AI agent in six weeks: where to start and when it pays off

Tuesday, ten in the morning. On the desk of the operations director at a building-materials distributor sits a proposal for RUB 4.2 million: an AI agent that carries an order from the customer's email all the way to shipment, eleven steps. One evening spent mapping those steps on paper produces a different plan — the agent takes the two steps with open-ended input, and integration plus the price list cover the rest. The first working version arrives in six weeks; the estimate is RUB 1.4 million.

The setup is typical. The figures below are calculated benchmarks for a company of this profile, not a report on a completed rollout. No aggregate public data on agent payback exists for the Russian market. So every figure comes with the method used to measure it — a method that can be repeated on a company's own data in a day or two.

What the breakdown shows. Seven of the eleven steps move data between email, the accounting system and the warehouse system. All three have a programming interface. Ordinary integration does the job here: cheaper and more predictable.

Half the work in those steps is matching item names: the same profile goes by three different names — in the email, in the accounting system and in the warehouse. Cleaning up a catalog of 11,000 items takes six weeks of in-house effort, and both the agent and the integration depend on it.

Two more steps check the order against the price list by fixed rules. A decision that genuinely requires a language model is made at two steps: parsing the customer's free-text email, and deciding what to do with an item missing from the catalog.

Work on those two steps costs roughly RUB 700,000 at development rates. The rest of the estimate goes to integration, access rights and an action log. Proposals rarely break the price down step by step.

A step-by-step breakdown leads to one of two workable outcomes

There are two outcomes, and both produce a result. Either integration closes the task — cheaper, and months sooner. Or the agent stays on one or two steps with open-ended input and delivers a gain that rules cannot capture. A full end-to-end agent is justified when there are more than two such steps. Pilots reach production when the four conditions in the section below hold, and those are checked before the estimate is drawn up.

An off-the-shelf assistant, an agent builder and a custom agent each fit a different part of the process

An off-the-shelf assistant wins where a text answer is what is actually needed. Public pricing from four vendors as of July 2026: RUB 25,000–40,000 a month for a team of up to ten people. Inside sits document search (RAG) and a model that writes the answer. A company can test its own case in a day: 200 requests are tagged as either closable with text or requiring an action inside a system. On a typical flow the second group runs 25–40% and rarely drops below that; the smaller it is, the better the off-the-shelf option looks.

An agent builder produces a demo in a day or two and a firm estimate in the second week. The demo is assembled on test data. The agent needs rights in the accounting system, in the warehouse system and in email: access, integration and logging take 60–75% of the estimate. Budgeted from the outset, they turn the demo into a working version without reopening the budget, while the log and the limits show exactly what the agent did with the data.

A custom-built agent is deployed on the part of the process that is written down in a procedure. Of the eleven steps in the example, six are documented, three live in the heads of two dispatchers, and two change from customer to customer. The pilot goes on the six documented steps and reaches production on schedule; the other five are written up by the client's own team. The seven data-transfer steps are closed by integration, which runs alongside the agent rather than instead of it.

The next step is chosen by a programmer, by a person or by the agent itself

If a step has one input, one output and a rule that fits on a single page, that is work for integration. An agent is needed where the input differs every time: the free text of an email, a scan of unclear quality, a set of seven documents with two missing. The right class of solution for a given task can be sized up with the solution picker.

Payback rests on four conditions, and all of them are checked before the estimate

First: volume, duration and hourly rate multiply into a concrete sum. Annual saving = operations per month × minutes per operation × 12 × the share of time the agent removes × the hourly rate. That share never equals one: review stays with a person, and on processes that reach production the agent removes 60–85% of the time per operation. At a rate of RUB 700 an hour including payroll taxes and an estimate of RUB 500,000, paying back within twelve months takes roughly 5,000 operational minutes a month — that is 1,700 operations of three minutes, or 300 operations of seventeen. The formula is sharper than a threshold like "300 operations and up": a long operation, or a rate of RUB 2,000 an hour and above, produces the same result on a modest flow. Estimates run from RUB 620,000 to RUB 2.1 million, with the middle around RUB 1.3 million — a benchmark, not a constant. The one-off and monthly parts for a given volume are calculated by the budget calculator.

Second: a rule is documented only to the degree that two employees agree on it. The check takes an hour: twenty borderline cases from last month are handed to two experienced employees separately. One or two disagreements is a green light, and the agent settles at an error rate below 5%. Five or more means the rule still lives in people's heads: the error rate sticks at 15–20% and moves neither through fine-tuning nor through rewriting the instructions. Something else works: bring the employees together, agree on the rule and repeat the check — an hour before the estimate instead of four months after the pilot. Between two and five disagreements there is no conclusion: on twenty cases that is noise, and fifty are needed.

