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

AI adoption in a business takes seven stages, from a process audit to ongoing 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, Sber's large language model, cost 0.065 rubles. The money goes elsewhere: on preparing data, integrating with accounting systems and keeping everything running.

Below we look at how many companies in Russia have already adopted AI, which stages the work consists of according to established methodologies, which prices are confirmed by official price lists, what usually drops out of the estimate, and why most pilots never reach the bottom line.

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 exceeded 71%, up from 54% a year earlier. 78% of them report an economic effect.

Across the market as a whole, the numbers are more modest: in the second quarter of 2025, about 43% of Russian companies used AI tools, up from 28% a year earlier. And a study by SberAnalytics and Sber Business Soft (November 2025, 559 respondents) shows what exactly gets automated: document workflows and request processing (70% of companies), accounting and finance (55%), HR (34%), customer support (30%).

The same study names the barriers directly: the high cost of solutions, integration complexity and a shortage of skilled staff. The first two are what this article is about.

Why do 95% of pilots never reach the bottom line?

Because a general-purpose tool does not fit into a specific workflow. The MIT NANDA report “The GenAI Divide” (August 2025) is based on 150 interviews with executives, a survey of 350 employees and an analysis of 300 public deployments. The conclusion: about 95% of organizations get no measurable impact on profit, and only about 5% of pilots deliver rapid revenue growth.

The wording is worth reading precisely. It does not say that 95% of projects broke technically: they did not produce an effect you can see on the income statement. The difference is fundamental: the model works, but there is no money, because nobody attached it to a process and nobody measured it.

The report's second finding is more useful than the first. Projects where the solution was bought from a specialized vendor or built with a partner reached results in about 67% of cases. In-house builds from scratch did so a third as often. The reason is not skill: an external team has seen dozens of similar deployments and is bound by a contract.

Gartner issues a similar warning: over 40% of agentic AI projects will be canceled by the end of 2027 due to rising costs, unclear business value and weak risk controls.

How do you bring an AI agent into a business, and where do you start?

Start with one measurable business problem, not with choosing a model or a platform. Describe the process and its “before” metric, pick a repeatable scenario for the pilot, prepare the data and knowledge base, connect the working systems, test the result within a limited scope, and only then scale the solution.

Goal and pilot process

Pin down the problem, the owner and the “before” metric, then pick one repeatable area of work.

Data and integrations

Check the documents, the access rules and how context is passed into the CRM or another working system.

Testing and scaling

Compare the before and after results, gather feedback, and expand the solution once the pilot is done.

What stages does AI adoption consist of?

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

1. Audit and prioritization of processes. What hurts, where the volume is, where the cost of manual work 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 you can measure before and after. “Cut request processing time from 40 to 10 minutes” is a metric. “Implement AI” is not.
3. Collecting and preparing data. Policies, price lists, request history, documents. The most underestimated stage; more on it separately below.
4. Prototype. Testing on real data whether the task can be solved at all with acceptable quality. A project has every right to die here, and that is cheap and correct.
5. Pilot with live users. A limited scope, a mandatory option to roll back, and measurement of that same metric.
6. Integration. Connecting to 1C, the accounting and ERP platform most Russian companies run on, as well as CRM, telephony and 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, knowledge base updates, 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 modeling stage it turns out there is not enough data, and you return to step three. An estimate that does not allow for these loops is an estimate that will slip.

How much does a call to a language model cost?

Less than people tend to think. The official GigaChat price list for legal entities (in effect since February 1, 2026, prices including VAT, per 1,000 tokens):

The minimum spend on the service is 600 rubles a month, VAT included. Asynchronous mode is half the price of synchronous, and that is a direct lever for saving money: anything that does not need a real-time response (overnight document processing, sorting email, labeling) is worth moving to an asynchronous queue.

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

How much does the team that builds it cost?

The same as ordinary software development, because that is exactly what it is. According to Habr Career's salary study for the second half of 2025, the median salary of an IT specialist in Russia is 183,000 rubles a month. In Moscow it is 230,000 rubles, in Saint Petersburg 200,000 rubles, and in the regions about 159,000 rubles. Software architects are the most expensive grade, at around 465,000 rubles.

From this you can calculate a team person-month, a benchmark the client can use to check whether a contractor's estimate is reasonable. If you are quoted four person-months, you know the lower bound of the cost and can see what the price is made of.

It also helps to remember the other side. An in-house team is not just salaries but also hiring, retention, idle time between projects and the risk of losing the one person who understood the system. We covered this choice in detail in the article “Custom system or off-the-shelf software”.

How much does a turnkey AI system cost?

