How to Calculate the ROI of an AI Agent: the complete method (formula, examples, metrics)

Introduction

Two figures, published the same year, seem to contradict each other. According to McKinsey, 87% of companies see a positive ROI within 6 months on their well-run AI projects. But a Gartner survey published in April 2026, covering 782 IT and operations directors, found that only 28% of AI projects fully meet their ROI targets, and 20% fail outright. MIT Sloan Management Review goes further: 95% of AI projects reportedly fail due to a lack of strategy and proper training.

These figures don't really contradict each other: they measure different things. A project can show a positive ROI on paper while never reaching the scale that was promised, simply because no one defined upfront what to measure and how. That's exactly what this guide is for: providing a rigorous method to calculate the ROI of an AI agent, before, during, and after its deployment.

‍

The basic formula for AI agent ROI

The standard method for calculating the ROI of an AI project relies on a simple formula:

ROI (%) = [(Total gains − Total costs) / Total costs] × 100

An ROI of 150% means that for every euro invested in the AI agent, the company generates €1.50 in net benefit, on top of the initial outlay. Simple on paper, this formula only holds if the "gains" and "costs" it includes are themselves complete and properly valued — this is where most rough calculations fall apart.

‍

The gains to include in the calculation

Time freed up for people

This is the easiest gain to put a value on: the number of hours freed up per week or month on a given task, multiplied by the fully loaded hourly rate of the role involved. An AI agent that frees up 15 hours a week on a task billed at €45 of fully loaded hourly cost represents roughly €35,100 in valued annual gain.

‍

Reduced error rate

Every error avoided carries its own cost: correction time, delays generated, sometimes a direct customer impact. This gain is often underestimated because it's harder to value precisely than simple time savings, but it frequently weighs more heavily over time than initially expected.

‍

Faster processing cycles

A case processed faster often generates an indirect commercial gain: customer satisfaction, conversion rate, shorter payment delays. This speed factor can translate directly into the ability to seize opportunities that a slow manual process would otherwise miss.

‍

Additional revenue generated

In some cases, an AI agent doesn't just cut costs: it directly generates new revenue (better prioritization of sales leads, faster responsiveness to customers, 24/7 availability on a task previously limited to office hours). This gain is generally the hardest to isolate with certainty, and should be estimated cautiously rather than promised as a guaranteed figure.

‍

The costs to include in the calculation (often forgotten)

• Scoping and development cost ("Build"): the initial investment to design and put the first agent into production.
• Usage-based AI model cost: generally billed by volume processed, a variable cost that grows as the agent scales up.
• Supervision and adjustment cost ("Run"): often 15% to 30% of the initial development cost, every year, to monitor and adjust the agent over time.
• Training and change management cost: training teams to work with the agent, defining who approves what, adjusting existing processes.
• The hidden cost of "shadow AI": when employees use unapproved AI tools alongside an official project, it fragments data, exposes the company to security risks, and skews the measurement of the actual ROI of the managed project.

Our guide on the price of an AI agent in the enterprise covers the observed cost ranges by project complexity in more detail.

‍

A complete worked example

Take the example of a document-matching AI agent deployed to automate supplier invoice checking:

• Development cost (Build): €25,000
• Annual supervision cost (Run): €5,000 (20% of the initial cost)
• Usage-based AI model cost: €3,000 per year
• Total cost in year one: €33,000
• Time freed up: 12 hours per week, at €40 of fully loaded hourly cost, roughly €24,960 per year
• Errors and supplier disputes avoided: estimated at €8,000 per year
• Total gains: €32,960 in year one

ROI = [(32,960 − 33,000) / 33,000] × 100 ≈ −0.1% in year one — a project that looks break-even over its first twelve months, but whose following years benefit from an already-amortized development cost: from year two onward, with an annual cost reduced to €8,000 (Run + usage) against stable gains of around €33,000, ROI exceeds 300%. This example illustrates an essential point: the ROI of an AI agent is rarely read correctly from the first year alone.

