Yandex B2B Tech has launched 'Alice AI Pro' — a corporate AI assistant that breaks down an employee's task into subtasks, forms a 'team' of specialized AI agents, and compiles a single response from them. The product features over 120 ready-made skills, custom skills can be created without writing code, and external services are connected via the MCP protocol and plugins. What is notable here is not the name, but the architecture: the agent's working environment — instructions and tools — becomes the foundation, and the LLM becomes a replaceable component.
What happened
Yandex B2B Tech announced the launch of 'Alice AI Pro' — a corporate AI assistant built on a multi-agent scheme: an employee gives a task, the main coordinator assistant breaks it down into subtasks and distributes them among specialized agents, some of which search for information and work with documents, while others interact with corporate systems. The product is aimed at marketing, design, accounting, HR, analytics, and finance. It can be deployed on a company's own infrastructure and connected to CRM and document management systems. 'Alice AI Pro' boasts over 120 ready-made skills (Skills) — instructions and templates for specific work tasks; a company can create its own skills without writing code. External services are connected via plugins: Phygital+ is responsible for image and video generation, AvangardAI works as a trainer for salespeople.
Context
Against the backdrop of the corporate AI market, this launch is a shift from the 'chatbot on a knowledge base' model to agentic systems, where the main thing is not the choice of model, but the agent's working environment: skills as a reusable layer of instructions, MCP as a layer of tools, and the orchestration of multiple agents. In terms of scientific novelty, there is almost nothing new here: multi-agent orchestration, reusable instructions, and tool-calling are known engineering patterns, and 'Alice AI Pro' is essentially an assembly of familiar components (planner, router, tool-calling). The significance of the launch lies elsewhere: the combination of 'instructions plus tools plus orchestration' was productized for the first time in the Russian-speaking segment specifically as a corporate product. A notable artifact of this approach is Skills as a separate reusable layer: the same skills work both in the assistant itself and in self-assembled agents in Yandex AI Studio. And the use of the open MCP standard in a corporate product is a verifiable signal of a course on ecosystem compatibility.
Why this matters for the industry
For the industry, the launch provides a reproducible architectural template: the value of an agent is made up of a reusable layer of instructions (Skills), a layer of tools (MCP and plugins), and the orchestration of multiple agents. The focus of development shifts from choosing a model to designing the working environment, and engineering work gradually moves to writing and versioning instructions, setting up evals for agentic chains, and observability. Companies that implement such systems are forced to formalize their own regulations as a set of instructions for agents — this forms a new market for 'agentic' corporate documentation and integrations, including future services for 'translating' regulations into a machine-executable form. For startups, the effect is twofold: on the one hand, the space around Skills packages and MCP integrations is growing, on the other — the value of thin wrapper products is falling, because customers are already being shown how to deploy an agent on their own infrastructure and connect it to CRM and document management without code. If the pattern takes hold, in a couple of years the LLM will finally become a replaceable component, and a company's valuable assets will be formalized skills and integrations.
Why this matters for users
For readers who are building their own agents in Yandex AI Studio, a practical step is available right now: skills from 'Alice AI Pro' can be reused in their own agents without rewriting instructions from scratch. Connecting their own services to the assistant is done via MCP or ready-made plugins (Phygital+ for image and video generation, AvangardAI as a trainer for salespeople) without writing code. Similarly, a corporate assistant can be deployed on their own infrastructure and linked to CRM and document management systems. Essentially, the launch provides a working reference for assembling an agent for real processes — 'skills plus MCP plus orchestration' — which can be studied and copied in their own products. For employees within companies, there will be a chance to delegate routine tasks from marketing, accounting, HR, analytics, and finance to an agent by describing them with rules once — almost like an onboarding instruction for a new employee.
What is still unknown / limitations
For now, this is a vendor announcement, and there are no confirmed production data in it: neither the share of successfully completed end-to-end tasks, nor a comparison of the orchestration of a 'team' of agents with a single agent, nor the cost of inference, nor latency, nor API limits are disclosed. It also does not follow from the sources what exactly is executed on the company's infrastructure — the models themselves or only the orchestrator and data, this needs to be checked separately. The list of connectable plugins at the start is narrow, and the fate of the Skills ecosystem and MCP connectors depends on whether the product will survive the announcement stage. The real cost of implementation and the quality of the 'team' of agents' work can only be assessed based on independent cases from the first customers.
Sources
Author
Look at AI, editorial team