The Real Cost of AI Adoption: Nine Budget Lines and Three Hidden Ones

Three vendor proposals for one task: RUB 540,000, RUB 1.6 million and RUB 3.5 million. A 6.5× spread. All three describe the task in identical words — process incoming supplier invoices and enter them into the accounting system.

The setup is typical: a packaging manufacturer, 260 employees, 2,400 invoices a month, St. Petersburg. The CFO takes the middle option — RUB 1.6 million. Here is what the first year actually consists of, counted from signed statements of work and invoices rather than from a line in a proposal:

• RUB 1.6 million — the project itself;
• RUB 210,000 — connecting to the accounting system: in the chosen proposal, integration is a line with no price attached, billed "as work is performed";
• RUB 300,000 — support, RUB 25,000 a month under contract;
• RUB 150,000 — model calls: an average of RUB 12,500 a month against a growing volume;
• RUB 352,000 — four months of an in-house accountant working full time to clean up the contract archive and the counterparty directory; calculated at the fully loaded cost of an hour of that accountant's time, taken from payroll.

That comes to RUB 2.6 million. The cheap proposal misses by 4.8×, the expensive one by 34% in the other direction. Both extremes were quoted before anyone had looked at a single document — the same mistake with opposite signs.

What follows is a line-by-line breakdown of that setup. The figures below are calculated benchmarks for a company of this profile, not a report on one specific rollout. Each one comes with the measurement behind it, so the measurement can be repeated on other data and produce a different number.

Why does the spread reach 6.5×?

Four questions explain it. What the one-time portion consists of and in what proportion. What the line nobody writes into a proposal actually costs. How to calculate the monthly column with a formula rather than a range. And why freed-up hours on their own do not reduce what a company pays out.

The proportions below apply to this profile: companies of 100–1,500 employees, document processing and customer support, 1,000–10,000 operations a month. For computer vision on a factory floor, or for loads in the tens of thousands of operations per second, the proportions shift — most of the weight moves into hardware.

What are the three market answers to the price question?

A ready-made subscription service — RUB 30,000–80,000 a month. The range comes from the published rate cards of such services for 10–20 seats and can be verified in half an hour on their own websites. Where a company lands within that range depends on seat count and storage volume, not on "AI complexity." The service does not know a given counterparty directory or a given approval chain: adapting it to the process is either impossible or sold as a separate project. Data leaves the perimeter — for some industries, the question ends right there.

Hourly billing — RUB 3,500–6,000 an hour, median around RUB 4,800. A contract like this has no ceiling: the hour estimate is given before the audit, and it diverges from actuals in one direction only — a 2× overrun on an hourly contract is unremarkable. A 900-hour project and a 2,400-hour project start with identical words in the meeting room.

A fixed price quoted before the audit. The vendor either builds in a buffer against the unknown — which is how the RUB 3.5 million from the opening of this article comes about — or trims the scope down to whatever fits the quoted sum, then makes up the difference with change orders in month four. Either way, the answer is to the question "what does the project cost," while an executive needs two other figures: year one and year three.

They are calculated like this. Year one is the one-time portion plus twelve monthly bills plus the in-house time spent preparing data. Year three is monthly bills only, plus one or two rounds of fine-tuning. In the setup at the top of this article, that is RUB 2.6 million against RUB 580,000. There is no universal multiplier between the two: the ratio depends entirely on where the model sits and how often the process changes.

What makes up the one-time portion?

The medians add up to 100 percent by design: this is the profile of a typical budget for a task of this kind. The ranges must not be added up — the extremes never occur in a single budget, and the upper bounds alone total 155 percent.

Three lines are missing from vendor proposals more often than they appear: "data cleanup," "fine-tuning," "infrastructure." A missing line does not mean the cost is absent. It has moved onto in-house staff and into next year's budget.

Cleaning up documents

The median is 22 percent of the one-time portion, and this line almost never comes up in the first meeting: before an audit, nobody can honestly estimate it, the vendor included.

A cheap in-house check: take 20 documents at random from the real incoming flow — from email and the scanner folder, not from an accounting-system export — and count how many have no text layer, exist in two contradictory versions, or carry handwritten edits. Three characteristic outcomes from the same pre-launch measurement:

Three points will not yield a formula. Something else is visible: above one-third, timelines stop holding; above one-half, the conversation is no longer about AI but about order in the data. Until the archive is sorted out, a search system over company documents (RAG) will confidently answer from an outdated version of a contract — and acceptance testing does not always catch it.

How should the monthly portion be calculated?

The range in the first line means nothing until the measurement is done: run 50 genuine operations, look at the actual input and output token counts, multiply by the rate and by monthly volume. In the setup at the top of this article, that works out to RUB 3.6 per invoice and RUB 15,000 a month across 4,100 invoices. The order of magnitude for a given process can be checked quickly with the budget calculator, but it is no substitute for the measurement.

Fine-tuning is needed wherever the process changes substantially: a new document type, a new product category, a new language of incoming requests. If an AI agent sits underneath, this line also covers process branches that were absent from the original specification.

