T-Bank: start by hiring analysts
KARPOV.CONF 2026: «Начните с найма аналитиков»
B · Useful, limited A first-hand account from T-Bank's head of analytics for its City business, with concrete before-and-after time figures for several LLM tools built by intern cohorts, but all at proof-of-value stage with self-assessed accuracy and no shown evaluation.
Summary based on the full transcript·Original: about 29 minutes to watch
In short
A 29-minute talk at KARPOV.CONF 2026 (28 October 2026) by Danil Nedbaev, who leads analytics for T-Bank's City (Gorod) business and the bank's push to apply AI in Shopping and across the analyst profession. Run work takes about 80% of analysts' time, yet the remaining 20% produced several billion rubles of extra revenue in Shopping last year, and adding 200-500 staff a year is no longer accepted. His answer was to hire intern cohorts, give them structure, and have them build LLM tools for analysts inside the bank's internal model perimeter. More than ten projects exist, all at proof-of-value stage.
Key numbers
- 80% / 20%estimated split of analyst capacity between run work and change work
- 15 hweekly time spent on leadership-team factor reports (about 3 h per analyst across five business lines) before automation
- 1.5 h + 1 week to ~10 mintypical commerce data request: analyst work plus queue before, waiting time with the text-to-segment tool after
- ~240 person-monthsestimated yearly saving from LLM post-analysis of A/B tests
- 9, 10, 15-20size of the first, second and planned third intern cohorts
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- Cohorts: nine interns, then ten, with the third expected at 15 or 20. Each gets onboarding, a buddy from the analyst community and regular stand-ups. Proof-of-concept agents took one to two weeks to build.
- Selection rule: move to agents the regular, well-structured tasks with clear instructions, where the time saved repays the build many times over. Where an LLM fits between two steps, consider redesigning the whole process so one person owns a domain end to end.
- Factor analysis for the weekly leadership-team meeting: about 3 hours per analyst per report, five business lines, so about 15 hours a week. With the tool the analyst saves "2 hours 50 or 2 hours 55 minutes" per report and spends the time on deeper hypotheses.
- Text-to-segment: a natural-language request goes to an LLM over a well-described semantic layer; it asks clarifying questions, runs dozens of SQL queries with cross-validation and returns the answer itself. A commerce request used to take 1.5 hours of work plus a week in the queue; now about 10 minutes. Analysts review queries; he says figures can be guaranteed "practically 100%".
- A/B pre-analysis: across thousands of tests, campaigns with zero effect and campaigns with 50,000 extra utility units took the same effort to prepare, and pre-analysis was usually skipped. The tool uses hybrid RAG to pick 20-30 similar past launches and returns MDE, sample needs, expected effect and recommendations. Post-analysis is estimated to save about 240 person-months a year.
- Other tools: bug reports written from event logs of sessions that ended in errors, a reviewer for Amazon-style two-page planning documents, an "assistant of assistants" that knows the other agents, and automatic descriptions of BI tables.
- Tooling (Q&A): OpenClaw is not used as unsafe. Staff use Nessy (name as heard), an internal whitelisted fork of Qwen Code, as their Claude Code equivalent, moving from n8n and notebooks to an agent harness. An internal MCP Hub already connects Jira, the wiki, the data perimeter and the corporate messenger.
How they got it
The speaker reports on his own programme. Time savings are estimates of analyst hours before and after; accuracy is backed by analysts reviewing outputs, not by a stated evaluation. The 240 person-months figure is an estimate with no method shown.
Limits
- All tools are at proof-of-value stage; productisation is described as long, hard and costly and has not happened yet.
- Accuracy claims ("practically 100%") are self-assessed, with no error rates or test set.
- The library summary pairs "about 10 minutes" with the weekly factor report. In the talk, 10 minutes is the text-to-segment result; for factor analysis the saving is given as 2 h 50 to 2 h 55 minutes per 3-hour report.
- Source is YouTube auto-captions: tool names are uncertain (Nessy, MCP Hub), and some terms are mangled.
- The host closed with a pitch for karpov.courses corporate training.
Why it matters for companies
The staffing logic will be familiar to any analytics function: most capacity goes to recurring reporting, and headcount growth is capped. Using interns, with onboarding, buddies and stand-ups, as a cheap way to test many agent ideas before handing winners to engineering teams is a model other firms can copy. The text-to-segment case depends on a well-described semantic layer and analysts reviewing queries, which is the work most companies skip. The pre-analysis point carries over too: if weak and strong campaigns cost the same to prepare, cheap screening before launch is where the time goes furthest.
Key passages in translation
Russian original
Но вот что здесь важно, что за 20% ресурса аналитиков в прошлом году мы в шопинге достигли несколько миллиардов рублей дополнительной выручки для продукта.
What matters here is that with 20% of analyst capacity, last year in Shopping we delivered several billion rubles of extra revenue for the product.
Russian original
Ну, делают её обычно за полтора часа. В каком случае мы видим колоссальную экономию? Ну, во-первых, что это делается уже не полтора часа, а там минут 10 с перерывами на кофе.
They usually do it in an hour and a half. Where do we see the big saving? First, it now takes not an hour and a half but about 10 minutes, coffee breaks included.
Russian original
На самом деле мы сейчас находимся только на этапе proof of value.
In fact we are only at the proof-of-value stage right now.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com
- Tags
- AI enablement