How to tell whether AI is worth it for your process

Most companies using AI do not yet see it in their profits. In McKinsey's 2026 global survey only 37% of respondents report any impact of AI on operating profit (EBIT), about the same as a year earlier. Only around 6% say that at least 5% of their profit comes from AI. That does not mean AI does not work. It means that choosing where to apply it decides almost everything.

In Bulgaria most companies still have that choice ahead of them. According to Eurostat, in 2025 AI was used by 8.5% of enterprises with at least 10 employees, against an EU average of 20%. The advantage of a later start is that you can choose better.

Where the money comes back

A 2025 study by MIT's NANDA project found that more than half of generative AI budgets go to sales and marketing, while the biggest return is in back-office automation: document processing, internal requests, accounting and administrative work. The study is based on interviews, a survey and an analysis of public deployments, so it is a pointer rather than a precise measure. But it matches practice: the effect is easiest to measure where the work is frequent, repetitive and has a clear outcome.

A good AI candidate has four marks. It happens often, every day or every week. It involves text, documents or classification. There is a person who can judge whether the result is right. And a mistake can be caught and fixed before it becomes expensive.

A simple calculation before the first euro

You do not need a complex model to judge whether it is worth it. You need five numbers about the process as it is today: how many times a month it happens, how many minutes it takes each time, what an hour of work costs for the people doing it (including social contributions), how often there are errors and what one error costs, and what share of cases AI can realistically handle while a person checks the result.

An example (the numbers are an illustration, not a statistic): an accounting team enters 2,000 invoices a month, 4 minutes each. That is about 133 hours, and at €15 an hour the process costs about €2,000 a month. If AI extracts the data and a person only checks it, the time drops to, say, 1 minute per invoice. The saving is about 100 hours, or about €1,500 a month.

From that you subtract the monthly cost of the solution: model, maintenance, monitoring. If it is €300, the net benefit is about €1,200 a month. A one-off investment of €12,000 then pays back in about 10 months.

The calculation is simple on purpose. If even the optimistic case takes years to pay back, there is probably a better task for a first project.

Benefits that are not in the hours

Not every benefit is time saved. Sometimes the value is a faster answer to a customer, fewer missed deadlines, or knowledge that does not leave with an employee. These benefits are real, but name them in advance and decide how you will measure them. Otherwise, six months later nobody will be able to say whether the project succeeded.

When not to do it

If the process happens a few times a month, automation rarely pays off. If everyone does it differently, AI will automate the chaos. If there are no examples to check quality against, you cannot tell whether it works. And if nobody in the business owns the result, the project will stay a technology experiment.

Gartner predicts that at least 30% of generative AI projects will be abandoned after proof of concept and names unclear business value as one of the main reasons. The calculation above is the cheapest way to stay out of that group.

A pilot with success criteria

Before you start, write down what success means. For example "90% of invoices go through without correction" or "a customer answer within 2 hours instead of 1 day". Pick a period of a few weeks and test on the real flow, not on hand-picked examples. Then "continue" or "stop" is a matter of numbers, not impressions.

In its 2025 survey McKinsey reports that the companies getting the most from AI are almost three times as likely to have redesigned the workflow itself rather than just adding a tool to the old one. So a good pilot tests two things: does the model work, and does the new way of working work.

If you would like to run this calculation together for one of your processes, write to us. One conversation is usually enough.