AI adoption in business: the stages and what it costs in 2026

Adopting AI in a business is seven stages from a process audit to continuous operation, and the cost is spread across them very unevenly. A call to a language model costs next to nothing: 1,000 tokens of GigaChat Lite, the language model service run by Sber, come to 0.065 rubles. The money goes elsewhere — into preparing data, integrating with accounting systems, and support.

Below: how many companies in Russia have already adopted AI, what stages the work consists of according to recognised methodologies, which prices are confirmed by official price lists, what usually falls out of the budget, and why most pilots never reach profit.

How many companies in Russia have already adopted AI?

More than half of large businesses, and less than half of the market overall. According to a study by Yakov and Partners and Yandex (December 2025), the share of large Russian companies using generative AI in at least one business function passed 71%, against 54% a year earlier. Of those, 78% report an economic effect.

Across the market as a whole the figures are more modest: in the second quarter of 2025 around 43% of Russian companies were using AI tools, against 28% a year earlier. And a study by SberAnalytics and Sber Business Soft (November 2025, 559 respondents) shows what exactly is being automated: document flow and request processing at 70% of companies, accounting and finance at 55%, HR at 34%, customer support at 30%.

The same study names the barriers directly: the high cost of solutions, the difficulty of integration, the shortage of specialists. The first two are the subject of this article.

Why do 95% of pilots never reach profit?

Because a general-purpose tool does not fit into a specific working process. The MIT NANDA report “The GenAI Divide” (August 2025) is built on 150 interviews with executives, a survey of 350 employees and an analysis of 300 public deployments. The conclusion: around 95% of organisations get no measurable effect on profit, and only about 5% of pilots produce a fast rise in revenue.

The wording is worth reading precisely. The point is not that 95% of projects broke technically — they produced no effect visible in the profit and loss statement. The difference matters: the model works, and there is no money, because it was never attached to a process and never measured.

The second conclusion of that report is more useful than the first. Projects where the solution was bought from a specialised vendor or built with a partner reached a result in roughly 67% of cases. In-house builds from scratch did so three times less often. The reason is not qualification: an outside team has seen dozens of similar deployments and has contractual obligations.

Gartner gives a similar warning: over 40% of agentic AI projects will be scrapped by the end of 2027 because of rising costs, unclear business value and weak risk control.

What stages does AI adoption consist of?

Six phases of work with data and the model, plus two operational ones. This is not an invention of ours: the first six are the industry standard CRISP-DM, which is more than twenty years old and has outlived every change of fashion. The last two come from the MLOps maturity levels described by Google.

1. Process audit and prioritisation. What hurts, where the volume is, where the price of manual labour can be calculated. The output of this stage is a list of processes with an estimate of the effect.
2. Choosing one process with a metric. One. With a number that can be measured before and after. “Cut request handling time from 40 minutes to 10” is a metric. “Adopt AI” is not.
3. Collecting and preparing data. Internal rules, price lists, request history, documents. The most underestimated stage — more on it separately below.
4. Prototype. A check on real data that the task can be solved at all, at acceptable quality. This is where the project has every right to die — that is cheap and correct.
5. Pilot with live users. A limited perimeter, a mandatory way to roll back, and a measurement of that same metric.
6. Integration. Hooking it up to 1C, the accounting and ERP platform most Russian companies run on, plus CRM, telephony and communication channels. This is where a pilot turns into a system, and this is exactly where budgets most often fail to reach.
7. Operation. Quality monitoring, updating the knowledge base, retraining as the data changes, support. Without this stage the system degrades within a quarter.

The phases are not strictly sequential. CRISP-DM explicitly assumes going back: at the modelling stage it turns out the data is insufficient, and you return to step three. A budget with no room for going back is a budget that will slip.

What does a call to a language model cost?

Less than people assume. The official GigaChat price list for legal entities (in force since February 1, 2026, prices include VAT, per 1,000 tokens):

The minimum spend on the service is 600 ₽ a month, VAT included. Asynchronous mode is half the price of synchronous, and that is a direct lever for saving: anything that does not need an answer in real time (overnight document processing, going through the mailbox, labelling) is worth moving into an asynchronous queue.

Work out the order of magnitude yourself. Processing one incoming letter and extracting the company details from it is roughly 2,000–4,000 tokens of input and output. At a thousand letters a month on the Pro model that is hundreds of rubles, not hundreds of thousands. Inference is almost never the main line of an AI project's budget.

What does the team that builds it cost?

The same as ordinary development, because that is what it is. According to the Habr Career salary study for the second half of 2025 (Habr is Russia's main tech publishing and job platform), the median salary of an IT specialist in Russia is 183,000 ₽ a month. In Moscow it is 230,000 ₽, in St Petersburg 200,000 ₽, in the regions around 159,000 ₽. Software architects are the most expensive grade, at roughly 465,000 ₽.

From this you get the cost of a team-month — a benchmark a client can use to check whether a contractor's estimate is realistic. If you are quoted a timeline of four team-months, you understand the lower bound of the cost base and you can see what the price is made of.

It is worth remembering the other side too. An in-house team means not only salaries but hiring, retention, idle time between projects and the risk of losing the one person who understood the system. We went through that fork in detail in the article “Your own system or an off-the-shelf product”.

What does a turnkey AI system cost?

There is no public “market price list” — there are the open price lists of individual studios, and they should be treated as a benchmark, not as an average price. By those price lists the order of magnitude is this: a simple bot answering frequent questions starts at a few tens of thousands of rubles, a search system over corporate documents starts at 200 thousand, and a full AI agent with integrations falls into a range from hundreds of thousands to several million rubles.

