Where AI genuinely pays off — and where it does not
The rule that holds up: AI pays back on high-volume repetitive work with text, speech and documents. Not on “digital transformation” in the abstract, but on specific processes where a person does the same thing day after day — answering the same questions, moving data from a document into a system, writing standard copy, screening hundreds of near-identical requests.
The payback points that hold up in small and mid-sized business:
- •Bookkeeping paperwork — reading invoices and delivery notes, entering them into the accounting system. Saves upwards of half an assistant’s salary (details),
- •First-line support and sales — answering routine questions around the clock, qualifying leads, following up on silent ones (support, sales),
- •Answering calls — a voice AI on the inbound line so nothing goes unanswered (AI operator),
- •Content — product cards, articles, posts and newsletters in the brand voice (AI copywriter),
- •Contracts — a first pass over risks and redline comparison (AI lawyer),
- •Reporting — a morning summary of sales and advertising instead of manual spreadsheets (AI analyst).
Where AI does not pay off: anywhere decisions are one-off and expensive — strategy, major deals, difficult negotiations — and anywhere the data is not digitised at all. If enquiries live in a paper notebook, you need a CRM first and a model second.
Three deployment models
Model 1. Cloud services
Subscriptions to ChatGPT, Claude, Gemini and specialised SaaS. Fast to start and a low barrier to entry, with two structural downsides: your data is processed on someone else’s servers, and the cost grows with usage. For non-sensitive tasks at modest volume, a perfectly reasonable choice.
Model 2. A self-hosted model on your own server
An open-weight model running on your hardware. Data never leaves the company, there is no per-token billing, and it works with the internet unplugged. It needs an investment in the server (from around 450,000 ₽) and competent setup. This is the option for anyone with trade secrets, personal data or regulatory obligations — banks, healthcare, law firms, manufacturing. The full breakdown is on the on-premise AI server page.
Model 3. Hybrid: local perimeter plus top models behind a scrubber
Our working setup where both confidentiality and maximum quality are required. The routine is handled by the local model inside the perimeter. When a top-tier cloud model is genuinely needed, the request goes through double scrubbing: rules and dictionaries strip names, registration details, amounts and company names, then the local model re-reads the text for anything missed. Only a de-identified task leaves, and the answer is reassembled inside your contour.
Five mistakes that sink deployments
1. Starting with the technology instead of the process
“Let’s buy an AI and work out what for later” is the most expensive route there is. The right order is the reverse: find the process where people burn the most time on routine, cost that time out, and only then pick a tool.
2. Expecting magic without configuration
Out of the box, a model knows nothing about your product, your procedures or your tone. Without loading a knowledge base and calibrating on real examples, the output is average-of-the-internet — and the project gets buried with “we tried it, it does not work”.
3. Automating everything at once
A deployment that works goes one process at a time, with a measurable result inside the first month. Success on the first process buys both the experience and the internal support for the next one.
4. Taking the human out of the loop immediately
For the first few weeks the AI works under supervision: a person reads the answers, adjusts the scenarios, marks up the errors. Autopilot comes on gradually, starting with the most reliable cases. Switch it on day one and you get public embarrassments instead.
5. Forgetting about the data
Nobody checked what staff are pasting into public chatbots — and meanwhile contracts, customer lists and payroll spreadsheets are going out that way. An AI usage policy needs to exist before the first corporate project, not after.
What it costs
| Format | Order of cost | Who it suits |
|---|---|---|
| Cloud subscriptions | 2,000-10,000 ₽/month per user | non-sensitive tasks, first experiments |
| An AI employee for one job | 15,000-30,000 ₽/month | one process turnkey: support, content, paperwork |
| Self-hosted server with a model | from 450,000 ₽ one-off | confidential data, many users |
| Hybrid perimeter | server plus scrubbing setup | when you need both top-model quality and data protection |
For comparison: one junior employee costs from 60,000-80,000 ₽ a month with taxes. An AI employee on a standard process costs two to four times less, works around the clock and scales without hiring.
A plan for the first 30 days
- •Week 1. Pick the single process with the most routine. Cost the current state: hours times rate. Gather the material — knowledge base, examples, procedures,
- •Week 2. Configure the AI on that process, run it against historical data, compare with how people actually do it,
- •Weeks 3-4. Supervised launch: the AI works, a person checks. Weekly review of errors and tuning,
- •Day 30. Measure the result: time, money, quality. If it pays back, switch on autopilot for the reliable cases and pick the next process.
That cycle removes the biggest fear — “what if it does not work”. After a month you have either a measurable result or an honest “it does not pay back on this process”, at minimal cost either way.