What AI implementation actually costs: six estimates with payback periods

Four things determine the cost.

How many systems need to be connected. A chatbot that answers in one messenger and looks nowhere else costs tens of thousands of roubles. The same bot that must see stock in 1C, orders in the CRM and listings on two marketplaces at once costs hundreds of thousands. Most of the work goes into four data exchanges that have to agree with each other.

Whether your accounting system can hand out data. 1C can do this through HTTP services, a built-in way of passing data to other programs. They have to be written, published, locked down against outsiders, and connected to the bot's server through a secure channel. In the ready-made solution for a construction company, I set aside a noticeable share of the seven weeks for this.

Whose personal data ends up inside. Russian citizens' data has to be stored in a database located in Russia. For tasks that involve that data I use Russian models, GigaChat or YandexGPT: that way the data doesn't go abroad, and no separate cross-border transfer procedure under 152-FZ is needed. The choice of model and hosting shows up in the quote.

Who is responsible for answer quality after launch. Your product range changes, prices and services change, new models come out. A system nobody has touched since launch will be quoting last year's price list six months later. Support is a separate line item, from ₽10K a month, and it's needed almost always.

And a rule I keep myself: if the numbers show the investment won't come back within eighteen months, I don't take the job. Small and mid-sized businesses usually work on small margins, and in eighteen months the task itself often changes. I include monthly support in the calculation: a quote without the monthly line looks attractive only until the first invoice.

• · Audit — ₽5,000. A 60–90 minute call and a written report afterwards: 3–5 tasks worth automating, payback calculated in hours and money, and the order to do them in. The report is yours even if we don't continue. If we do, the fee counts toward the first project.
• · Pilot — from ₽60,000, 1–2 weeks. One process, one channel. A test of the idea on a specific task before you invest in a large system. Estimated payback: 2–3 months.
• · Project — from ₽150,000, 2–4 weeks. Several channels at once, connections to amoCRM, Bitrix24, YClients or 1C, and team training. Complex projects start at ₽400,000.
• · Support — from ₽10,000 a month. The system is monitored around the clock, and if it fails, I get an alert. I adapt it to new scenarios and move it to newer models when they fit better. During business hours I respond within two hours.

The guideline I follow: automation makes sense from roughly ₽1M of monthly revenue. Below that, the hours saved rarely pay back even a pilot. All plans are on the pricing page.

For the estimate, take a home goods shop: 280 orders a day, 12 product categories, selling on its own website, Wildberries and Ozon at the same time, a team of 14. It hires almost every month, and revenue grows more slowly than headcount.

The audit takes two weeks. I watch each key role at work, export 90 days of logs from every system and time the 20 most frequent operations. The result is a list of tasks worth automating, each with a payback estimate. For a shop like this, three fixes come first: stock sync, AI for first-line support, and documents generated in 1C. I allow about four more weeks to build them.

Two-week in-depth audit ₽380K + implementation ₽220K = ₽600K. By the estimate, 142 hours a month are freed up, worth ₽89K in payroll. On that alone, the investment comes back in about 7 months: 600 ÷ 89 ≈ 6.7.

What this estimate doesn't include: the other tasks on the audit list, and the benefit of selling out-of-stock goods less often and paying fewer marketplace penalties for it. That benefit depends on the shop, so the payback counts payroll only, a figure you can recalculate yourself. Ready-made solution: process audit for an online shop →

One of my projects already has a similar check: a revenue reconciliation across three sources runs every day with a 3% threshold. Why stock drifts apart across marketplaces, and how a sync that brings it back together works, is covered separately: inventory sync between 1C, Wildberries and Ozon.

The estimate is for a company that fits out commercial premises and runs 14 sites at once. Each foreman calls the office 6–8 times a day: “when does the tile arrive?”, “has the subcontractor been paid?”, “what's left in the budget?”. The office manager spends 4–5 hours a day on the phone.

The ready-made solution is a Telegram bot with login by phone number from an approved list: two taps to an answer and almost no typing. Behind it are an intermediate service and HTTP services in 1C:UNF, joined by an encrypted tunnel, so 1C itself can't be reached from the internet. A foreman photographs the delivery note on site, and the goods receipt is posted in 1C straight away.

Budget guideline: ₽480K one-off + ₽12K a month for support. About seven weeks of work. By the estimate, the bot frees up 0.5–0.6 of the office manager's full-time workload: on a ₽80K monthly salary that's ₽40–48K a month. Payback without support: 10–12 months. With support, the net saving is ₽28–36K a month, and payback takes 13–17 months. Plan with the second figure. It's close to my eighteen-month limit, so for your company I'd recalculate it from your actual number of sites and calls.