The tenfold spread is explained by three things, and these are exactly what is worth asking about in negotiations:

1. The number of integrations. One link to a CRM, versus links to CRM, 1C, the warehouse, telephony and a payment provider — these are different projects with the same name.
2. Data readiness. If the internal rules live in employees' heads, they first have to be extracted and structured. That is work, and it is paid for.
3. Quality and accountability requirements. A demo that runs beautifully on five examples and a system with error handling, escalation and monitoring differ several times over in effort.

We keep our own benchmarks for stages and cost open on the page “What development costs” — you can see there what the estimate is made of.

What is usually left out of the budget?

Everything that happens after the demo. Inference is cheap, development is predictable, and the budget is eaten by the items nobody thinks about at the start.

• Data preparation and cleaning. Usually the largest item in the project. The data is scattered: some requests arrive by email, some in a group chat, some are simply spoken aloud. The data is contradictory: one internal rulebook says leave is 10 days, another says 14.
• Integrations. Every link to 1C, a CRM or telephony is separate work with someone else's API, someone else's limitations and someone else's updates.
• Inference at volume. The price per thousand tokens is multiplied by real load. Calculate it on your peak month, not the average one.
• Monitoring and support. Someone has to notice that answer quality has dropped, and fix it.
• Retraining for data drift. The price list changed, the rulebook was updated, the product range grew — the system has to be brought up to date.
• Your own employees' working hours. The product owner and the subject-matter experts spend hours on acceptance and on explaining the process. That is real money, it just is not in the contract with the contractor.

It is precisely on unready data that projects die most often.

“A model launched on data like that does not simply work badly — it gives wrong answers with great confidence” — SimpleOne, “Why your AI adoption project will die at the start”, Habr, June 29, 2025.

How do you calculate payback before you start?

On one process, in units you understand. The general frame is simple: the ratio of the difference between the benefit and the total costs to the costs themselves. All the substance is in what you put into each part.

1. Calculate the price of manual labour. The volume of operations per month, multiplied by the average time per operation and by the employee's rate. Add an estimate of the losses: requests that went cold because the answer was slow, errors in documents, fines.
2. Add up the total costs. Development, integrations, inference at peak load, a year of support, your own people's working hours.
3. Set a metric and a threshold. What exactly has to change, by how much, and by when. Without that you can neither declare the project a success nor close it in time.

There is no universal payback period for AI projects: it depends on volume, on the price of manual labour and on what an error costs. Anyone who names a period before looking at your process is naming it at random. A worked example of that calculation on a specific first-line scenario is in the article “When a first-line AI agent pays for itself”.

Are there grants and tax breaks for AI adoption?

Yes, both for the client and for the developer. Pilot deployments of Russian AI solutions at client companies are supported by grants under the programme of Russian Government Decree No. 767, operated by the Skolkovo Foundation. RFRIT, the Russian Fund for IT Development, runs a separate line of grants for adopting Russian IT solutions. The Foundation for Assistance to Small Innovative Enterprises issues grants to developers under its “Start-AI” and “Development-AI” lines. Separately, since 2025 there have been tax measures: a multiplier applied to spending on licences for Russian software from the official registry, and an accelerated depreciation rate for AI equipment and IT solutions.

An important caveat: the specific amounts, thresholds and selection conditions change from year to year and from decree to decree. Before you build a grant into a project's financial model, check the version currently in force with the programme operator and with a tax adviser. Planning a project on the assumption of a grant that may not materialise is a reliable way to halt the work halfway.

How do you stay out of the PoC graveyard?

By not starting without answers to three questions. The antipatterns repeat so consistently that you can see them before the start: adopting AI because it is fashionable, the absence of a numeric metric, automating a process nobody has described, and buying an “AI agent” with an old chatbot under the hood.

Three questions to close before signing the contract:

1. Which process and which number? One process, one metric, a known current value. If there is no current value, the first stage is measurement, not development.
2. Who owns the system after handover? A specific person in your company who has the time and the authority to update the knowledge base and watch the metric. Without them the system quietly degrades.
3. What happens when the system gets it wrong? The answer “it won't get it wrong” means nobody thought about operation. There have to be escalation to a human, logs and a way to roll back.

Every failed pilot costs more than its budget. Management settles on the belief that “AI doesn't work here”, and the next project — a sensible one this time — becomes three times harder to defend.

Where do you start?

With a process that hurts and can be counted in money. If you have not chosen one yet, look at the general map in the breakdown “Business automation: where to start”. If you are unsure whether you need an agent or a scripted scenario is enough, the article “AI agents for business” works through that boundary.

After that the order is this: measure the current metric, estimate the volume, collect the data, build a prototype on real documents, and only then calculate a full budget. At IncubeAi we start exactly there: we go through the process, find where hours and leads are lost, and fix the goal in numbers — before the first line of code is written. The system is built around your own rules, under contract, with the data staying in Russia and with support after handover: discuss your project.

Sources

• Yakov and Partners and Yandex, study of generative AI in Russia, write-up on ComNews, December 8, 2025.
• SberAnalytics and Sber Business Soft (November 2025, 559 respondents), write-up on CNews — which processes are being automated and what gets in the way.
• MIT NANDA, “The GenAI Divide: State of AI in Business 2025”, write-up in Fortune, August 18, 2025.
• Gartner, press release of June 25, 2025 — the forecast that over 40% of agentic projects will be cancelled.
• CRISP-DM — the six phases of the industry standard for working with data.
• Google Cloud, MLOps: continuous delivery and automation pipelines — the maturity levels of operation.
• Sber, official GigaChat API tariffs for legal entities — prices per 1,000 tokens, in force since February 1, 2026.
• Habr Career, IT salaries in the second half of 2025 — medians by city and grade.
• SimpleOne, “Why your AI adoption project will die at the start”, Habr, June 29, 2025.
• RFRIT, grants for adopting Russian IT solutions — the terms of support.