There is no public “market price list”: there are published price lists from individual studios, and they should be treated as a reference point, not as an average price. According to these lists, the order of magnitude is as follows: a simple FAQ bot 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 somewhere between hundreds of thousands and several million rubles.

A spread of several dozen times comes down to three things, and these are exactly what to ask about in negotiations:

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

We keep our own benchmarks for stages and cost public on the “How much development costs” page, which also shows what goes into the estimate.

What usually gets left out of the estimate?

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

• Data preparation and cleanup. Usually the largest item in the project. Data is scattered: some requests come in by email, some via a shared chat, some are discussed verbally. Data is contradictory: one policy gives 10 days of leave, another gives 14.
• Integrations. Every connection 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 production volume. The price per thousand tokens is multiplied by the real load. You need to calculate for the 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, a policy was updated, the product range grew — the system has to be brought up to date.
• Your employees' working time. The product owner and subject-matter experts spend hours on acceptance and on explaining the process. That is real money; it just isn't in the contract with the contractor.

Unready data is what kills projects most often.

“A model launched on data like this doesn't just perform 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 and in units everyone understands. The overall framework is simple: the difference between the benefit and the total costs, divided by those costs. All the substance lies in what you put into each part.

1. Calculate the cost of manual work. The monthly volume of operations, multiplied by the average time per operation and the employee's rate. Add an estimate of losses: leads that went cold because of a slow response, errors in documents, penalties.
2. Add up the total costs. Development, integrations, inference at peak load, a year of support, your own people's working time.
3. Set a metric and a threshold. What exactly should change, by how much and by when. Without this, you can neither declare the project a success nor shut it down in time.

There is no universal payback period for AI projects: it depends on volume, on the cost of manual work and on how much a mistake costs. Anyone who names a period before looking at your process is guessing. This calculation is worked through for a specific first-line scenario in the article “When a first-line AI agent pays off”.

Are there grants and incentives for AI adoption?

Yes, for both the client and the developer. Pilot deployments of Russian AI solutions at client companies are supported by grants under the program of Russian Government Decree No. 767, operated by the Skolkovo Foundation. RFRIT, the Russian Fund for the Development of Information Technologies, runs a separate grant line for deploying Russian IT solutions. The Innovation Promotion Fund gives grants to developers under its Start-AI and Development-AI tracks. Separately, since 2025 there have been tax measures: an increased coefficient on expenses for purchasing licenses of Russian software from the official registry, and an increased depreciation rate for equipment and IT solutions for AI.

An important caveat: specific amounts, thresholds and selection criteria change from year to year and from decree to decree. Before building a grant into the project model, check the current version with the program operator and with a tax advisor. Planning a project around a grant that may not materialize means stopping work halfway through.

How do you stay out of the PoC graveyard?

Don't start without answering three questions. The anti-patterns repeat so consistently that you can spot them before the project begins: adopting AI because it's fashionable, having no numerical metric, automating a process nobody has described, and buying an “AI agent” that turns out to be an old chatbot under the hood.

Three questions to settle 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 with the time and authority to update the knowledge base and keep an eye on the metric. Without them, the system quietly degrades.
3. What happens when the system gets something wrong? The answer “it won't” means nobody has thought about operation. There must be escalation to a human, logs and the ability to roll back.

Every failed pilot costs more than its budget. Leadership comes away convinced that “AI doesn't work here,” and the next, well-thought-out project becomes three times harder to defend.

Where do you start?

With a process that hurts and can be measured in money. If you haven't chosen one yet, look at the overall map in “Business automation: where to start”. If you're unsure whether you need an agent or a scripted workflow will do, that boundary is covered in the article “AI agents for business”.

From there, the order is: measure the current metric, estimate the volume, gather the data, build a prototype on real documents, and only then calculate the full budget. At IncubeAi, this is exactly where we start: we break down the process, find where hours and leads are being lost, and set the goal in numbers before the first line of code is written. The system is built around your workflow, under contract, with data kept in Russia and with support after handover: discuss your project.

Where the numbers come from

• Yakov and Partners and Yandex, a study of generative AI in Russia, summary on ComNews, December 8, 2025.
• SberAnalytics and Sber Business Soft (November 2025, 559 respondents), summary on CNews — which processes are being automated and what gets in the way.
• MIT NANDA, “The GenAI Divide: State of AI in Business 2025”, summary in Fortune, August 18, 2025.
• Gartner, press release of June 25, 2025 — forecast that over 40% of agentic projects will be canceled.
• CRISP-DM — the six phases of the industry standard for data work.
• Google Cloud, MLOps: continuous delivery and automation pipelines — maturity levels for operation.
• Sber, official GigaChat API rates for legal entities — prices per 1,000 tokens, in effect 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 deploying Russian IT solutions — terms of support.