Third: access exists now. What is required is a programming interface plus read and write rights; an email to the IT department confirms it, and the date in the reply is what matters. With a date in hand, the pilot goes into the plan; "next quarter" means starting with a system where the interface already exists. Without an interface the agent clicks through the screen the way a person does, and a vendor update brings the work to a halt — such stretches are left for later.

Fourth: the cost of an error is calculated before the pilot. Error rate × cost of one error × volume must come out below the saving. A separate question is whether the action can be undone: an email already sent to a customer and a payment already made cannot be. The condition is met in two ways: the agent is confined to reversible steps — a draft email, a request marked "pending review" — or human approval is kept at the irreversible step and the saving is calculated on the lower share of time. The project comes out ahead either way.

Three checks that make the calculation match the facts

Where the freed-up hours go is decided before the start. A steel trader, order-entry department, four people, a project worth RUB 900,000. 84 orders out of 100 pass without corrections, and time per order falls from six minutes to a minute and a half — stopwatch measurements on 200 orders before launch and 200 three months later. Before the measurement the department quotes three minutes, off by a factor of two. Six months on, payroll is tallied: the same amount, the same four people — volume did not grow, nobody was reassigned, and the hours never turned into money. One line in the contract before the start rules this scenario out: where the freed-up hours go — into higher volume without new hires, into reassigning people, or into cutting headcount. With that line in place, the RUB 900,000 comes back as freed-up hours.

The pilot goes on the steps where two employees agree. A logistics company, eleven steps, four months, RUB 1.2 million. On six steps the agent holds under three errors per hundred — measured on 400 orders over two months. On three steps the error rate sits at 18% and does not come down: month-by-month measurement, four months, the same orders. The reason is visible in advance: decisions are made from experience, and on twenty borderline orders the two dispatchers agree on eleven. Two further steps require access to the carrier's system. The one-hour check from the second condition separates six workable steps from five that are not ready: the pilot on the six reaches production while rules are written and access is requested for the rest.

Irreversible actions stay with a person. A retail chain, processing returns. The agent is allowed to process a return of up to RUB 5,000 without approval. A full audit of 900 returns over three months: three go through on the wrong grounds and the money reaches the customers. The loss is smaller than one month's saving, but the finance department does not allow write-offs of that kind, and approval is put back. The share of time the agent removes drops from 70% to 35% — measured on 200 returns before and 200 after. The project still pays off, just at half the scale of the first calculation. Half an hour with the finance department before the estimate settles the question: it names which actions it will hand over without approval, and the saving is calculated on the lower share.

Processing an order takes 1.6 minutes instead of 8.5 when the agent stays on a single step

The same process, split up differently. The seven data-transfer steps go to integration, the price-list check goes to a rule, and the agent takes one stretch: turning emails and attachments into a structured order. For the first two months, the stop before the order is created in the system stays with a person. The flow is 1,900 emails a month.

The first working version ships in six weeks, and another six weeks before that go into the catalog, handled by the client's own team — that work is not part of the estimate.

What it costs after launch. Model calls run about RUB 7,400 a month at 1,900 emails. Support runs 12% of the one-off cost per year, another RUB 14,000 a month.

Over nine months the parsing is revised about four times. Two suppliers change the format of their emails, one customer starts sending orders as images, and the model vendor raises its price. Some steps are moved to a cheaper model. The unit cost of RUB 3.9 changes twice over that period. Each revision takes four to twenty hours under the support contract; that line item is what holds the share of manual corrections at the level of the first months.

The saving is calculated against hiring that did not happen. Three autumn months against the same months a year earlier: 4,200 orders against 3,230, with the same four people. Growth like that would otherwise take two temporary staff — RUB 190,000 a month including payroll taxes. After RUB 21,000 of monthly running costs, RUB 169,000 remains, and the RUB 1.4 million comes back in about eight months.

The whole scheme is driven by volume. At 300 emails instead of 1,900 the same RUB 1.4 million would not pay back even in three years: at that flow, the place to start is integration. If the catalog and the substitution rules change every week, the catalog is stabilized first.

What to do this week

1. Write out one process step by step on a single page: where a person makes a decision, and where they move data from one window to another.
2. Time twenty cases with a stopwatch, count the operations per month and feed them into the formula with the company's own hourly rate. Timings quoted by the department will not do.
3. Hand last month's borderline cases to two employees separately and count the disagreements: the steps where they agree go into the pilot straight away.
4. Ask the IT department about the programming interface and write permissions. Note the date in the reply — it sets the order of the stages.
5. Put a figure in rubles on a single error, answer whether the action can be undone, and write down where the freed-up hours will go.

All four conditions can be walked through on a specific process with the readiness test.

If the decision is made at one step out of eleven, that is where to start: a working version for that step ships in six weeks. A data-transfer step costs RUB 60,000–180,000 as an integration against RUB 400,000–700,000 as an agent in the same place. Both rest on catalogs that have been put in order — work a company does itself, and starts the same week.