Monthly AI Budget

How the calculation works

There is a single starting point: how many person-hours the process eats today. The number of employees is multiplied by hours per week and by 4.33 — the average number of weeks in a month. The hourly cost comes from the salary divided by 164 working hours.

Multiplied together, these give what the process costs the company today. That is the baseline for comparison, and it almost always turns out larger than expected: routine work does not look like a line item until someone adds it up.

Next comes the share AI is ready to take over. A "one hundred percent" option is deliberately absent: in tasks involving documents and clients there is always a tail of non-standard cases that needs a human. Even in mature deployments the automated share rarely exceeds seventy percent of the original volume.

What makes up the monthly bill

Three line items.

Model and tokens. Every call to the model costs money — that is inference. The calculation assumes 55 RUB per hour of routine work removed: the coefficient is derived from the typical volume of text that passes through a model when processing documents and correspondence. For tasks with long documents it will be higher; for short replies to clients, lower.

Infrastructure and monitoring. Server, storage, watching the system run. In the cloud that is around 28,000 RUB per process; inside a private perimeter, roughly 2.4 times more, because of the hardware and its upkeep.

Support and retraining. Twelve percent of the costs removed. This is not abstract insurance: the system meets new document types, wording changes, exceptions appear. Without ongoing support, answer quality degrades within months.

Why payback can mislead

The period is calculated by dividing the one-off investment by the net saving. The one-off figure is not fixed: a 500,000 RUB base plus the monthly cost of the process. A fixed sum would produce nonsense at large volumes — it would show a hundred-person process paying back in a month.

There is a trap here worth stating outright. Saving hours is not the same as saving money. If after automation the people stayed in place and moved on to other tasks, the company gained capacity, not a smaller payroll. That is value too, but it does not arrive as a reduced bill. Only what is genuinely no longer being paid — or no longer needs hiring — counts as money.

That is why the lower figure in the calculator is the more honest one: the monthly budget will certainly be paid, whereas the saving depends on what is done with the freed-up time.

What shifts the numbers most

The order of impact is usually this.

First, the state of the documents. Scattered formats, scans instead of text, contradictory versions of internal rules push the project's cost up more than any model does.

Second, the cost of an error. If a wrong answer is expensive, every step needs a human check, and the saving halves. If an error is cheap, the system can be left to work on its own.

Third, volume. Automation starts to pay back above a certain flow: infrastructure and support costs barely depend on the number of documents, while the saving grows with it. Hence a simple corollary: the larger the flow, the faster the return.

When there is nothing to calculate

If the process changes every month, stabilise it first and automate afterwards: otherwise the system is retrained faster than it delivers value. A month or two spent settling the rules comes back as a shorter implementation.

If the volume is under a few hundred operations a month, the monthly bill will eat the saving. At those volumes it is cheaper to put the process itself in order.

The calculation rests on two figures: the volume of the process and the fully loaded cost of an hour. Until a process has been chosen, start with the readiness test — it will show which area to cost out first.