On September 30, 2026, AvitoTech released a 16-minute video, “Will AI Replace the Analyst?”, in which three of the company's analysts discuss their own experience working with neural networks. The main conclusion was sober: fully delegating analysis to the model doesn't work — the answers turn out to be too general, and a significant part of the work has to be polished manually. As a result, the pricing of skills shifts: generating SQL and code becomes cheaper, while domain expertise and the ability to properly frame a question, gather context, and check whether the agent distorted the result become more valuable. The video breaks down three real cases: migrating 200 data marts in six months, automating the checking of employee performance, and training an LLM to analyze metrics.
What happened
In the video, three Avito analysts take turns describing how they built collaborative work with neural networks on real company tasks. By the end of the discussion, the participants arrive at a common denominator: pure delegation of analysis to the model doesn't work, the results are too generalized, and the analyst still has to bring them to a working form themselves. At the same time, the hard skills themselves are becoming cheaper — writing a SQL query or generating code today is possible for almost anyone.
Context
Avito has one of the largest data teams in the Russian market, and the classifieds platform itself is among the largest in the country, so its conclusions are the proven practice of a large company, not a thought experiment. The video is not the first material on this topic: AvitoTech has already published on “Habr” a breakdown of the corporate agent platform “Vitalik,” which includes an LLM portal, a factory of internal agents based on ADK, and RAG on top of Elasticsearch. From this article, an important detail follows: the infrastructure connecting the analyst with AI has long been assembled at the company and works in real production processes, not shown as a demo. Against this backdrop, the pragmatic conclusion of the video looks logical: the models took over routine work, but where a domain-accurate result is needed, a human is still indispensable.
Why this matters for the industry
For the industry, the story reads as a map of value shifting. Routine generation of SQL and code is turning into a commodity, and this puts pressure on vendor products in the “query and report generator” category: large companies like Avito close such pain points with their own LLM-based integrations, not purchased solutions. At the same time, niches open up: building not an “AI analyst,” but the wrapper around the human-agent pair — tools for context, checking, and auditing results, which today every company makes by hand. If the trend holds, formalization should be expected: internal eval frameworks for analytical agents, metrics like the share of manual rework, and standard outlines for checking conclusions. Judging by the architecture of “Vitalik,” the template “assistant portal plus a factory of specialized agents plus retrieval on top of internal data” will most likely become the standard corporate analytical infrastructure.
Why this matters for users
If you are an analyst or data engineer, the checklist is simple: what to develop is not the speed of writing queries, but the ability to correctly frame a task for the agent, gather context for it, and check that the conclusions are not distorted. A useful step would also be a local audit: find places in your processes where the model is already writing code and queries, and close the points from which the result goes further without checking. The three cases from the video — migrating marts, automated checking of employees, and analyzing metrics — are convenient to compare with your own processes. You can repeat the basic integration at your company: RAG on top of a corporate knowledge base plus internal agents plus mandatory manual validation. The video itself is short, it will take 16 minutes, the link is available in the sources.
What is still unknown / limitations
The material is a corporate case, not a study: the sources have no metrics, benchmarks, or reproducible experiment setup. The number “200 marts in six months” cannot be used as a benchmark for migration speed — it is unknown how long the migration would have taken without an LLM, what size the team was, and what share of the results was accepted without rework. From the implementations, it is also impossible to extract specific models, APIs, prices, and measured gains in time or quality. Finally, the links to the video are equipped with erid advertising markers, meaning this is native advertising; the production use of LLMs at Avito is plausible and indirectly confirmed by the article about “Vitalik” on “Habr,” but the specific effects of the cases do not follow from these sources.
Sources
- Will AI Replace the Analyst? — AvitoTech video (YouTube)
- From an LLM portal to a factory of internal agents: how Avito builds the corporate assistant “Vitalik” — Habr, AvitoTech
Author
Look at AI, editorial team