Many companies have already tried GenAI. What can now be considered true value, rather than just a technology demonstration?
In enterprise, value begins with a product's compliance with existing customer constraints. Often, foreign commercial models cannot be used, the solution must operate within a closed loop, and be compatible with available infrastructure. If, at the same time, the AI product delivers the required quality, speed, and helps solve a real problem, then it is no longer an experiment, but a useful tool. Therefore, it is important to evaluate not only time savings or error reduction, but also readiness for industrial implementation without compromising data and reliability.
Do you agree that competition now needs to be at the context level, not at the fundamental model level?
Yes. Competition at the fundamental model level is the domain of very large players with corresponding resources. For most companies, the realistic path is to use existing open-source or commercial models and enrich them with their own context, data not publicly available: regulations, documents, history of requests, business rules, accumulated expertise. But data alone solves nothing. The context needs to be formed so that the model provides a correct answer or invokes the necessary tool within the business process.
What does this “year of context” look like in enterprise?
First, companies need to gather knowledge stored in disparate wiki systems, in notes, internal documents, and even “in the heads” of experts. All this needs to be not only collected in one place, but also learned how to store, update, and effectively search for answers within this knowledge to then enrich the context of language models. When creating AI agents based on language models, it is important to understand that the system prompt or system instruction is also part of the context. In essence, it is a brief description of the business function performed by the assistant: what it should do, within what boundaries, with what data, and according to what rules. Therefore, the profession of a prompt engineer or context engineer, which was recently perceived almost as a joke, is becoming truly important in 2026. The quality of such a setup directly determines whether the assistant will simply answer in general terms or be able to help in a specific business process.
Do companies need to train their own fundamental models?
Not always. Training a fundamental model is a very expensive endeavor; a company must understand if such investments will pay off. Many business cases today can be covered by context engineering: RAG, well-crafted prompts, scenario configuration, and data integration. Fine-tuning or separate ML models are justified where there is a specialized task, a stable data flow, and a clear economic effect.
Which tasks benefit from GenAI, and where does it only complicate the architecture?
Generative AI is not suitable where strict explainability is required. For example, in scoring models or default probability assessment. If a document needs to be classified into a small, fixed set of classes, a large language model may also be redundant: classical ML, rules, or search will suffice. GenAI provides value where there is unstructured text, ambiguity, a need to prepare a summary, reconcile data from different sources, or assist a person in a task where a simple “yes/no” is not enough.
If a customer says, “We will assemble this ourselves from open source,” what is the vendor's value?
This question applies not only to artificial intelligence but to software development in general. Assembling a solution from open source is still internal development. It needs to be managed, it requires IT resources, architecture, testing, and support. These resources are expensive, and often the team goes through the same trial and error path that others have already gone through dozens of times. The vendor's value lies in ready-made and proven software, implementation experience, debugged models, integrations, and responsibility for the result. In enterprise, not only algorithms are important, but also security, access management, auditing, SLAs, and industry specifics. For example, at GreenData, we train our own document classification and document fragment recognition models so that solutions work better specifically in the context of Russian legislation and Russian document types. This is the practical value of the vendor: they shorten the path from idea to industrial result and take on some of the risks that, with internal development, remain entirely with the customer.
Which GenAI scenarios are already mature, and which are still in the experimental zone?
Mature scenarios include Meeting Intelligence: speech recognition, meeting transcription, summarization, protocol generation. The second block is RAG and corporate search: working with large unstructured repositories of documents or code and generating an answer based on the found context. AI agents as a replacement for traditional interfaces are still developing. A more mature example is first-line support, where chatbots with LLMs under the hood already work quite stably. But if the process is formalized, it is better to use classical workflow, RPA, rules, or an ML model.
What practical formula would you give to businesses?
It's worth starting not with an abstract desire to “implement GenAI,” but with a narrow applied use case. For example, document recognition. Then, study what solutions are already on the market, where they make mistakes, what open-source alternatives are available, and what can be refined in-house. It's important to calculate the economics immediately. In the cloud – token consumption and cost control; on-premise – infrastructure and support. If the weaknesses are clear, data is available, the economics align, and the effect can be measured, then the project is worth developing. GenAI in 2026 is not the magic of a large model, but a precise task selection, quality context, and only then scaling.