What AI implementation actually costs: six quotes with real payback periods

Four things drive the cost, and none of them is which model you pick.

How many systems have to talk to each other. A bot that answers in one chat and looks nowhere else costs tens of thousands of roubles. The same bot that has to see stock in 1C, orders in the CRM and listings on two marketplaces at once costs hundreds of thousands: the work is in four data flows that must reconcile.

Whether your accounting system can expose data at all. 1C can, through HTTP services — but someone has to publish them, lock them down and connect them over an encrypted tunnel to the server where the bot lives. On my construction-company project that wasn't a warm-up task; it was a real share of the seven weeks.

Whose personal data ends up inside. For Russian citizens' data, the database stays inside Russia and the workloads that touch it run on GigaChat or YandexGPT rather than foreign models — that's what 152-FZ requires. It's a choice of model and hosting, and it shows up in the quote.

Who owns answer quality after launch. Your range changes, prices and services change, new model generations ship — and a system nobody has touched since launch starts quoting last year's price list six months in. Support is a separate line item, from ₽25K/month, and it's needed nearly always.

And the rule I apply to myself: if the numbers don't return the investment within eighteen months, I don't take the job. Margins in small and mid-sized business are thinner than people assume, and in eighteen months the problem itself usually changes. I count that with the monthly support included — a quote without the recurring line looks better only until the first invoice.

• · Audit — ₽9,900. A 60–90 minute call plus a written report: 3–5 places where automation actually makes sense, payback calculated in hours and money, and a running order. The report is yours even if we stop there. If we continue, the fee is credited toward the first project.
• · Pilot — from ₽60,000, 1–2 weeks. One process, one channel. A reality check before you commit to a large system. Typically pays back in 6–8 weeks.
• · Project — from ₽150,000, 2–4 weeks. Several channels at once, integrations with amoCRM, Bitrix24, YClients, 1C, and team training. Complex builds start at ₽400,000.
• · Support — from ₽25,000/month. Round-the-clock monitoring, changes for new scenarios, migration to newer model generations, response within two hours during business hours.

The honest limit I keep on the pricing page: an audit only makes sense above roughly ₽1M of monthly revenue. Below that, almost any automation pays back more slowly than the hypothesis behind it survives.

A homeware store: 280 orders a day, 12 categories, selling on its own site plus Wildberries and Ozon simultaneously, a team of 14. The owner was hiring on the theory that “a few more people and we'll dig ourselves out”, and revenue wasn't keeping up.

Two weeks of audit: shadowing six key roles for 90 minutes each, pulling 90 days of logs from every system, timing the 20 most frequent operations to the minute. The output was 11 automation candidates ranked by payback. Only the three fastest-returning ones were built — four more weeks.

Audit ₽380K + build ₽220K = ₽600K. 142 hours a month freed — roughly 1.1 FTE, or ₽89K monthly, which is the investment back in about 7 months on payroll alone. Stock discrepancy across the site, marketplaces and 1C: 7.4% → 0.3%. Monthly close: 9 days → 2. Customer response time: 47 minutes → 6.

What that quote does not include: eight of the eleven candidates — skipped because they paid back more slowly. The case page states 4.2× over 12 months: that figure adds avoided penalties for selling out-of-stock goods on top of payroll. I don't publish the penalty figure, so this quote sticks to the part you can recompute yourself. Full case →

How the stock exchange itself works, and why the drift sat at 7.4% in the first place: inventory sync between 1C, Wildberries and Ozon.

A construction company doing commercial interior fit-outs, 14 sites running in parallel. Foremen called the office 6–8 times a day: “when does the tile arrive”, “has the subcontractor been paid”, “what's left in the budget”. That was 4–5 hours of pure phone time a day for the office manager.

We built a Telegram bot: login by phone number from a pre-approved list, two taps to an answer, almost no typing. Behind it sits a middleware service and 1C UNF HTTP services connected by an encrypted tunnel — 1C itself never faces the internet. Photographing a delivery note on site now posts the goods receipt: it used to take 1–3 days, now it averages four minutes.

₽480K one-off + ₽12K/month support. Seven weeks of work. Calls: 6–8 a day → 0.4. About 0.6 FTE of the office manager freed — ₽48K a month. Payback: ₽480K / ₽48K ≈ 10 months if you ignore support, and with support the net saving is ₽36K a month, so ≈ 13 months. The second number is the one to plan with.

