AI adoption by industry: where it works in Russia and where to start

AI adoption by industry means choosing the process that delivers a measurable effect first in your particular industry, and building it around the company's accounting system. There is no universal “AI for business”: in manufacturing, visual quality control goes first; in logistics, parsing shipping documents and planning replenishment; in retail, order processing and demand forecasting; in finance, risk management; in services, the first line of customer requests. The industry determines not only the process but also the obstacle: your own item catalogue, your own internal rules, your own accounting system and your own cost of error.

Below: how far apart industries have drifted in AI adoption in Russia, and where those numbers come from. A table of “industry → the process handed over first → what it takes to make it work at your company.” A detailed look at six areas: manufacturing, logistics and warehousing, retail and wholesale, construction and development, finance, services. And why an industry package covers the common part of a process but not yours.

How far apart have industries drifted in AI adoption?

The gap is measured in multiples, not percentage points. According to the release of March 18, 2026 from HSE University's Institute for Statistical Studies and Economics of Knowledge (ISSEK), prepared on the results of a survey of large and medium-sized organizations (excluding small businesses) that Rosstat, Russia's federal statistics service, conducted in 2025 using the institute's methodology, the average share of organizations using AI technologies is 4.8%. Within that same population, groups differ by a factor of 3.6: 14.9% among organizations with more than 500 employees and 4.1% in the “100 employees or fewer” group. AI is in demand most often in marketing and sales, and least often in logistics and transportation planning.

Among large businesses the picture is fundamentally different. According to the study “Artificial Intelligence in Russia — 2025” by Yakov & Partners and Yandex (December 8, 2025; 150 CTOs of large companies across 16 industries, 150 solution vendors and more than 3,500 users were surveyed), 71% of large Russian companies use generative AI in at least one function, and 78% of them report an economic effect. Five industries apply the technology most actively: IT and technology, telecom and media, e-commerce, banking and insurance. They have one thing in common — their data is digital from the start and collected in one place.

These two numbers cannot be compared with each other: the studies have different populations, different survey methods and different definitions of the technology. The practical conclusion comes not from their difference but from what they agree on. What an industry determines is not so much the desire to adopt as the readiness of the data: where an operation is already recorded in a system, AI plugs into a ready stream; where it is agreed over the phone and written in a notebook, the project is preceded by work on documenting the process.

Which process gets handed to AI first in each industry?

The one with many repetitions, where the reference answer is written down and an error is visible before it reaches the outside world. By industry it looks like this.

The right-hand column is the real project budget. The model is equally available to all six industries, but bin addressing in the warehouse, labeled defect examples or call transcripts only appear once the company has created them.

Manufacturing: where do companies start and what do they run into?

With video-based quality control — the only technology in the industry that has reached mass use. According to the study “Promising digital technologies in industry” by Strategy Partners and Tsifra Group (November 5, 2025), the share of Russian industrial plants using machine vision more than doubled between 2020 and 2024: from 18.9% to 41.6%. The other technologies lag far behind: the industrial internet of things at 7.1%, machine learning and big data at 5.5%, generative AI at 0.3% as of 2024. The gap between the first line and the rest has a simple explanation: a camera is installed at a work station and runs on its own, whereas a forecasting model needs linked data from several systems.

The same study names the reason the next step is hard: industry has accumulated large volumes of data, but up to half of it goes unused. The sensors are in place and telemetry is being recorded, yet it is not joined with failure history, the maintenance log and the spare parts catalogue — and predictive maintenance lives precisely on that join.

The second most common scenario is more modest and pays off faster: parsing incoming documents. A specification from a supplier, a request from a customer, a completion certificate from a contractor — the model pulls out the line items and matches them against the catalogue. This is where industry specifics show up: every manufacturer has its own part-numbering system, its own units of measure and its own rules by which a supplier's “50 by 50 angle bar” turns into a warehouse item. The ceiling of recognition accuracy and the ways around it are covered separately, in the article on recognizing source documents in 1C.

One constraint is specific to the industry — where the data may reside. Strategy Partners estimates (as reported by CNews on March 3, 2026) that 79% of large industrial plants rule out public generative AI services on data security grounds. That means the question of where the model is hosted is settled before the model is chosen, not after.

Logistics and warehousing: why is the industry at the back?

Because planning relies on data that, in this industry, more often than not disagrees between systems. The lag is visible in two different measurements, and they should not be confused. The first is process-level: HSE ISSEK, based on the Rosstat survey, names logistics and transportation planning as the business process where AI technologies are in least demand, and that statement is about a process inside organizations in any industry. The second is industry-level: Strategy Partners estimates (January 25, 2025) that in transport and logistics 6% of companies are adopting generative AI, another 16% plan to do so within the next two to three years, and 45% intend to use AI over the same horizon. These are two independent measurements, and they point the same way.

Strategy Partners lists the application areas in the industry in descending order: logistics, freight delivery, customer service, IT support, finance. The agency did not disclose the base for these shares — it is unknown whether the percentages are of all companies in the industry or only of those using AI — so what matters here is the order of the areas, not the numbers themselves.