‍

The payback period

Alongside the ROI percentage, a second metric helps with decision-making: the payback period.

Payback (in months) = Initial investment / Average monthly gains

In the example above, with an initial investment of €25,000 and average monthly gains of roughly €2,747, the payback period comes to just over 9 months — a figure consistent with the McKinsey benchmark mentioned in the introduction (positive ROI observed within 6 months for well-run projects).

‍

Why so many AI projects fail to hit their ROI target

According to the Gartner survey cited in the introduction, 57% of executives who experienced an AI project failure attribute it to expectations set too high, too quickly. Three structural causes come up consistently:

• An overly ambitious initial scope, which spreads effort across too many use cases at once instead of securing one measurable first win.
• No baseline measurement before the project launches: without an initial point of comparison, it becomes impossible to objectively demonstrate the gain achieved.
• Success metrics defined after the fact, rather than before deployment, which opens the door to a biased or contestable evaluation.

The golden rule, widely shared by practitioners in the field: define success metrics and measure a baseline before launching the project, not after. The ROI of an AI agent isn't guessed retrospectively — it's measured from the design stage of the project onward, which is exactly what the scoping phase covered in our comprehensive guide on AI agents in the enterprise is for.

‍

How to make your ROI calculation reliable from the start

1. Measure the current state before any deployment: how long the task takes today, at what error rate, at what cost.
2. Choose a first use case with high volume and quickly measurable impact, rather than an ambitious project that's hard to isolate statistically. We recommend starting from our list of the 15 AI agents to deploy first to identify this type of use case.
3. Include all costs, including Run, rather than only comparing the initial development cost to the gains achieved.
4. Track ROI over time, not just at launch, since an AI agent keeps generating both costs and gains after its initial deployment.
5. Stay cautious about indirect gains (additional revenue, customer satisfaction), which are real but harder to isolate than time savings or error reduction.

‍

Formulas to remember

ROI — Formula: [(Total gains − Total costs) / Total costs] × 100. Purpose: measure the overall profitability of the project.

Valued time savings — Formula: Hours freed up × Fully loaded hourly rate. Purpose: isolate the "time" component of the total gain.

Payback period — Formula: Initial investment / Average monthly gains. Purpose: know how many months it takes for the project to become profitable.

Annual Run cost — Formula: 15% to 30% of the initial development cost. Purpose: anticipate the ongoing operating cost over time.

‍

FAQ

What counts as a good ROI for an AI agent project?

There's no universal threshold, but a positive ROI within the first or second year, with a payback period under 12 months, is generally seen as a sign of a well-scoped project.

‍

Should the time spent by internal teams be included in the cost calculation?

Yes, the time internal teams spend on scoping, testing, and training represents a real cost, even if it never appears on an invoice, and should be valued just like external costs.

‍

Why is the ROI of an AI agent often negative or flat in year one?

Because the initial development cost (Build) is generally concentrated in the first few months, while the gains materialize gradually. ROI often becomes much more favorable from year two onward, once the initial investment has been amortized.

‍

How do you measure the ROI of an AI agent when the gains are hard to isolate?

By focusing first on the gains that are easiest to measure objectively (time freed up, errors avoided), and treating more indirect gains (customer satisfaction, additional revenue) as supplementary indicators rather than the main basis for the calculation.

‍

Does an AI agent's ROI need to be recalculated after deployment?

Yes. An AI agent keeps generating costs (supervision, usage) and gains after going into production; tracking ROI over time, not just at launch, is necessary to properly manage the project.

‍

Take action

A convincing ROI isn't calculated after the fact: it's built from the moment you choose the first use case and define the metrics to track. That's exactly what the scoping workshop we run at Qolaig is for, before any AI agent deployment — so that calculating ROI is never an open question after going into production. Let's talk about your project.