How Can Mid-Sized Businesses Adopt AI Without Big-Business Budgets?

Vorontsov describes companies from his practice: annual revenue of two to five, sometimes ten to fifteen billion roubles, and up to five to seven hundred employees. These are boundaries for this conversation, rather than a formal business classification. Headcount reveals little about organisational complexity. In custom software development, hundreds of people participate in internal processes; in staff augmentation, most specialists work for clients while a small team runs the provider’s operations. Similar headcounts and revenue can therefore require different changes. Understanding how daily work is organised, who makes decisions and which skills are available internally matters more than company size alone when discussing AI.

Owners usually arrive with a specific problem. An events organiser must rebuild logistics after a cancellation; a trading company loses margin to marketplace commissions and struggles with its product range and unsold stock. It may have billions in revenue, a few dozen employees and one systems administrator. The owner needs to sell twenty tonnes of oil or resolve a problem with stranded rail wagons. AI looks like a quick saving: buy subscriptions, send employees to courses, and expect efficiency. Polomodov identifies the missing step: people keep their familiar routines unless somebody explains what must change. Tool access alone establishes neither a new objective nor a way of working. The conversation then examines two implementation stories.

A marketplace trading company offers an example of a considered start. Its two partners, one a former development manager, clearly divided responsibilities. Without an operations director, they asked employees themselves which tasks consumed time: commercial proposals, manually merging regional sales spreadsheets, and category analytics. They set aside an experiment budget from net margin and treated it as an investment project. Vorontsov contrasts this with buying subscriptions, then complaining about token spending and making four marketers share one account. The partners hired a systems analyst, a former consultant familiar with data, and a specialist to teach AI tools. A prototype contest offered prizes of 350,000, 200,000 and 150,000 roubles. The analyst checked prototypes against actual spreadsheets and 1C: goods were recorded after the fact, date corrections followed particular rules, and contractors’ warehouse balances arrived on different schedules. Personal and margin-related data stayed out of the cloud; staff learned handling rules instead of using a masking gateway. This describes that company’s arrangement, rather than a universal information-security model.

The less successful case involved an IT staffing firm with three to four hundred employees. Rapid growth in 2022–2023 had slowed; its owner was enthusiastic about AI. First came CV summaries for recruiters, then summaries of available candidates sent to clients. Personal screening remained necessary, so interviews also needed summarising. Automation addressed real pressures: more competitors and candidates, fewer client openings, idle recruiters and growing numbers of employees awaiting another assignment. But there was no internal IT partner capable of bringing the scattered initiatives together. Discussing this failure, Vorontsov broadens the conversation to constant firefighting: urgent matters displace systematic work on data, processes and training. Initiatives spread across functions, the platform never became a working system, people lost confidence, and the effort was abandoned. Polomodov connects this to process maturity, comparing AI to a stock cube expected to improve any dish.

Owners may assign AI to an enterprising executive, the IT director or the operations director. A larger company, with roughly fifty billion roubles in revenue, illustrates the difficulty. Its IT director spends around $5,000 monthly on tokens, has built Power BI dashboards and an agent that retrieves and consolidates data. Yet category managers in another department do not use it: they lack a method for analysing inventory alongside sales, and their workflow remains unchanged. The transformation message reached IT but missed the business. Vorontsov suggests bringing in people to cover missing expertise: a project owner, part-time support, or help with a difficult technical decision. Managers must explain the benefits, talk to users and follow actual adoption. A technical team alone cannot do all this.

Vorontsov recommends spending one or two weeks observing how employees actually work. Documents may travel from Wildberries through Yandex Disk and Telegram, despite a nominally working integration between 1C and warehouse software. He says one to three immediately solvable tasks always emerge. During discovery for a property developer, five to seven of fifty to sixty employee suggestions proved viable; discovery took two weeks instead of five or six. Automation has become more accessible: faced with rising receivables, an employee built a bot within hours to check 1C documents and send Telegram reminders. Speed carries risks, though: serious trouble can appear by the second or fifth task. Generated documents can also replace discussion and agreement. Vorontsov recalls a call where participants exchanged twenty-eight-page documents about a new process for concluding contracts. Nobody had read the other side’s proposals, and the conversation stalled in painful silence.