How much is an AI agent worth: total cost, KPIs, and return on investment

An AI agent can draft responses in seconds, check documents, and execute steps in a process. For a small or medium business, however, the speed shown in demos doesn’t reveal the true value of the investment. What matters is which results are accepted, how much work remains for people, and what expenses or revenues actually change.

Saving ten hours a week can be valuable even if it doesn’t immediately reduce any bills. It might allow the team to respond faster or handle more requests. But for an investment decision, you need to distinguish between freed-up capacity, cash savings, and the extra margin gained.

In the article on choosing a platform, we discussed implementation options and their associated costs. Here we build a framework for measuring value: the initial benchmark, total cost, performance indicators, and a thorough profitability calculation.

Sources checked as of September 27, 2026. Percentages from studies refer to the tested contexts. Figures in the financial example are hypothetical and are not drawn from research results.

Four terms for the same decision

TCO, Total Cost of Ownership, is the overall cost of owning and operating a solution over a specific period. For an agent, this might include implementation, integration, subscriptions, usage, infrastructure, verification work, maintenance, and switching vendors. The time frame should be chosen before comparing solutions: a large upfront cost looks very different in a three-month versus a three-year calculation.

KPI, Key Performance Indicator, means key performance indicator. A useful KPI shows whether the process achieves the desired outcome—for example, the number of requests resolved correctly, human work minutes, or cost per accepted case. The number of generated messages might describe activity but doesn’t measure its value.

ROI, Return on Investment, expresses the ratio between net benefit and the baseline cost. Since different conventions are used in practice, the period, included benefits, and the denominator need to be specified. In our example, we’ll calculate the first-year cash ROI relative to all cash costs of the project in that year.

Payback period shows when cumulative net benefits cover the initial investment. On its own, it doesn’t indicate how good the investment is after that moment and does not replace analysis of quality or risk.

We start with a verifiable result

“The agent is done” isn’t a sufficient definition of success. The result must be described in the terms of the company’s process.

Let’s take a distributor receiving inquiries about products. The agent classifies the message, searches the catalog and approved commercial terms for information, drafts a reply, and sends it to a colleague for validation. It doesn’t change prices or make promises about deadlines the system hasn’t confirmed.

For this process, an accepted outcome might mean the product is identified correctly, the information is current, the reply complies with company policy, the request is handled on time, and the necessary interventions are recorded. A fluent message with an incorrect price fails the test. Nor is it a success if a case is closed automatically without actually being resolved.

We separate results accepted without rework from those corrected by humans. Both can have value, but incur different costs. Escalating an exception to the right person can be the agent’s correct behavior; nonetheless, in the dashboard, this should be distinguished from final resolution of the request.

This focus on results aligns with the FinOps Foundation, which differentiates between technical unit costs—like cost per token—and business-related costs, such as cost per resolved case. A token is a unit of text or other content processed by the model. It's useful for billing and optimization, but doesn’t directly represent a resolved request.

Measure the current process before promising savings

The initial benchmark, also called the baseline, describes how the process works without the new solution. For the selected period, record the volume, case complexity, active human time, rework, errors, and costs. Include tough cases and incomplete ones, not just the examples that look good in a demo.

It’s useful to distinguish between two timeframes. Active time shows how much time a person actually spends on a case. Time to close also includes waiting for a colleague, supplier, or client. The agent may shorten one timeframe without affecting the other. For instance, drafting may get faster, but approval remains a bottleneck.

The comparison must use the same acceptance criteria and cases of comparable difficulty. Where feasible, random assignment of cases between options or a control group helps separate the solution’s effect from volume changes and seasonality. If you have only a “before-and-after” comparison, present the result as an estimate and explain what else changed between periods.

Compare it to a simple automation too. A fixed rule, a better form, or a connection between two apps might solve the problem at lower cost. The seven processes well-suited for a first pilot offer starting points for choosing a sufficiently narrow domain.

The total cost also includes the work surrounding the agent

For an initial estimate, we recommend four categories. The list adapts to the process—it doesn’t assume every project has all possible costs.

1. Initial costs

Process inventory, data preparation, integration with applications, access setup, building tests, and team training all take time. Transition and the period when the new process runs alongside the old one should also be considered. If you use internal hours, decide how to assess them and keep actual payments separate from internal cost allocations.

2. Ongoing technology costs

Subscriptions and model usage are just one part. There may also be costs for search, extracting text from documents, databases, storage, monitoring, and hosting. API is the interface through which the agent communicates with another system; some services charge for calls or operations made through it.

