AI ROI: How to Measure It in Your Company

AI ROI is the ratio between the measurable benefits of an AI investment and its total costs. The formula is simple: (benefits minus costs) divided by costs. The problem is not the calculation but what goes into the numerator and how to prove it.

For this reason, AI ROI should be built per well-defined process, with a measurement before and one after. A single number for the whole company usually tells a more convenient story than a true one.

Why AI ROI is hard to measure

Four obstacles recur almost every time.

• Diffuse benefits. AI helps with many small activities, each with a modest gain that is hard to isolate.
• No starting measurement. Often nobody had timed the process beforehand. After the tool is introduced, comparison is impossible.
• Time saved is not value captured. If a person saves an hour, the economic benefit depends on what they do with it: more work, better work, or nothing detectable.
• Underestimated costs. Beyond licenses, training, integration, time spent reviewing outputs and change management all count.

On top of this comes a framing error common in investment plans: asking "what does AI return?" as if it were one thing. AI is a set of tools used in dozens of different ways.

The three levels of measurement

It helps to keep three levels separate. They answer different questions and should not be summed into a single indicator.

Level 1: real adoption in processes

Question: is the tool really used, by whom and for what? It is measured with usage data by department, role and use case, complemented by surveys. It is the starting condition: without adoption there is no impact. The method is described in what AI adoption is and how to measure it.

Level 2: impact on processes

Question: does the process work better than before? It is measured with operational indicators (time, volumes, errors, quality) compared against a baseline. It works well only on well-defined processes, where the effect of AI can be isolated from other variables.

Level 3: impact on people

Question: how does the work change? It is read through surveys and sentiment indicators, always in aggregate by role or department, never as an individual evaluation. It includes perceived workload, satisfaction and trust in the tool.

An excellent result at level 2 with a deteriorating level 3 is a signal not to ignore, because a gain obtained at the expense of people rarely holds over time.

Before-and-after baseline on well-defined processes

The most solid method is also the simplest to describe: measure the process before, introduce AI, measure it after.

1. Choose a well-defined process. One with a clear start and end, repeated volumes and an identifiable owner. For example, preparing replies to standard requests or the first review of recurring documents.
2. Define the indicators. Few of them: time per unit, volume handled, error or rework rate, review time.
3. Capture the baseline. For a period long enough to cover normal variability, before changing anything.
4. Introduce the tool with training and clear rules of use.
5. Measure again once it is in steady use, not in the first weeks when use is still unstable.
6. Compare and document the conditions, to distinguish the effect of AI from seasonality, staffing changes or volume changes.
7. Calculate the process ROI, including all the costs of that use case.

Where possible, a comparison group that does not use the tool makes the estimate more credible. It is not always practicable, and in that case the limits should be declared.

Sensible indicators

Indicators should be chosen on the process, not on the tool. Some useful families:

• Cycle time per unit of work, or average handling time.
• Volume handled with the same resources.
• Quality: errors, rework, corrections requested in review.
• Review time of AI outputs, often overlooked yet decisive for the net benefit.
• Satisfaction and perceived workload of the team, in aggregate.
• Total cost of the use case: licenses, integration, training, time spent.

A hypothetical example: a team handles 400 requests a month with an average time of 30 minutes. After the tool is introduced, average time falls to 22 minutes, but each reply needs 4 extra minutes of review. The net gain is 4 minutes per request, not 8. If quality is stable, the benefit is real but smaller than writing time alone suggests. The numbers are invented for illustration.

From time saved to value

Converting minutes into euros requires caution. Multiplying hours saved by hourly cost is a widespread but optimistic calculation, because it assumes every freed hour is reused productively. It is more prudent to reason in scenarios.

• Freed capacity: more volume handled without new hires.
• Quality: fewer errors and less rework, with an estimable avoided cost.
• Speed: shorter response times, with an effect on customers or other processes.
• Reduced cost: only if actually reallocated or avoided.

Declaring which of these scenarios you assume makes the number open to useful challenge, which is better than a flawless number that cannot be verified. It also helps to set the expected impact in advance for each use case: comparing expected and realized is the most honest way to judge whether the initial estimate was reasonable and to correct later ones.

The cost side

ROI also has a denominator. License spend is only part of total cost. It is worth keeping spend under control by tool and by seat, and identifying unused licenses and overlaps between tools, because they directly affect the result. The method is in how to reduce AI license spend. A profitable use case on an oversized tool can end up in the red.

Honest limits

Declaring the limits makes the measurement more credible.

• Causality. Even with a baseline, other factors may explain part of the change.
• Learning effects. Performance changes as people learn to use the tool: measuring too early understates the benefit.
• Unmeasurable benefits. Some advantages, such as removing frustrating workloads, are captured through surveys and remain subjective.
• Not everything can be measured precisely. For creative or non-repetitive activities the baseline is fragile and indirect indicators have to do.
• Generalization. The ROI of one process does not automatically transfer to another.

The result is a set of estimates per process, with ranges and explicit assumptions, more useful than an apparently precise aggregate. To organize the use cases to measure, an AI use case register can help. Presenting leadership with a set of estimates and declared assumptions, rather than a single percentage return, reduces the risk of expectations that the next measurement cannot confirm.

Where Delphis fits

Delphis is a vendor-independent platform that connects spend, adoption, impact and regulatory obligations. For ROI, it compares expected and realized impact on use cases, with before-and-after measures on well-defined processes, alongside aggregated indicators on people and the cost of the use case. See the adoption and ROI page.