monthly benefit into build cost for payback. Replace every input with measured values from a pilot, not vendor claims.
How do I measure AI productivity in my team?: Time the same real tasks with and without AI, record error rates, and compare on a frozen evaluation set. Do not rely on self-reports. In METR's 2025 trial, developers expected to be 24% faster yet took 19% longer. Measure output, quality and review time, then check whether saved time was redeployed.
How long does an AI automation take to pay back?: It depends mostly on review time and how much saved time you redeploy. In our invoice example with a €4,000 build, payback is 2.8 months when review takes 2 minutes and all saved time counts, 6.5 months at half realisation, and 18.7 months with 4-minute reviews. All inputs are stated assumptions, not measurements.
How big should an AI evaluation set be?: For a first pilot, 100 to 200 real, anonymised cases including hard ones, labelled by two people and frozen. Size sets your precision: 95 correct out of 100 gives a 95% Wilson interval of about 89% to 98%, while 190 of 200 narrows it to about 91% to 97%. Add cases for failure types you find.
When is AI automation not worth it?: When volume is too low to repay the build, when mistakes are costly and cannot be checked by a person, when the process or data keeps changing, when a rule or script does the job, or when review takes nearly as long as the task. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027.