Claude’s pricing documentation, for example, separates model input and output, cache operations, and certain tools. The cache allows reuse of already processed information, as per the provider's conditions. A retry, extra check, or a new step in the agent can change the overall consumption. For budgeting, use measured executions and the rates of your actual contract, not the price of a single isolated call.

For a local installation, costs include hardware, energy, operations, and available capacity. An already purchased server may reduce the project's initial outlay, but administrator time and consumed resources remain relevant.

3. Verification, correction, and exceptions

Measure time spent reading and approving, documents redone, manually handled cases, and error investigation. If an agent saves four minutes on drafting but adds five minutes to checking, the net result is unfavorable for that task.

Include unsuccessful attempts too. The cost of a process doesn’t disappear just because the result is rejected. A human verification step may be warranted by risk, even if it reduces time savings. Permission and audit rules must be considered as part of the solution.

4. Maintenance and change

Data sources change, models are updated, and applications change interfaces. Plan for reevaluation, fixing integrations, and support. For a longer timeline, factor in data export or migration as well. Don’t double-count a service already included in your subscription, and document how you share common costs across processes.

Six indicators to support decision making

The following set is a proposal for a pilot, not a mandatory standard. Definitions and the responsible person for each indicator are set in advance of the test.

1. Percentage of cases completed correctly and accepted. Count accepted results out of all eligible cases that entered the pilot. Failures, abandoned cases, and those not validated on time remain visible in the record. At the end, distinguish ongoing cases from those whose deadline has expired.
2. Percentage of results accepted without correction. Shows how many cases passed validation without human redo, reported as a share of all eligible cases in the pilot. A result checked by a colleague may count here; checking and correction are different activities. Results recovered through intervention are reported separately.
3. Minutes of human work per case. Add up time spent on verification, corrections, and handling exceptions for all eligible cases, divided by their number. Agent run time alone does not equal employee labor time.
4. Total cost per accepted result. Divide the costs attributed to the process by the number of accepted results in the same period. Costs include failed attempts, too. If implementation is partly allocated, explain the calculation. If there are no accepted results, the unit cost is not usefully reported as zero.
5. Time to closure. Track the median (the middle value of the distribution) and the 95th percentile, the duration within which about 95% of measured cases are closed. Keep ongoing case ages separate so the slowest don’t disappear from the report.
6. Errors and their consequences. Distinguish a vague formulation from an incorrect price sent to the client, or from an unauthorized action. A good average rate does not offset a single high-impact incident.

A useful formula is:

Cost per accepted result = total cost attributed to the period / number of accepted results for that period.

When comparing two options, use the same scope and the same definition of result. Otherwise, the version that excludes review work may seem artificially cheaper.

Anthropic’s guide to agent evaluation, published on January 9, 2026, recommends verifying results and combining evaluation methods. For an SME, the practical application is simple: don’t let the agent mark its own success. Use evidence from operating systems and human checks appropriate to risk.

Freed-up hours, savings, and revenue are different things

Released capacity means people have available time. If salaries and other payments stay the same, the budget doesn't decrease just because some tasks take less time. The benefit may be better service, extra volume, or less pressure on the team.

Cash savings occur when the company actually avoids making payments: for example, overtime or outsourced services that are no longer needed. These savings need to be compared against the extra costs of the solution. The distinction between spending reduction and available capacity also appears in the Government Efficiency Framework. Here, we use the conceptual distinction, not the UK public sector’s reporting rules.

Additional revenue should be analyzed based on the contribution left after the extra activity's costs. If freed-up time enables more sales, we don’t automatically attribute all revenue to the agent nor do we ignore the cost of goods or delivery. It also needs to be shown that there was demand and that the process change contributed to the result.

We don’t add the estimated wage value of the same hours to the savings or margin gained by reusing them. That would mean double-counting the same benefit.

A complete calculation, with explicit assumptions

Let’s return to the distributor’s request process. All the following figures are hypothetical. They are not an i8.ro outcome, a market price, or a performance guarantee. We assume the volume and quality remain comparable, and that the process passes acceptance criteria.

From minutes to capacity

The company receives 1,200 requests per month. The current process takes an average of 10 minutes of human work per request, meaning 200 hours per month. After introducing the agent, the average drops to four minutes, including checks and exceptions, calculated across all eligible requests. That results in 80 hours per month and a released capacity of 120 hours.

At a reference labor cost of 60 lei per hour, the equivalent of 120 hours is 7,200 lei. This assessment describes capacity, not proof that company payments will decrease by 7,200 lei.

From capacity to cashflow

In the central scenario, we assume that 60 of the released hours replace outsourced services previously paid at 60 lei per hour. The company can actually stop these expenses, so it avoids 3,600 lei per month. The remaining 60 hours release internal capacity, without immediate payroll reduction.