Two benefits are left out of the estimate: fewer budget overruns on sites (a foreman sees the remaining budget before making a decision) and faster turnover of materials. Nobody can calculate these in advance. Ready-made solution: 1C and a Telegram bot for foremen →

Parts of this setup already run in my projects. My AI knowledge-base assistant is a Telegram bot with three access levels. A converter for Federal Treasury statements builds files for import into 1C. How the bot for 1C is built inside, where this approach doesn't work, and what to check before you start: Telegram bot for 1C.

The example for this estimate is an online school with 20,000 students. An AI tutor is already built into its product: it sees each student's progress, their message history and sometimes personal data from sign-up forms. Before launch, nobody tested how the bot would behave if someone tried to trick it.

Full audit: ₽400K to ₽1.1M, 2–3 weeks. The price depends on the scope and the number of AI agents in the system. I attack a copy of the system following the OWASP LLM Top 10, a widely used list of risks for AI applications. Every finding comes with a remediation plan, and once it's applied, I run the same attacks again.

This estimate works differently, and that should be said plainly. For leaking the data of 10,000 to 100,000 people, part 13 of article 13.11 of the Code of Administrative Offences sets a fine of ₽5–10M for a company. Even the lowest fine is 4.5 times the top price of a full audit. But a fine is only a possibility: nobody knows whether it will be imposed, so I don't calculate the payback from it. The tangible benefit is this: a report with the re-test date can go into the document package for an investor, a buyer of the company or a large client. If their requirements include an item on data protection, the report answers it.

If a full audit is beyond your budget, start with the one-day express diagnostic for ₽45K. It shows whether there is anything to worry about at all. It doesn't replace a full audit. Service: AI security audit and the ready-made solution for an online school →

Already done in my projects: in the same AI knowledge-base assistant I measured on the live index how many private fragments the “team” access level could see. There were 1327; after the permissions fix, there were 0. What such an audit looks for, what fines the law sets, and how to run a basic check yourself in twenty minutes: Personal data and LLMs.

A marketing agency had eight clients, and each expected a report for the previous week on Monday morning. Account managers collected the numbers from four sources by hand: 4–6 hours a week in total, about 20 hours a month. And that was if no client asked a follow-up question.

About a week and a half of work, so the pilot rate applies: from ₽60K. Manual time: 4–6 hours a week → 0. Reports go out to all eight clients by 09:00 on Monday, with no more “I'll send it before lunch”. Valued at the cost of the hours saved, the work pays back in about 1.5–2 months.

This is the least expensive of the six estimates and the fastest to pay back. The reason is the narrow task: one repeating process whose manual time you can measure before work starts. If you're looking for a place to start, look for a task like this in your own business. Case: weekly client reports →

Estimate 5. An iPhone app — from ₽1.2M, about nine weeks. Take an online English school with 4,800 active students. It needs an app that reminds students about lessons, trains pronunciation and takes payments. Within that time, the first SwiftUI version reaches the App Store.

Two things keep it within this budget: the scope of the first version, and one deliberate decision. I put the server side on Firebase, a ready-made cloud platform. It's a simplification with a migration plan written in advance: moving to the school's own server will take 3–4 weeks. If local law requires student data to stay in the country, the backend goes on a local server from day one, which adds about four weeks.

Estimate 6. A web app for fitness studios — from ₽450K for a first working version, 4–5 weeks. The example is a studio with group classes, booking through YClients and memberships. The web app installs on the iPhone home screen straight from Safari, works offline and updates without App Store review. Memberships are paid for through YooKassa, a Russian payment service. I build the app on Astro, the same framework behind the website I made for an electrical panel manufacturer.

Ready-made solution: an app for an online school · ready-made solution: a web app for fitness studios →

Four cost items worth asking about before you sign a contract.

• · Support after launch. At my rates it starts at ₽10K a month; in the construction estimate it's ₽12K. It also covers moving to new models, which come out faster than people get used to them.
• · Getting access. Publishing HTTP services in 1C, granting permissions in the CRM, agreeing all this with your existing 1C contractor: it takes weeks, and those weeks count toward the project timeline.
• · 152-FZ requirements. They narrow the choice of tools and sometimes add work. That is the price of a smaller risk: for the online school in estimate 3, a leak would mean a fine of ₽5–10M.
• · Changes in scope along the way. I price each change in scope separately and agree it with you before the work starts. If a vendor answers this question with “we'll sort it out as we go”, you'll be the one sorting it out. The terms are in the payment FAQ.

And four situations where I tell you plainly: don't.

• · An hour saved is worth too little. My guideline is revenue of about ₽1M a month: below that, the hours freed up rarely pay back even a pilot.
• · The projected payback is longer than eighteen months. Then I turn the job down myself, even if the client is ready to pay.
• · The process isn't written down and changes every week. Automating chaos gets you faster chaos. Write the process down first, then add AI.
• · You own a significant critical information infrastructure object. Penetration testing there requires a contractor licensed by FSTEC, the Russian technical regulator. More on this in the article on personal data.