What the quote deliberately excluded, and what I said up front: reduced budget overruns (a foreman now sees the remaining balance before the decision), faster material turnover, and the capacity to take on two more sites with the same headcount. All of it happened — but you can't promise it in a quote, because you can't compute it in advance. Full case →

An online school already had an AI tutor in production. The CTO kept noticing odd behaviour, and due diligence was four months away. I spent two and a half weeks attacking the system following the OWASP LLM Top 10 methodology, plus a week re-testing after the fixes.

Fixed price ₽620K, timeline 2.5 + 1 weeks. 33 findings: 8 critical, 11 high, 14 medium. Among the critical ones — one student's conversation leaking into another student's answer, and personal data going abroad without anonymisation. Every finding came with a concrete fix. A week later the re-test confirmed all eight were closed.

This quote is counted differently, and that deserves to be said plainly. The 152-FZ exposure for a personal-data leak was assessed at ₽15M: ₽620K against ₽15M is a ratio of 1 to 24. But a fine is a probability — nobody knows whether it would have landed, and I won't multiply by it. The checkable part is different: the clean report was a condition of the deal and went into the due-diligence package. If a buyer, investor or enterprise client is about to audit you, that document is a line in their requirements list.

If a full audit is out of budget, there's a one-day express format from ₽45K. It doesn't replace the full one, but it tells you whether there's anything to worry about. Service page and full case →

A marketing agency ran eight clients, each expecting a report on Monday morning for the previous week. Account managers assembled the numbers from four sources by hand: 4–6 hours every week in total, about 20 hours a month — and that assumed no client asked a follow-up question.

About a week and a half of work — the pilot rate, from ₽60K. Manual time: 4–6 hours/week → 0. Reports reach all eight clients at 09:00 Monday, with no “I'll send it before lunch” slippage. Payback ≈ 1.5–2 months at the cost of the hours saved.

This is the cheapest of the six quotes and the fastest to return, because the task is narrow: one repeatable process with manual time you can measure before starting. In scope it's exactly the pilot format. If you're looking for a place to start, look for a task like this one — not for the most visible one. Full case →

An English-language online school needed an iPhone app. Studios quoted ₽4–5M. I shipped the first working version in nine weeks for ₽1.4M, and App Store review passed on the first attempt in 19 hours.

The gap isn't greed. It's the scope of a first version plus one deliberate decision: I built the backend on a managed platform instead of writing my own. That's conscious debt with a documented exit — moving to our own backend later takes 3–4 weeks. A studio prices its own backend in from day one; that's fair, just slower and more expensive.

A second example of the same logic. A fitness studio wanted an app but didn't want a store review on every update, or Apple's 30% commission. We built a web app that installs to the phone's home screen: 1.2 MB instead of 47, updates ship in 30 seconds, membership payments go direct. Commission savings alone came to roughly ₽890K in the first year. App case · web app case →

Four cost lines that rarely appear in a proposal.

• · Post-launch support. ₽12K/month on the construction project; from ₽25K/month in my price list. It also covers migration to new model generations — they ship faster than you get used to them.
• · Getting access. Publishing HTTP services in 1C, granting CRM permissions, coordinating with your existing 1C contractor — that's weeks of work, and it lands inside the project timeline.
• · 152-FZ requirements. They narrow the toolset and occasionally add work — and remove the risk that was valued at ₽15M in the case above.
• · Scope changes mid-flight. I agree a fixed price for every scope change before work on the new scope begins. If a vendor answers that question with “we'll figure it out as we go”, you will be the one figuring it out, and rarely in your favour. The terms are in the payment FAQ.

And four situations where I say “don't”.

• · Revenue below roughly ₽1M a month. There's nothing to compute: the hours saved are worth too little to return even a pilot.
• · Projected return longer than eighteen months. I decline these myself, even when the client is willing to pay.
• · The process isn't documented and changes weekly. Automating chaos gets you faster chaos. Write the process down first, then add AI.
• · Critical infrastructure — banking, oil and gas, telecom, government. There I work as an architect and consultant; I don't hold FSTEC licences for deep penetration testing and refer those engagements to a partner.