Own agent or platform: Winny customer support
ВИКИБОТ ГОВОРИТ #7: свой ИИ-агент или готовая платформа? Опыт Winny
B · Useful, limited A support lead at the Russian AI-tutor firm Winny gives a candid single-builder account with response times, test costs and failures; the figures are his own, from a five-week window, on a stream hosted by a support-bot platform vendor.
Summary based on the full transcript·Original: about 109 minutes to listen
In short
A 109-minute Russian live stream (27 October 2026) on the channel of Wikibot, a vendor of support-bot platforms, with Andrey Ovchinnikov, director of customer experience at Winny, an AI tutor for school pupils that also serves teachers and parents. Since early August 2026 he has built the support agent "Fixie" mostly alone, by giving specifications to Claude coding agents, using the open-source helpdesk Chatwoot as a shell. In about five weeks the bot closed 60% of tickets, and a scheduled Claude job reviews escalations every three days to fill knowledge-base gaps. He is open that the code is not maintainable and says engineers will have to rebuild it before it scales.
Key numbers
- 60%share of support tickets closed by the bot in about five weeks of full operation, including tickets auto-closed after the user went silent
- 12-15 s to 2-3 sbot reply time per message before and after the rebuild (maximum 5 s)
- ~47 centsmodel cost of a run of 100 test tickets
- <1%the support bot's share of the company's total LLM spend
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- Escalation rules: payment problems ("I paid and nothing appeared") and blocking bugs go to a human. Before handing over, the bot collects device and browser facts (sometimes from screenshots) and writes a summary capped at 600 characters.
- "Closed" has two meanings: the user goes silent and the ticket auto-closes after 8 hours, or (more than half of closures, he says) the user thanks the bot and the model judges the issue resolved. The hosts question whether the first counts as resolution.
- Every three days Claude reads all bot-to-human escalations and adds knowledge-base entries where the bot could have answered.
- Reply time fell from 12-15 seconds per message to 2-3 seconds, maximum five. The knowledge base is a small JSON knowledge graph on the same server: the model goes to a topic block such as payments and reads about 10 rows instead of 200. Claude wrote the graph and search in about two hours.
- A tiered multi-model setup was tried and dropped as slower and costlier. He settled on one model (name garbled in captions) that at medium reasoning costs half as much as GPT-4.1 and is "7% or so" more accurate. Yandex's Alice did badly on the main product's prompts.
- Costs: 100 test tickets came to about 47 cents; the bot is under 1% of total LLM spend, so it is not tracked. Development time cost more than inference.
- Failures: a male persona kept adopting the user's gender through six or seven prompt iterations, even when hard-coded. Asked to make the bot multilingual, the coding agents spent a weekend writing about 1,000 prompt lines for language detection the model already does. A wrong answer found in production led him to have the bot reopen four closed dialogs and apologise.
- Group chats: the bot joins Telegram and Max chats as admin and moves sensitive or face-losing questions to private messages within the same ticket; Claude wrote that integration in an evening, the hand-off took about three days.
- He plans a self-built helpdesk (Chatwoot has more than a thousand open bugs on GitHub) and says people, not agents, must build it.
How they got it
A hosted conversation with a short live demo; the guest describes his own system. The 60% rate has no baseline, no error rate and no split between silent and confirmed closures. The digest is based on cleaned Russian auto-captions in which model names and some numbers are garbled; the guest's name comes from the video description.
Limits
- Self-reported by the builder; five weeks of operation; no outside check.
- The headline closure rate includes users who stopped replying.
- Model names are garbled in captions, so the model comparison cannot be verified.
- The hosts sell support-bot platforms, the build-or-buy topic of the episode.
- The system depends on one person and on generated code he calls unmaintainable.
Why it matters for companies
The case shows what one non-engineer with coding agents can ship in a quarter: a working support agent, fast replies and very low model cost. It also shows the bill that follows: one person as the only maintainer, generated code nobody can maintain, and a closure metric that flatters. The useful practices are the scheduled review of escalations to fill knowledge gaps and the narrow topic-blocked knowledge base that keeps replies fast. Firms letting staff build their own agents should plan for the rebuild by engineers before relying on them.
Key passages in translation
Russian original
А за это время 60% обращений закрыл бот. А при этом раз в 3 дня у меня стоит автоматическая задача, что мой бедный клод просматривает все обращения, которые были, просматривает все эскалации
In that time the bot closed 60% of tickets. And every three days I have a scheduled task where my poor Claude goes through all the tickets and all the escalations.
Russian original
Никто не будет разбираться аа в сотни тысяч строк нейрослопа.
Nobody is going to dig through hundreds of thousands of lines of AI slop.
Russian original
это всё, будем честными, это навайпкоженная штука, и я просто даю ТЗ, грубо говоря, и моя армия агентов это всё делает.
Let's be honest, it's all vibe-coded: roughly speaking, I just write the spec and my army of agents does all of it.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com