How to Measure AI ROI: Formula and Concrete KPIs

Every time you evaluate an AI project, there's really only one question that matters: what does it return compared to what it costs? It's not a trivial question, because AI has an inconvenient trait. Costs are immediate and visible (setup, licenses, consulting), while the benefits are often diffuse, indirect, and spread out over time. That's why so many companies start out excited and then end up with no numbers to show management.

In this guide you'll find an explicit formula, the KPIs to track from day one, and a way to estimate payback (the time it takes to recoup the investment) honestly, without inflating the results. This is one of the operational pieces we cover in our complete guide to AI consulting for businesses; here we go deep on ROI alone. If you're still deciding what to automate, it's worth reading which processes are worth automating with AI first, because the return depends largely on choosing the right process.

One premise that will save you time and money: roughly 85% of generative AI pilot projects fail to make it into production. Measuring ROI properly isn't just a bookkeeping exercise — it's how you figure out in advance which projects to keep and which to stop. If you want the root causes behind these failures, we cover them in why AI projects fail.

The AI ROI formula, explained line by line

Let's start from the basic return-on-investment formula, valid for any project, and adapt it to AI:

ROI (%) = (Annual net benefit / Total investment) × 100

The tricky part is the numerator: the net benefit. In AI it's made up of three items that need to be kept separate and added together carefully:

Net benefit = (Hours freed up × Hourly cost) + Extra revenue − Recurring costs

Let's look at each one, because each hides a trap.

1. Hours freed up × hourly cost (the efficiency saving)

This is the most tangible component and the easiest to defend in front of a CFO. If an AI agent reads incoming emails, extracts the data, enters it into the CRM, and drafts a reply, it's freeing up hours of human work.

The calculation is simple: how many hours per week are freed up, multiplied by the person's real hourly cost. Careful, though: the real hourly cost isn't net salary divided by hours worked. You need to use the fully loaded company cost, which includes social contributions, severance provisions, the extra month's pay, and other charges. As a rule of thumb, the fully loaded cost runs about 1.5-1.8 times the gross annual salary. An employee on a €30,000 gross salary costs the company roughly €45,000-52,000 a year, so their real hourly cost over about 1,600 hours worked lands around €28-32, not €15.

The real trap is elsewhere: freed-up hours only have value if they're actually reallocated. If you free up 5 hours a week for someone but that person just does the same tasks at a more relaxed pace, the saving is only theoretical. ROI becomes real when those hours turn into more clients handled, more quotes issued, or one fewer hire needed during a growth phase.

2. Extra revenue (the commercial upside)

This component is harder to isolate, but it's often the biggest one. A few concrete examples:

• An automated sales follow-up system that recontacts leads within 5 minutes instead of within a day, raising the conversion rate.
• Reactivating dormant customers in the database, generating sales from contacts who already paid you in the past.
• An AI assistant on the website that answers customer questions after hours and captures inquiries that would otherwise be lost.

The trap here is attribution. Not all the extra revenue is thanks to AI. To be honest, estimate only the incremental share — the part you wouldn't have gotten without the system. A clean way to do this is comparing a period with and one without it (or a test group against a control group) and measuring the difference.

3. Recurring costs (the item almost everyone underestimates)

This is where the gap between a real ROI and a fantasy one hides. Costs aren't just the initial setup. The typical recurring items are these:

• Licenses and APIs: cost of the AI models (per token or subscription), platforms, subscriptions.
• Maintenance: no system stays stable on its own. There are prompts to update, integrations that break, edge cases to handle.
• Model drift and quality: a model's performance degrades over time, or when the input data changes. It needs monitoring and correction.
• Human oversight: the time spent by whoever checks outputs, handles exceptions, and maintains the guardrails.
• Training: people need to know how to use the tools. Only 22% of companies have structured employee AI training programs, which is one reason projects underperform.

If you want a transparent breakdown of the expense items for a typical scenario, we've detailed the numbers in how much it costs to automate business processes.

A complete numerical example

Let's take a realistic case: an SMB automating lead qualification and first contact for incoming leads. Here's how the calculation is built, with plausible figures.

At this point two numbers matter:

• First-year ROI = (16,800 / 12,000) × 100 = 140%. If you also count the setup as a first-year cost, the actual net benefit drops to 16,800 − 12,000 = €4,800, and the net first-year ROI is roughly 40%. From the second year on, the setup no longer weighs on the calculation, so annual ROI climbs.
• Payback = Investment / monthly net benefit = 12,000 / 1,400 = roughly 8.5 months.

This example falls within the realistic range: a well-chosen AI project has a payback of between 4 and 12 months. If your numbers show a payback of 2 months, you've probably underestimated the recurring costs. If they show more than 18 months, the project is fragile and at high risk of being abandoned before it pays for itself.

The KPIs to track from day one

ROI is the final number, but you can't calculate it if you haven't measured the operational KPIs along the way. The classic mistake is launching the project and then trying to reconstruct the numbers afterward. That doesn't work: without a baseline (the situation before AI), you have nothing to compare against.

First rule: measure the "before" before you start. Record how long tasks take today, how many errors there are today, what conversion rate you have today. Without a baseline, any later result is just an opinion.

Useful KPIs fall into three families.

Efficiency KPIs (measure hours freed up)

• Average time per task: minutes to handle a case, a reply, a quote, before and after.
• Volume handled per person: how many cases, tickets, or leads one person can handle in the same amount of time.
• Automation rate: percentage of cases resolved by AI without human intervention (useful for automated customer care).
• Response time: how long the system takes to respond to or react to an event.

Quality KPIs (measure whether the AI is doing a good job)

• Error/hallucination rate: percentage of wrong or fabricated outputs that need correction.
• Escalation rate: how often a case needs to be handed off to a human.
• Satisfaction (CSAT): the perception of customers and internal users.
• Internal adoption rate: how many people actually use the tool. A great tool nobody uses has zero ROI.

Business KPIs (measure the extra revenue)

• Conversion rate for leads worked by AI versus those worked manually.
• Cost per lead or cost per acquisition, which should go down.
• Incremental revenue attributable to the system (measured with a control group where possible).
• Retention or reactivation rate of customers.

You don't need all of them. Pick 2-3 per family, the ones most tied to the specific project's goal, and track them consistently. Better to have a few well-measured KPIs than a dashboard full of numbers nobody looks at.