Business cases
How companies use AI to grow.
How Sberbank moved 60% of contact-center load to GigaChat in two years
First year: 47 million inquiries handled automatically — 60% of inbound traffic. Average response time dropped to 15 seconds (from 8 minutes). Contact center OpEx reduced by $120M annualized. NPS among customers whose queries were resolved by AI without escalation: 71 (higher than for live-agent calls at 64). Unexpected win: agent turnover fell from 40% to 19% — those who stayed handle interesting complex cases instead of "what's my balance". The biggest risk to close was financial hallucination. During Samara pilot in February 2024, the model confidently quoted a wrong deposit rate. The fix: every number affecting a financial decision must come from a function call to the source-of-truth system, never from generation.
Wildberries: how ML demand forecasting saved 12 billion rubles on returns
7-day forecast accuracy rose from 68% to 91% (MAPE). For electronics: 54% to 89%. Returns dropped 23% — customers get what they ordered because the item is in stock, and delivery takes 1.4 days vs 3.7. Warehouse turnover improved 35%: dead-stock capital fell from 23B to 14B rubles. Direct savings: 12 billion rubles annually — 7.3B from reduced return logistics + 4.7B from freed capital. Side effect: GMV grew 4.1% from pure availability (when items are in stock, conversion is higher). One key takeaway for the WB team: feature engineering matters more than architecture. The same CatBoost with the right features beat a custom transformer by 12pp accuracy.
Alice in smart homes: 41 million devices and the on-device model shift
Latency for smart-home commands dropped from 1.8s to 240ms — 7.5× faster. Smart-home CSAT rose from 6.8 to 9.1 on a 10-point scale. 99.7% of smart-home queries are now local — STT cost down 91%. That's $1.15M/month savings on STT alone, plus a similar amount on NLU. Unexpected effect: engagement growth. Users now use voice control 2.3× more often because "you don't have to wait". An active voice user brings the platform 3,700 rubles more per year (via subscription and plus services) than a touch-only user. Key problem: model updates. Twelve million devices across diverse firmware versions. The team built differential updates: new models roll out by user segment with rollback on p95-latency regression. Average update: 48MB, 90 seconds over WiFi.
T-Bank: how a graph neural network stopped 4.2 billion rubles of fraud per year
Direct fraud losses dropped 4.2 billion rubles per year (from 6.1B to 1.9B). False positive rate fell from 47% to 8% — customers no longer get "your transfer is paused" when sending money to grandma. NPS among customers with at least one alert rose from 31 to 64. The antifraud team shrank from 280 to 110 — the remaining handle complex cases with LLM explanations. Compensation payouts dropped 2.4×: the model catches attacks before money leaves the card, not after. Main production challenge: graph drift. Fraudsters adapt in 2-3 weeks: new patterns, yeast accounts, social engineering vectors. The team automated retraining: every week the model sees a subsampled prior week + newly confirmed fraud cases, and weights update.
Botkin.AI: how computer vision detects lung cancer 11 months earlier than the radiologist
Average diagnostic lead time: 11 months. From 4.7M processed scans, the system detected 2,380 stage-I lung cancer cases missed by human radiologists. Of these 2,380 — 88% survived 5+ years (vs expected 6% in late diagnosis). Mathematically: 1,950 lives saved. Economic effect for clinics: average stage-I lung cancer treatment in Russia costs 320K rubles (surgery), stage-IV costs 4.1M rubles (chemo + surgery + palliative). Early detection saves the public health system about 8 billion rubles per year across the network. The company obtained FSTEC certification and Roszdravnadzor registration (RU-2024-0481), allowing official inclusion of the AI tool in the clinical pathway. Since 2024, 11 regional OMS funds reimburse Botkin.AI usage at 380 rubles per screening — the first reimbursed medical AI in Russia.
