The price of a useful answer
Цена полезного ответа: считаем unit-экономику ИИ-сервиса
B · Useful, limited A Cloud.ru product owner builds a transparent, reproducible cost model for a support agent, but the case is modelled from assumptions and the vendor's own price list, with no measured deployment.
Summary based on the full text·Original: about 17 minutes to read
In short
A Habr post (19 October 2026) by Timur Shakirbekov, senior product owner for the Cloud.ru marketplace. He argues that buyers and sellers of AI agents compare the wrong number. Using a modelled first-line support agent for an online shop, he counts the cost per "useful answer": a dialogue closed without a person, with no return on the same question within 72 hours and no negative rating. On that unit, model price nearly drops out; automation rate, error rate and handover quality decide the economics.
Key numbers
- 25x to 1.4xtoken cost gap between the cheapest and dearest model per dialogue, and what remains of it per useful answer (51 to 73 roubles)
- 5%share of tokens in the price of a useful answer on GigaChat-3-Pro (2.8 of 53 roubles)
- 71%share of escalations and wrong answers in the price of a useful answer
- ~3,500dialogues a month below which the modelled agent does not pay off on any model (about 7,000 with pessimistic fixed costs)
- 149 roublesassumed cost of a Moscow support ticket handled by a person, the benchmark the agent is compared with
As reported by the source.
Russian·Web page·opens habr.com
What it reports
- Modelled case: 40,000 contacts a month, 90% routed to the agent (36,000 dialogues), three LLM calls per dialogue, about 12,000 input and 800 output tokens, a guard model, RAG and order-status tool on Cloud.ru services.
- Token cost per dialogue ranges from 0.43 roubles (Qwen3-30B-A3B) to 10.99 roubles (Claude Sonnet 4.6), a 25x gap. Per useful answer the range is 51 to 73 roubles, a 1.4x gap.
- On GigaChat-3-Pro, escalations and wrong answers make up 71% of the price; tokens are 5%; all cloud infrastructure is 0.8 roubles. On Sonnet, tokens rise to 32%.
- Ten points of automation rate save about as much as dropping one model class (14 versus 15 roubles on Sonnet).
- Error share drives cost harder than tokens: 3% gives 45.0 roubles per useful answer, 15% gives 79.1.
- Handover time: at 3 extra operator minutes the answer costs 77 roubles, at 6 minutes 127, above a regional operator.
- A model 25x dearer in tokens pays back if it adds ten points of automation and two points of accuracy.
- Fixed costs of 163,000 roubles a month mean the agent does not pay off below about 3,500 dialogues a month (7,000 with pessimistic fixed costs).
- Pricing: Intercom Fin from $0.99 per resolution; Salesforce Agentforce launched at $2 per conversation, then $0.10 per action. Under outcome pricing, a vendor on Sonnet at 35% automation earns about zero margin.
How they got it
A linear spreadsheet model. Inputs are the Cloud.ru price lists of 28 September and 2 October 2026, an hh.ru salary figure for support staff, Ozon's published bot automation rates (27% in 2020, 62% in 2022), and Lorikeet's collection of 195 public vendor cases. The author lists every assumption, runs sensitivity checks and a pessimistic scenario, and gives the formula so readers can substitute their own figures.
Limits
- The case is modelled, not measured; the author says any of the four core inputs may differ by a factor of two.
- Answer quality is held constant across models in the main comparison.
- Prices and infrastructure come from the author's employer's catalogue, which also hosts the marketplace that would benefit from outcome-based pricing.
- Benefits such as peak-season scaling, night shifts and staff turnover are left out.
Why it matters for companies
This is a usable template for anyone approving an AI support or service agent: measure automation rate, 72-hour repeat contacts, handling time of escalated tickets and logged input tokens before negotiating model price. The model shows why a pilot can look cheap per token and still lose money at low volume or with cold handovers. It also warns against counting "freed" operators as savings while they remain on the payroll, a common error in pilot business cases.
Key passages in translation
Russian original
Две верхние строки, эскалации и ошибочные ответы, дают 71% цены. Ни та, ни другая не выглядят как «расходы на ИИ» ни в одном бюджете
The two top lines, escalations and wrong answers, make up 71% of the price. Neither looks like "AI spend" in any budget.
Russian original
Модель, которая дороже в 25 раз по токенам, в цене результата дороже на 40%.
A model that is 25 times dearer in tokens is 40% dearer in the price of the result.
Russian original
Что считать не надо, так это экономию от «освобожденных» операторов, пока они не освобождены в штатном расписании.
What you should not count is savings from "freed" operators while they have not been freed in the staffing plan.
Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com
- Tags
- Model cost