---
id: "lib-2026-10-26-how-the-tri-sestry-clinic-adopted-ai-ai"
type: "library"
title: "How the Tri Sestry clinic adopted AI: AI supervision and quality control"
title_original: "Как клиника Три сестры внедрила ИИ: ИИ-супервизия, контроль качества"
lang: "RU"
org: "Tri Sestry clinic (Purrweb)"
source_type: "Other"
date: "2026-10-26"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-26-how-the-tri-sestry-clinic-adopted-ai-ai/"
original_url: "https://www.youtube.com/watch?v=0TPwt0gHCMM"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "The medical director of a JCI-accredited Russian rehab clinic describes a concrete record-audit automation and two named failures; all figures are his own, with no measured effect of the AI tools, on a development agency's sales webinar."
summary_basis: "transcript"
original_minutes: "60"
tags: ["business transformation", "healthcare", "record audit", "data de identification"]
frames: ["verification-capacity"]
---

# How the Tri Sestry clinic adopted AI: AI supervision and quality control

Original title: Как клиника Три сестры внедрила ИИ: ИИ-супервизия, контроль качества

Tier B (Useful, limited): The medical director of a JCI-accredited Russian rehab clinic describes a concrete record-audit automation and two named failures; all figures are his own, with no measured effect of the AI tools, on a development agency's sales webinar.

Summary based on the full transcript.

## In short

A 60-minute Russian webinar (26 October 2026) hosted by Sergey Ponomarev, CTO of the Purrweb development agency, with Anton Klochkov, medical director of the Tri Sestry early-rehabilitation clinic in the Moscow region (about 400 staff, 120 beds, JCI-accredited in 2025). The main case is his own monthly audit of patient records. He taught a model the clinic's SMART-goal method by scoring about 400 goals alongside it for several months, then let it run the check. The model now flags weak goals, banned abbreviations and copy-pasted daily notes. Fall detection by computer vision failed, and a scheduler built into the clinic's records system goes unused because of poor usability.

## Key numbers

- ~12 hours: medical director's monthly manual review of rehab goals before the audit was handed to an AI model
- ~400: rehab goals scored in parallel by the director and the model over several months before automation
- 3,000 to 5,000: incidents reported per year at the clinic
- ~46 million roubles: losses from patients leaving treatment early in 2025; early departures since fell 2.5 times

## What it reports

- Audit of rehab goals against SMART criteria took him about 12 hours a month (one point per criterion, five-point scale). He gave the method to one model, which wrote the prompt for a second; both scored goals, and the model asked him about every disagreement. After several months and about 400 goals the check was automated.
- The live audit checks team-meeting protocols, short and long-term goals, abbreviations (more than five outside the approved list costs a star; the demo record had up to ten) and note uniqueness as "anti-plagiarism", for example a respiratory rate of 16 copied for six months. It also checks goals against the occupational-therapy interview (COPM) and suggests rewrites. Next step: feedback during the team meeting itself.
- Field checks run on DeepSeek on a server rented from the records-system vendor under a non-disclosure agreement. An earlier in-house assistant, "Filantrop", ran on ChatGPT and was trained on clinic policies, SOPs and clinical guidelines. ChatGPT is used on de-identified data (record number only, matched later in Excel). He calls de-identification a main drag on AI in medicine.
- The clinic code says each employee should use technology and AI to automate their work and leave to people only what only a person can do. There is no central automation department.
- Clinical debriefs after deterioration, falls or emergency transfers are recorded and transcribed; a second "devil's advocate" role in the prompt looks for gaps in the doctors' reasoning.
- Failures: computer vision for fall detection found no trainable pattern (falls happen sliding along walls or during transfers) and too few examples. The scheduler in the records system is unused ("an extra click is a failure"), so the clinic paid another firm for a custom one.
- Money: in 2024, about 200 missed rehab days in a month. Early departures cost about 46 million roubles in 2025 and have fallen 2.5 times since staff began tracking red flags, "including with technological solutions".
- KPI dashboards (falls, hospital infections, stoma removals, patients who started walking, average bill) are shown on screens in corridors and the restaurant for patients.

## How they got it

A hosted webinar with screen-shared demos of the audit output and dashboards. The guest describes his own clinic; no before-and-after measures of the AI audit itself (accuracy, time saved after automation, agreement rate) are given. The digest is based on cleaned Russian auto-captions in which the guest's surname and the records-system vendor's name are garbled.

## Limits

- Self-reported by the medical director; no outside verification.
- The 2.5-times drop in early departures and the 46 million rouble loss are tied to red-flag tracking in general, not to AI.
- No tool costs and no error rates for the model's scoring.
- The host's agency inserts a sales pitch for its own services and promotes its community, so the event is lead generation.
- Auto-captions garble names and some terms.

## Why it matters for companies

The calibration method is the transferable part: an expert scores the same items as the model for months, the model asks about each disagreement, and the check is automated only once scores align. It is a cheap way to turn one person's judgement into a repeatable audit. The failures are equally useful: computer vision stalled on too few, too varied examples, and a built-in feature lost to a custom tool on clicks alone. Data residency was solved by renting a server from an existing vendor under contract.

## Key passages in translation

> Каждый сотрудник старается использовать современные технологии и искусственный интеллект для автоматизации своей работы. Мы оставляем за человеком только то, что может делать только человек.

Every employee tries to use modern technology and artificial intelligence to automate their work. We leave to people only what only a person can do.

> Потом он сравнивал свою оценку с моей и задавал вопросы, если они отличаются, что я оценил не так, почему, в чём мы расходимся.

Then it compared its score with mine and asked questions when they differed: what I had scored differently, why, where we disagreed.

> Лишний клик, всё это провал. Это значит, что это не будет использовать.

An extra click, and it is a failure. It means nobody will use it.

Original: https://www.youtube.com/watch?v=0TPwt0gHCMM