Gazprom Neft: predictive maintenance cut drilling downtime by 41%
Total downtime across the 1,240-rig fleet dropped 41%. In absolute terms: 188,000 rig-operating-hours per year recovered. At 7.7M rubles average daily revenue per rig, that's +60 billion rubles of additional production per year. The largest ROI of any ML project in the Russian oil & gas sector. Emergency "between-plan" failures dropped from 68/year to 9 — the top downtime killers (bearings, pumps, top drive) are now predicted 200+ hours out. Parts spending decreased 12% (less overmaintenance), though the project initially feared the opposite — that predictive replacement would increase spend. In practice — replace surgically and less often. The biggest challenge: migrating engineering culture. Experienced rig managers drilling since the 1990s initially rejected model recommendations. "AI as intern" framing helped: a junior engineer on shift can cite "AI recommends" as a second opinion, giving political cover against a senior. After 14 months the culture flipped — a rig WITHOUT predictive maintenance now feels under-equipped.
Ozon: how a 1.2B-parameter transformer boosted GMV by 14%
GMV grew 14% — about 340 billion rubles annualized. Main-screen conversion rose from 4.1% to 6.8%. Average order value: +8% (the model better understands which up-sell items are relevant). Main-feed CTR: +31%. The most dramatic effect: cold start for new users. Previously, a new user's first purchase happened on average 12 days after registration (couldn't find what they wanted in "generally popular"). With PRISM — 4 days. Three times faster monetization of a new user. The contextual model revealed an interesting pattern: "evening sessions" (after 21:00) give 1.8× more conversions on personal recommendations than "morning" (before 12:00). In the morning, users goal-search for specific items; in the evening, they relax-browse and are receptive to recommendations. This changed push-notification marketing strategy: "evening picks" became the priority communication.
Rosatom: nuclear plant digital twin shaved 19 days off reactor refueling
Average PPR duration shrank from 53 days to 34 — 19 days saved. For 9 repairs in 2024 — 171 days of additional generation worth 38 billion rubles. Specialist utilization rose from 88% to 96% — same people do more work. Number of "cascading" delays (one operation slides 5+ others) dropped from 47 per PPR to 6. The 15-minute re-planner doesn't let slippages accumulate — they're isolated and handled locally. Side effect: safety. The anomaly detector with 12-minute reaction prevented two potentially dangerous incidents in 2024: a crew started disassembling one auxiliary circuit without verifying the adjacent one was depressurized. Alert fired in 67 seconds, site stopped in 4 minutes. Without AI this could have escalated. Key implementation challenge: culture. Brigadiers with 30 years of experience don't like "the computer" reordering their work. The team built an important frame: "the computer doesn't order, it offers alternatives — the brigadier decides". Recommendation acceptance grew from 23% (first 3 months) to 81% (after 14 months of operation).
M.Video-Eldorado: computer vision across 1,080 stores lifted conversion by 23%
Network-wide conversion rose from 14% to 17.2% in the first year. Top factor: consultant schedule optimization. Before: 3 people in the washing-machine zone scheduled 10:00-18:00. Now: 5 in peak hours (Wed 17-20, Sat 13-18) and 1 in dips (Tue 11-13). Customer waiting time fell 47% — visitors who don't get a consultant in time leave without buying in 73% of cases. Real-time incident detection revealed three things: 14% of thefts are internal (staff), unexpected scale; 31% of "lost" expensive items (left unattended) were cases where the consultant got absorbed with one client and forgot another; checkout conflicts — 22 cases/day on average, of which 4 escalating (previously only escalating ones were tracked). Key challenge: privacy compliance. The team worked with Roskomnadzor for 7 months on approval: face blurring on edge before frames leave anywhere, no direct identification, anonymous session IDs reset every 4 hours, all "events" auto-deleted after 30 days. All documented in a standard now used as reference for the industry.
Russian Post: OCR + AI router cut parcel processing time by 38%
Average parcel processing time at a sorting center dropped from 47 to 29 seconds (-38%). Address-reading accuracy rose from 96.3% (manual entry) to 99.4% (OCR). Manual-sort operator headcount cut from 8,200 to 2,300 — the remaining work only on hard cases the ensemble couldn't decode. Key metric: % of parcels delivered within SLA. Before: 81%. After: 96%. This restored user trust in the state post: in 2024 Russian Post NPS turned positive for the first time in a decade (+14, was -23 in 2022). The December 2024 crisis went almost unnoticed: peak load processed by the same capacity, no queues. Project economics: 4.7 billion rubles invested (hardware + development + integration). Direct savings: 8.1 billion rubles per year from reduced manual work. ROI payback in 7 months. Additional benefit: "address not found" returns dropped, cutting logistics costs by another 1.2 billion rubles.