The first process companies take on is usually not routing but parsing shipping documents and requests: waybills, transport orders, acceptance certificates. It is a stream, the reference answer exists, and an error shows up during reconciliation. Routing and replenishment forecasting come next and require more: bin addressing, honest stock levels and a history of actual — not planned — trip times.

This is where it most often turns out that the project is blocked by something other than the model. If receiving is entered at the end of the shift and mis-picks are corrected by hand, the forecast is built on a stock picture that does not exist. What has to be in the warehouse accounting system before automation is covered in the article on the WMS system, and the connective layer between the warehouse, 1C and sales channels is covered in the piece on how to connect systems into a single loop.

Retail and wholesale: how does online differ from offline?

In data readiness and in having an in-house technical team. The gap inside what looks like a single industry was measured in an earlier study by the same pair of authors: according to work by Yakov & Partners and Yandex published in May 2024, generative neural networks were used by 67% of e-commerce companies and less than 10% of traditional retail. More than a hundred CTOs of the largest companies across 15 industries were surveyed, including representatives of 90% of the largest retailers and 60% of the e-commerce market. The figures relate to 2024; in the December 2025 study, e-commerce remains among the leading industries.

The reason for the gap is not budget. In e-commerce, every action a buyer takes is already recorded: a view, a cart, an abandonment, a return. Offline, a significant part of the same information is not captured at all, and sales history by item is distorted by stockouts and promotions — the product did not sell not because nobody wanted it, but because it was not on the shelf.

In wholesale and retail the first process handed over is usually the processing of incoming orders and counterparty price lists: emails and files in all sorts of formats are turned into order line items matched against your own catalogue. Demand forecasting and replenishment come as a second step — they deliver a larger effect but require a cleaned-up item catalogue and flags on the periods that must be excluded from the calculation.

Construction and development: why is AI still in marketing?

Because that is where the data is digitized, and on site it is not. In a survey by Avito Real Estate (June 2026; 30 development companies, 89 agents and independent specialists) — Avito being Russia's largest classifieds marketplace — 94% of surveyed developer employees say AI is used at their company, 45% report regular use in their work, and 12% have such tools deployed at a systemic level. A sample of thirty companies is small, so the numbers show a direction rather than a precise market level. The distribution across tasks is telling: marketing, content and advertising — 75%, analytics — 53%, sales and contact center — 47%, work with documentation — 28%, and construction supervision and management decisions — 9%.

AGM's study (as reported by RBC Real Estate on May 5, 2026) shows the same picture from another angle. It is based on interviews with heads of commercial functions, marketing and digital development at 24 development companies across different housing classes and regions. AI initiatives exist at 88% of developers, around 80% use or are adopting speech analytics on calls, more than 90% use generative models for content, but only 5% of deployments reach the expected metrics. Up to 80% of failures there are explained by data quality, the absence of formalized processes and a lack of in-house skills.

The practical conclusion for the industry: calls and inbound requests go first, because a recorded conversation is ready-made data. Everything to do with the site — work acceptance, as-built documentation, volume control — requires the rules to be digitized first, and that stage is written into the project as a separate line.

Finance: what gives the industry a head start?

A labeled history of decisions and an explainability requirement the industry has learned to meet. The Bank of Russia, the country's central bank, reports in its consultation paper of November 20, 2025 that one financial market organization in five uses AI, and another third plans to adopt it into business processes within three years; the regulator's survey was conducted in 2025 among 252 financial organizations. AI is used most often in customer interaction, identification, anti-fraud, complaint handling, analytics and forecasting, and risk management.

The industry's head start comes from how its data is built. A credit application, the decision on it and the outcome make up a ready labeled dataset accumulated over years. A manufacturing or construction company has no such set: the outcome of a decision is known, but the initial conditions are recorded in email threads and in people's heads.

The flip side is accountability, and the industry answers it with a move anyone can copy. According to the same Bank of Russia review, more than 80% of financial organizations using AI on a permanent basis give the customer the option to switch to a human operator. The applied rule for a company in any industry follows directly: the higher the cost of an error, the earlier logging, decision explainability and mandatory handover of a disputed case to a human appear in the project.

Services and B2B service: which process pays off first?

The first line of requests and request handling — the process with the most repetitions. According to a study by SberAnalytics and Sber Business Soft covering 2025 (as reported by ComNews on January 22, 2026; 38% of the sample are large businesses of 250 employees or more, about a third are mid-sized, 22% small), 39% of organizations use AI agents and assistants. The most commonly automated areas are document handling and request processing — 70%, accounting and financial records — 55%, HR processes — 34%, customer support — 30%.

In services the barrier to entry is lower than in manufacturing: no cameras, sensors or telemetry are needed. What is needed is a knowledge base in machine-readable form and a person who maintains it. Hence the industry's typical mistake — launching a bot on the website before the terms, pricing and exceptions have been written down. The model starts inventing what is missing, and the very first error goes straight to a customer.

