AI ROI: how to calculate it properly

Calculating the ROI of artificial intelligence takes mathematical rigour: (annualised net benefit − total cost of ownership) ÷ total cost of ownership × 100. The danger sits in the denominator. Ignoring integration, data governance, maintenance and model evolution turns any projection into a sterile figure that will not survive serious budget scrutiny.

In short

• The formula is simple; the hard part is not inflating the numerator or trimming the denominator.
• Hours freed are only a benefit if they are reassigned to billable work or if they absorb growth without hiring.
• The correct horizon is 24 to 36 months, with the first year penalised by the ramp-up.
• The RAO (Operational Autonomy Ratio) is the variable that moves ROI: without real autonomy, there is no return.
• An honest calculation includes the pessimistic scenario and the point at which the project would stop being justified.

Which ROI formula actually works for an AI project?

ROI (%) = (Annualised net benefit − Annualised total cost of ownership) ÷ Annualised total cost of ownership × 100

And alongside it, the metric a committee really looks at:

Payback (months) = Initial investment ÷ Monthly net benefit once in production

Both are calculated per process, never per company. An ecosystem with three superagents has three different ROIs, and aggregating them hides the fact that one of them probably does not stack up.

What counts as benefit and what gets overcounted?

Four legitimate sources of benefit:

1. Hours freed, valued at the fully loaded hourly cost, and only if those hours are reassigned to value-adding activity or avoid a planned hire.
2. Errors avoided: credit notes, late delivery penalties, mis-shipped orders, complaints. Estimated from history, not from intuition.
3. Attributable additional revenue: more quotes sent, shorter commercial response time, lower drop-off. It requires a comparison group or a clean before and after.
4. Capacity absorbed: growing 20% in volume without adding headcount is a benefit, even if it does not show up as a saving.

And three things that get overcounted far too often: hours freed that nobody reassigns, "satisfaction" improvements with no metric, and revenue attributed to AI when the sales team also changed strategy that quarter.

What goes into the total cost of ownership?

The five line items of the cost of implementing AI: context, integration, governance, maintenance and models. To those you have to add a sixth almost nobody budgets and which is very real: your own people's time during implementation. The people who know the process have to spend hours explaining it, validating answers and correcting. On a three-month project, that internal effort is a significant line item.

What time horizon should you use?

24 to 36 months. A 12-month horizon unfairly penalises AI because it concentrates the whole initial investment in a period when the system is only in production part of the time. A five-year horizon unfairly favours it, because it assumes a technological stability that does not exist.

The rule we apply at NEXTAI: 36 months for ROI, with the first year calculated on the real months in production, and a recalculation at 6 and 12 months with real data instead of estimates.

A worked example: a tax and accounting firm

A firm of 45 employees processing clients' accounting documentation, in southern Europe. The process chosen: intake, classification, extraction and posting of purchase invoices.

Starting point

After the finance superagent

The agent extracts, validates against each client's chart of accounts and posts. It reaches a RAO of 72%: almost three in four documents complete with no intervention. The remaining 28% (1,820 documents) goes to human review, at 3 minutes per document.

Total cost of ownership

Result

Pessimistic scenario. If RAO settled at 50% instead of 72%, the gross monthly saving would drop to around €5,958 and the payback would slip by about three months. The project would still be defensible. That is the exercise to run every time: below what RAO does this stop making sense?

This example is illustrative and built with realistic parameters from NEXTAI projects. It is not a forecast of results and not financial advice: every company should calculate it with its own figures and validate it with its finance function.

How does ROI connect to RAO?

Directly. The RAO (Operational Autonomy Ratio) measures the percentage of process runs completed without human intervention. It is the variable the saving depends on: if RAO is low, the human work does not disappear, it only changes shape, and the ROI collapses.

That is why the correct order is: first instrument the process's RAO, then calculate the ROI. A project that cannot measure its RAO cannot demonstrate its return either; it can only assert it. And an AI that answers well but requires every output to be reviewed has a RAO close to zero, however impressive it looks in a demo.

What mistakes invalidate an ROI calculation?

1. Counting hours freed that nobody reassigns to anything.
2. Budgeting only the development and leaving out context, governance and maintenance.
3. Ignoring the internal team's effort during implementation.
4. Calculating the ROI of "AI" as a whole instead of per process.
5. Using the best month as if it were the average.
6. Not defining the failure threshold in advance: at what figure do you stop.
7. Not repeating the calculation with real data at 6 and 12 months.

How NEXTAI does it

At NEXTAI the ROI is defined before building, not after. In the Mapping phase of the NEXT-5 Roadmap we measure the current process — volume, time per unit, hourly cost, error rate — and set the target RAO. On that baseline we calculate the expected return over 36 months, pessimistic scenario included, and agree the threshold below which the project gets rethought. Superagents are deployed instrumented from day one on top of the Enterprise Brain, so RAO and real cost are data, not estimates.

Frequently asked questions

What is a reasonable ROI for an AI project? There is no universal figure: it depends on the process, the volume and the sector's hourly cost. What is reasonable is to require an identifiable payback within the first 12 to 18 months on high-volume processes with rule-based decisions. If the maths only works over five years, review the scope before investing.

How do you measure ROI when the benefit is qualitative? By converting it into an observable intermediate metric. "Better service" translates into response time, first-contact resolution rate or number of complaints. Those metrics do have history and can be compared before and after.

Should internal staff cost be included in the calculation? Yes. The hours your team spends explaining processes, validating answers and correcting are a real cost of the project. Leaving them out is the fastest way to produce an ROI that will not survive the finance director.

Does Kit Consulting change the calculation? It can improve it. Kit Consulting is a Spanish public grant scheme for smaller companies, and its voucher of up to €6,000 funds the advisory work and the roadmap, reducing what the company carries in the initial phase. It does not cover implementation. Conditions and funding rounds can change: check them with Red.es, the public body that runs it.

Recalculating the ROI at six and twelve months with real data is essential before settling into an annual review. Inference costs, operational flow and the assisted return all shift the initial conditions; a static calculation loses its technical validity very quickly. Assessing the real return of a specific process starts with a free Digital Audit.