AI ROI for businesses: how to measure it, key indicators, and pitfalls

What is AI ROI and how is it calculated?

AI ROI (return on investment) measures the value created by adopting artificial intelligence, relative to the total cost incurred [13][27]. Concretely, it compares measurable gains (time saved, productivity, cost savings) with the full cost of the project: tools, setup, and training [15]. Another guide defines it in the same way, as the ratio of the gains generated by a solution to the costs incurred [26].

The standard formula

The basic formula is as follows [6][10]:

ROI (%) = [(Gains − Costs) ÷ Costs] × 100 [6]

On an annual basis, the calculation takes annual gains minus annual costs, divided by the initial investment [17]. A positive ROI means the value created exceeds the expense. Everything therefore depends on what is counted as "gains" and what is counted as "costs."

A worked example (commercial source)

Asana, a commercial software vendor, illustrates the calculation this way: for an AI investment of €20,000 (licenses, training, configuration), the ROI reaches 380% [6]. This example serves to explain the mechanism. It is not a result to expect in general.

What about classic financial methods?

Net present value (NPV), internal rate of return (IRR), and payback period also apply to AI projects. However, they require estimating future cash flows [21]. For a technology whose gains emerge gradually, this estimate is the main source of uncertainty.

Which costs should be included in the AI ROI calculation?

The cost to include in the AI ROI calculation is the total cost of the project, not just the price of the tools: it covers tools, setup, and training [15]. The Asana calculation example, from a software vendor, confirms this: the €20,000 investment combines licenses, training, and configuration, for an ROI of 380% [6]. Since ROI compares gains to this cost, any item left out of the denominator mechanically inflates the result.

Direct costs

• Tools: licenses and subscriptions [6][15].
• Setup: configuration and deployment [6][15].
• Training: supporting teams in using the tools [6][15].

Hidden costs

Hidden costs are those that do not appear in the initial budget and that contradict productivity-gain promises. Converteo explicitly contrasts these promises with "the reality of hidden costs" [28]. The available excerpt does not detail them: it is up to each company to identify them for its own scope, before launch rather than after.

The cost of inaction

Not deploying also has a cost. According to a guide on AI in human resources, this cost of inaction weighs on profitability, through lower productivity and higher operating expenses [33]. This source is an HR-oriented web content piece: it serves to raise the question, not to quantify the gap.

Tracking costs over time

When usage is spread across several tools, each subscription adds a line item to consolidate. Oupi groups usage into a single subscription, which simplifies cost tracking.

Which indicators should be tracked to measure AI productivity gains?

To measure AI productivity gains, you need to track indicators from several families: financial (additional revenue, cost savings), operational (time saved, reduced errors), and strategic [20]. The most frequently cited indicators are cost savings, revenue generation, and customer satisfaction [30]. Another framework rests on three pillars: business value creation, direct or indirect business impact, and revenue or margin growth [5].

Linking the model to business results

A useful indicator bridges the performance of the model (success rate, processing speed, input volume) and business results [29]. A high success rate proves nothing until it is converted into cost, time, or revenue. Tangible gains, such as time saved on repetitive tasks, are the easiest to translate into euros [32].

Example, from a commercial provider (Plug-Tech) and therefore not generalizable: an e-commerce client reduced its cost per interaction from €4.20 to €0.38 (−91%) after deploying an LLM for customer support [23]. The indicator divides the monthly cost of AI and its maintenance by the number of interactions [23].

Don't forget the cost of poor quality

Including the cost of poor quality changes the assessment. According to analyses by Tredence reported by Décision IA, companies that include it see a perceived value of their AI investments 15 to 25% higher than companies that focus only on productivity gains [37].

Tracking is already widespread: 86% of French companies regularly measure the impact of their AI use cases (Deloitte), and productivity and time saved are among the most tracked indicators [19].

How do you set up measurement before deploying AI?

Measuring ROI is prepared before deployment, in three steps: set the objective, establish a starting point, then compare quantified results.

1. Define the objective before the project. You need to know why you are launching the project (cost reduction, productivity improvement, etc.) before implementing it [34].
2. Establish a baseline. An audit conducted before the first projects provides a reliable reference point against which progress can be measured objectively. It avoids "gut-feeling" assessments based on impressions rather than data [31].
3. Measure results with figures. A practical book recommends launching a first project on the tasks best suited to AI, then measuring its results in quantified terms [2].

Distinguishing time horizons

Not all effects appear at the same time. At six months, you can track the impact on client deadlines, deliverable quality, and team satisfaction [24]. These indicators should be set at the outset, with their measurement dates.

Estimate upfront, with caution

Before the project, employees' self-reported estimates provide initial trends. You can also observe the results of comparable companies or consult studies by consulting firms such as McKinsey or BCG [16]. These are estimates: the baseline should replace them as soon as internal measurements are available.

Two methodological pitfalls

• An indicator must be valid. A study on AI adoption by Quebec companies questions the validity of the indicators used to measure the concepts studied [4]. A poorly constructed indicator produces a precise but misleading figure.
• Counting pilots is not enough. AI acculturation cannot be measured by the number of pilots launched [1].

On the cost side, Oupi groups usage into a single subscription, which makes it easier to track this expense line within the baseline.

What ROI do studies observe, and can they be trusted?

Published studies give very scattered results, from 2% of companies with a return to an average ROI of 340%, and none is sufficient to establish a reference ROI. Here is what the sources say, along with their nature.

These discrepancies stem from different methods, samples, and dates. The median ROIs (159% and 17%) do not measure the same thing as the shares of companies with "positive" results (75% and 82%), or as the share of companies with any return (2%). Figures from commercial players, such as the 340% over 12 months, should be read with caution. As for the 75% attributed to Wharton [8] and the University of Pennsylvania [34], they are reported by third-party sites, under two distinct formulations: the excerpts do not allow tracing back to the original study.

These figures therefore mainly show the dispersion of published results. They do not provide a benchmark that can be directly transposed to a given company.

Why do so many companies struggle to prove the ROI of their AI?

Companies struggle to prove the ROI of their AI because usage is advancing faster than their ability to link it to a financial result.

A gap between adoption and financial effect. In its 2026 global survey (1,719 respondents, 97 countries), McKinsey observes that nearly nine out of ten organizations regularly use AI in at least one function. 44% say they are scaling up, but the share attributing an effect to EBIT remains stable at 37% [18]. These figures are reported here by a specialized site, AI Agent, and were not read in the survey itself.

A poorly framed measurement from the start. According to electe.net, only 16% of companies manage to deploy AI at scale, largely because of erroneous measurement at the task level [32]. A gain observed on an isolated task does not yet show what the company gains overall.

Returns on generative AI remain rare. According to a KPMG study from late November 2025, reported by Genia, only 2% of Canadian companies would see a return on their generative AI investments [22]. This scope is Canadian and limited to generative AI, and it does not transfer directly to the French market.

Organizational prerequisites. A research paper on SMEs concludes that AI integrates successfully only if the company has already reached a certain level of digital maturity and is ready for organizational change [3]. A disappointing ROI may therefore stem from these conditions, not only from the tool.

A management science publication also notes that AI acculturation cannot be measured by the number of pilots [1]. It describes the move from experimentation to software integrating AI, industrialized and steered by ROI [1]. For a finance department, this means setting indicators before launch: moving from isolated pilots to ROI-driven industrialization is only possible once measurement is already in place.