The payback calculation for a specific first-line scenario is covered in the article “When a first-line AI agent pays off”, and the choice between a scripted bot, an assistant and an agent is covered in “AI agents for business”.

What does it take to make this work in your industry?

Six prerequisites, and none of them is about choosing a model. The list looks the same for manufacturing and for services, but what goes into each item differs by industry.

1. A documented process with a written reference answer. Acceptance rules, item-matching rules, tariff conditions. If a rule lives only in the head of a foreman or a senior manager, documenting it is the first stage.
2. Access to data in machine-readable form. Not scans in a shared folder or a messenger thread, but an export that has a structure. This is usually the largest line in the project.
3. Integration with the accounting system. 1C, WMS, CRM, MES, telephony — the system where the operation actually lives. Without writing back, the project stays a demo.
4. Access control. What the model sees, what it never sees, which fields never leave the perimeter. Settled before design work, because it rules out entire hosting options.
5. A “before” baseline. Operations per month, minutes per operation, the hourly cost of an employee. Without those three numbers there is nothing to show for the effect, even if it is there.
6. An owner after launch, and support. A specific person with the time and the authority to update the knowledge base and the catalogues. Price lists change, rules get amended — without an owner, the system starts answering by last year's terms within a quarter.

The first two items require no development budget, but they are exactly what determines whether the project reaches an effect. The general framework of stages and budget is in the article on the stages and cost of AI adoption; where to start if the process has not yet been chosen is covered in “Business automation: where to start”.

Why doesn't an industry package cover the whole task?

Because the package knows the industry but not your company. A ready-made product covers the common part well: standard forms, standard reports, a familiar interface, statutory requirements that are identical for everyone in the market. Companies diverge below that level — in the structure of the item catalogue, in the approval route, in access rights and in the list of exceptions that make employees work around the system in the first place.

It is telling how those who have a choice behave. According to the HSE ISSEK release of March 18, 2026, organizations with more than 500 employees, alongside buying technology, are the most active users of free solutions (38%), develop solutions in house (34%) or customize them (23%). Organizations in the “100 employees or fewer” group more often buy ready-made (62%) or modified (56%) solutions. It is worth reading this half a step further: where the process is already documented, the choice between “take the package” and “build around our process” goes to the second option, and what to build it with — your own team or a vendor — is a question of available resources, not of approach.

Hence the working order: take an industry product as a foundation where it matches your process, and build the difference to fit. INCUBE AI works under contract, keeps data in Russia and builds the system around your rules — with integrations into accounting systems, access control and support after handover. We covered the fork between a ready-made solution and a system built around your own process separately — a custom system or an off-the-shelf package. If you are applying this to your own industry, book a consultation: we will look at the process and tell you straight where a ready-made product is enough and where you will have to build your own.

Sources

• HSE ISSEK, release of March 18, 2026 — a survey of large and medium-sized organizations (excluding small businesses) conducted by Rosstat in 2025: an average share of 4.8% using AI, 14.9% among organizations with more than 500 employees and 4.1% in the “100 employees or fewer” group; the least covered process is logistics and transportation
• ComNews, December 8, 2025: the study “Artificial Intelligence in Russia — 2025” by Yakov & Partners and Yandex — 150 CTOs across 16 industries, 150 vendors, more than 3,500 users: 71% of large companies, 78% report an economic effect
• RBC Companies, November 5, 2025: the study “Promising digital technologies in industry” by Strategy Partners and Tsifra Group and its coverage in Kommersant — machine vision from 18.9% in 2020 to 41.6% in 2024, the industrial internet of things at 7.1%, machine learning and big data at 5.5%, generative AI at 0.3%
• Expert, January 25, 2025: Strategy Partners' estimate for transport and logistics — 6% are adopting generative AI, 16% plan to, and 45% intend to use AI within two to three years; the agency did not disclose its company selection criteria
• CNews, March 3, 2026: Strategy Partners' estimate — 97% of large companies are adopting or plan to adopt AI, one in four has a strategy, and 79% of large industrial plants rule out public generative AI services
• CNews, May 14, 2024: the study by Yakov & Partners and Yandex — generative neural networks are used by 67% of e-commerce companies and less than 10% of traditional retail; more than a hundred CTOs across 15 industries were surveyed
• Avito Real Estate survey, June 2026 — 30 development companies, 89 agents and independent specialists: 94% say AI is in use, systemic deployment at 12%, construction supervision and management decisions at 9%
• RBC Real Estate, May 5, 2026: AGM's study — interviews at 24 development companies: 88% of developers have initiatives, 5% of deployments reach the expected metrics, up to 80% of failures are tied to data quality and the absence of formalized processes
• Bank of Russia, consultation paper of November 20, 2025 and the page “Use of artificial intelligence in the financial market” — a 2025 survey of 252 financial organizations: one in five uses AI, and more than 80% of permanent users of the technology offer a switch to a human operator
• ComNews, January 22, 2026: the study by SberAnalytics and Sber Business Soft — 39% of organizations use AI agents and assistants; document handling and requests 70%, accounting and financial records 55%, HR 34%, customer support 30%