Support gets cancelled after a year — usually to save money. What happens next shows up in a single metric: the share of documents falling through to full manual entry. Five months without oversight and it climbs back into double digits: the typical regression runs from 7–9 percent to 19–24. The cause is always the same: suppliers change their form layouts and there is nobody left to fix the parsing. The metric calculates itself from the system log and is worth tracking month by month.

How do saved hours turn into money?

The system frees up 40 hours a week for the legal department. The lawyers stay in their seats, salaries unchanged — what the company pays out does not move by a single ruble. What the company gains is headroom: the same department can handle twice as many contracts. That is valuable, but it does not appear in the P&L.

Freed-up hours convert into money along exactly two paths: a hire that never happened as volume grew, and revenue that did not exist before. Nothing else should be called savings.

Hence a limit on the arithmetic worth keeping in mind when reading any payback calculation. A drop in the cost per operation and a hire that never happened are the same saving counted from two sides. They must not be added together in one model — yet in the calculations brought forward for sign-off, that addition shows up regularly.

Where do estimates come apart from reality?

Processing completion certificates: six weeks estimated, eleven actual. A construction company, 2,000 certificates a month. The audit looks at an export from the accounting system — where everything is tidy. A random week of the incoming flow, 465 certificates, shows something else: 34% are phone photographs, some with a stamp printed across the table. That is the "34%" row in the table above. What needs looking at is the incoming flow, not the showcase inside the system.

A model bill that missed by 2.4×. The average request length — 1,200 input tokens — is taken from an export of 500 emails covering the month before launch, based on the body text without attachments. Across the first thousand requests it emerges that 31% arrive with an attached document of 8–15 pages. Instead of the expected RUB 48,000, month one comes to RUB 115,000. The fix is architectural: long attachments are processed in a separate cheap step and the bill returns to RUB 52,000. Hence the rule for measurement — count not the average operation but the heaviest one, and its share of the flow.

A pilot at 94 percent accuracy that was rejected. An insurance broker, field extraction from application forms: 94% of fields correct across a sample of 600 documents over a three-week pilot — by prior agreement, that was enough. At acceptance it emerges that the remaining 6% still have to be found by eye, which means reading all one hundred percent. Timing 40 documents before and 40 after yields a 19% time saving instead of the projected 70. A step is then added to flag low-confidence fields: three more weeks and money that was never budgeted.

What comes back in the setup from the opening?

Same profile — the packaging manufacturer, fourteen months from the start of the project. Each line is measured by its own method, given in the last column.

The last two rows sit side by side deliberately. Vendor proposals headline the cost per unit without the company's own monthly bill — here the difference is nearly 1.5×.

Nine minutes per invoice describes a profile with no electronic document exchange with any supplier, where every invoice is matched against a contract by hand. Where half the flow arrives as structured files, the same figure comes out at half that. Cost per document is not comparable across companies — only within one company, before and after.

The money here is counted one way only: through the hire that never happened. Under the old arrangement, 4,100 invoices would have required 615 hours a month against the original 360. The difference is 255 hours, or 1.8 FTE, and people are hired whole: two of them. The hire never happened. Two people at fully loaded cost run RUB 176,000 a month; less the RUB 40,000 bill for the system, RUB 136,000 remains.

The saving does not appear on acceptance day. For the first five months the flow has not yet grown, and the extra people would not have been needed in any case. Across the nine months when the saving is running, roughly RUB 1.22 million comes back against RUB 2.16 million invested (project, integration, in-house accountant). By month fourteen, payback has not arrived: at this rate it falls in month twenty-one. The vendor's deck put it at month eleven.

Where this does not work. At 500 invoices a month, the same arithmetic yields a saving of 55 hours, or RUB 34,000 a month. The monthly bill barely depends on volume (RUB 25,000 support plus RUB 1,800 for the model), leaving around RUB 7,000 net against a one-time outlay starting at RUB 500,000. That is a payback horizon over which the process itself will change twice. The second stop signal is document formats that change every quarter: fine-tuning eats the saving whole.

What six checks come before the call to a vendor?

1. Time one operation with a stopwatch: 20 consecutive cases, with two different employees. Multiply by the fully loaded cost of an hour — salary with contributions and leave, divided by 1,700 working hours a year.
2. Break the proposal received into a one-time column and a monthly column. Check the "data cleanup" line separately: its median share is 22%, and if the line is absent, the work has moved onto in-house staff.
3. Open 20 random documents from the incoming flow, not from an export. An unusable share above one-third means the timeline will not hold.
4. Ask not for the price of the project but for the price of one operation in production — together with the measurement behind it: how many operations were run, and when.
5. Answer in writing what will happen to the people whose hours are freed up. No answer means no saving in money, and no reason to calculate payback at all.
6. Run the process through the automation readiness test and the solution class selector — that clarifies what order of magnitude is even in play.

These six points will not force a vendor to talk in budget lines: everyone can defend their own number persuasively. What they give is something else — the ability to compare proposals against each other and see which lines each of them left out.