Every AI project sooner or later runs into one question from the CFO: "And when does it pay off?" If you don't have a clear answer backed by numbers, the project will stall at the budget-approval stage, no matter how promising it is. The good news: calculating the payback of an AI system is no harder than calculating the payback of a new employee. This article gives you a working framework you can apply to your own process in half an hour.
Why "AI is innovation, you can't calculate ROI" is an excuse
Let's clear up the main myth right away. If a vendor tells you the payback of an AI project can't be calculated because "it's innovation," that's a warning sign. They either don't believe in the result or don't understand your business. Any automation produces a measurable effect, and it can be estimated in advance with reasonable accuracy. The only question is which of three models to use.
Three ways to calculate the effect of an AI project
In enterprise projects the effect usually comes from three sources. One of them, or several at once, will fit your task.
Method 1. Freeing up working hours. The most transparent method and the easiest to defend to a CFO. The logic is simple: employees spend N hours on routine work, AI takes over part of it, and the freed-up time turns into money at the employee's hourly cost. It suits processes where people do repetitive work: parsing documents, filling in the CRM, searching archives, preparing standard paperwork.
Method 2. Revenue growth through conversion. If AI speeds up sales or improves their quality, we calculate the increase in revenue. For example, AI lead qualification lets a manager handle several times more inquiries, and call analysis raises the share of deals closed to script. Here the effect is larger, but harder to calculate: you need an honest link to the funnel.
Method 3. Reducing the cost of errors. For processes where an error is expensive: anti-fraud, compliance monitoring, detecting anomalies in finance. We calculate how much the company loses on errors today, and what share of them AI will catch in real time instead of "after the fact, two weeks later."
The formula for freeing up hours
Let's take the most common case: freeing up time. The formula looks like this:
Where the hourly cost is calculated from the fully loaded payroll cost, that is, salary plus taxes and contributions (in Kazakhstan this is roughly +30% on top of the gross salary), divided by the number of working hours in a month (~168).
A worked example with real numbers
Suppose a department has 10 lawyers, each spending 3 hours a day parsing tender and contract documentation. The average salary is 500,000 tenge. An AI system with a RAG architecture takes on 70% of this routine (initial review, extracting requirements, compliance checklists), and the lawyer only checks the result.
Let's calculate: hourly cost ≈ (500,000 × 1.3) / 168 ≈ 3,870 tenge. Freed up: 10 × 3 × 21 × 0.7 ≈ 441 hours per month. In money that's about 1.7 million tenge per month, or roughly 20 million tenge per year. With a platform subscription starting at 12 million tenge per year, this effect pays it back in about 7 months, well within a healthy range.
You can run these calculations for your own process on our ROI calculator: enter your numbers and get an estimate in a minute.
What counts as a good payback period
For an enterprise AI project in Kazakhstan, a healthy benchmark is payback within 6 to 9 months. If your calculation comes out lower, great, that's a quick win. If it's more than a year, take a closer look: either the process was poorly chosen, or the volume of routine is too small for automation to justify itself. In that case it's better to pick a different process for the first rollout.
Three common mistakes in the calculations
Mistake 1. Counting only the implementation price and forgetting about running costs. Implementation is often just a third of the real costs. Then come tokens (the charge for requests to the model), support, and further development. Calculate the full cost of ownership over 3 years (TCO), not the price of getting started. A typical story: a company bought a solution for 5 million, then spent another 20 on running it over two years, because no one asked about TCO.
Mistake 2. Overstating the automation share. AI rarely covers 100% of a process; there's almost always a portion that needs a human (checks on critical operations, non-standard cases). Assume a realistic 60 to 80%, not 100%, or the calculation will fall apart at its first encounter with reality.
Mistake 3. Ignoring "soft" benefits. Freeing up hours is only the direct effect. Behind it there are often things that also have value: employees stop burning out on routine, decisions are made faster, there are fewer errors. These benefits are harder to quantify, but they're worth mentioning in your business case.
What to ask a vendor about ROI
When a vendor proposes an AI solution, ask three questions. First: which of the three models will be used to calculate the effect on our process? Second: what automation share are you assuming, and why that one? Third: what is the full cost of ownership over 3 years? If there are no clear answers to these questions, the payback calculation is built on thin air.