Implementation costs 18,000 lei, paid upfront. Additional operating costs are 2,400 lei per month and include all new recurring payments assumed in the example. Remaining internal work and ongoing salaries are included in the operational analysis and TCO but are not presented as cash savings.

The monthly net cash benefit is:

3,600 − 2,400 = 1,200 lei.

If the benefits start immediately and remain steady, simple implementation payback is:

18,000 / 1,200 = 15 months.

In the first year, we have benefits of 3,600 × 12 = 43,200 lei and costs of 18,000 + 2,400 × 12 = 46,800 lei. The convention used here is:

First-year cash ROI = (first-year benefits − all first-year project costs) / all first-year project costs × 100.

The result is −7,7%. The project generates a positive monthly benefit after recurring costs, but does not recover the initial investment within the first 12 months. The two conclusions are compatible.

The calculation is simplified, with no financial discounting and no tax effects. It assumes no other costs or benefits. If adoption is gradual, benefits will appear later or if implementation exceeds the budget, payback will take longer. For large or multi-year investments, the timing of cash flows and the cost of financing should also be considered.

Three scenarios for the same decision

We keep the implementation at 18,000 lei and the extra monthly cost at 2,400 lei. The only variable is how much of the freed-up capacity can actually avoid external payments, with all other assumptions unchanged:

• Cautious: 30 external hours avoided, i.e., 1,800 lei per month. The difference compared to operating cost is negative, −600 lei per month. The investment is not recovered through this constant cash flow; first year ROI is approximately −53.8%.
• Central: 60 external hours avoided, i.e., 3,600 lei per month. The net benefit is 1,200 lei per month, simple payback takes 15 months, and first year ROI is approximately −7.7%.
• Favorable: 90 external hours avoided, i.e., 5,400 lei per month. The net benefit is 3,000 lei per month, simple payback takes six months, and first year ROI is approximately 38.5%.

The scenarios show how sensitive the calculation is to a business assumption: how much spending can actually be avoided. It's not enough to improve the model if the company can't capitalize on the freed-up time. A project may still have operational reasons to continue—such as faster response times—but these must be explicitly stated.

What the research tells us—and what it can't tell us

In Generative AI at Work, published in 2025 in the Quarterly Journal of Economics, the authors analyze the introduction of an assistant for 5,172 support staff. The number of issues resolved per hour increased on average by 15%, with different effects depending on experience. People controlled the conversations. The study does not measure the ROI of an autonomous agent and does not allow the percentage to be converted into salary savings for any company.

Navigating the Jagged Technological Frontier, published in Organization Science in March 2026, draws on a 2023 experiment with 758 consultants. For tasks suited to the model, participants with AI completed more tasks and worked faster. On a difficult problem outside the tested capabilities, accuracy dropped. The key takeaway is that task choice influences outcomes.

In a METR experiment published in July 2025, 16 experienced developers worked on 246 real-world tasks. Tools from early 2025 increased time-on-task by 19% in the test context. The METR update from February 2026 explains why later data does not reliably estimate the current effect, including due to participant selection and working with multiple agents at once. We do not use the old result as a general statement about programmers in 2026.

These results support the need for a local pilot, not a universal profitability percentage. We measure the company's process with its own tools, people, and rules.

When to continue, narrow, or stop

Before the pilot, the process owner and the person monitoring the budget set the decision criteria. Quality and authorization of actions have their own requirements; we do not automatically offset these with an average time saving.

Continue or expand when results meet criteria, costs are explainable, and benefits repeat across enough comparable cases. Expansion should be gradual, to observe whether exception rates or supervision work change.

Narrow when value appears only in one category: for example, product information requests but not negotiated-offer cases. An agent that's useful in a small domain may be a better investment than one trying to cover the whole process.

Stop or redesign when checks consume more than they save, errors exceed acceptable limits, or there is no credible mechanism to capitalize on the freed-up capacity. High usage volume alone is not a reason to continue.

A good decision can also be to keep an assistant with human approval or return to simpler automation. The goal is a more efficient, auditable process, within the company's economic constraints.

Main sources

• FinOps Foundation, Unit Economics, accessed on 27 September 2026.
• Claude Platform, Pricing, accessed on 27 September 2026.
• Anthropic, Demystifying evals for AI agents, 9 January 2026.
• UK Government, The Government Efficiency Framework, accessed on 27 September 2026.
• Brynjolfsson, Li, and Raymond, Generative AI at Work, published online on February 4, 2025.
• Dell’Acqua et al., Navigating the Jagged Technological Frontier, published on March 11, 2026.
• METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 10, 2025, and methodology update, February 24, 2026.