AI Atlas Weekly Report — 2026 Week 31
Use Case Highlights
SK Telecom, Korea's #1 customer service provider for 27 consecutive years (NCSI), deployed Claude in Amazon Bedrock to power customer support for millions of Koreans. The deployment lifted in-call LLM response quality ratings by 34%, cut low-quality responses in SKT's Telco fine-tuned model by 68% versus the base, and reached ~89% of human agent quality in post-call processing — combining SKT's in-house RAG with Claude for end-to-end retrieval-augmented generation.
Insilico Medicine's Rentosertib — the first antifibrotic small-molecule TNIK inhibitor, with both target and structure generated by generative AI — showed +98.4 mL mean change in forced vital capacity versus -20.3 mL on placebo at 12 weeks in the GENESIS-IPF Phase IIa trial (n=71), published in Nature Medicine. Patients on 60 mg QD not on concomitant antifibrotic therapy showed +187.8 mL. The first randomized clinical proof-of-concept that an AI-discovered target and AI-designed compound can del
Japan's leading DIY home improvement chain Cainz deployed Google Cloud's Vertex AI Forecast across 209 stores, cutting sales data preprocessing time to 50 minutes regardless of store count and replacing a fixed-order-quantity system that couldn't predict seasonal or short-term demand. The architecture — Vertex AI Forecast + Cloud Run jobs + BigQuery + Workflows — was refined twice through Google Cloud's Tech Acceleration Program (TAP) and scaled from a flagship pilot in Isesaki (June 2022) to al
Bayer China built Dr. Bei, an LLM-based AI training system for pharmaceutical medical representatives running on Claude (Amazon Bedrock), going from concept to launch in 72 days and scaling to 1,000+ users within 6 months — saving 6,000–7,000 hours of training time per year and lifting 1-on-1 role-play coaching efficiency by over 50%. The system is now rolling out globally, with deployment underway in Europe and additional Asia regions.
Latvian fintech Sun Finance (4M evaluations/month across 9 countries) rebuilt its identity verification pipeline with the AWS Generative AI Innovation Center. The new solution, built on Amazon Bedrock (Claude Sonnet 4), Textract, and Rekognition, pushed extraction accuracy from 79.7% to 90.8%, cut per-document costs by 91%, and reduced processing time from up to 20 hours to under 5 seconds — while adding a parallel visual-pattern + background-similarity fraud detection layer.
Trends
Insurance AI: from chatbot pilots to claims-automation scale
Insurance emerged as the top industry this week with 15 use cases (12.4% of the 121 ingested), spanning claims triage (Tencent WeBao 40% efficiency / 98% accuracy), underwriting automation (AXA, China Pacific), customer service (Groupama virtual assistant), and fraud detection. The category has clearly moved past FAQ-bot pilots to production-grade document/claims automation with quantified ROI.
Industrial AI quality inspection at scale — Foxconn, CIMC, Huaxiang/Taisteel
Three high-confidence China-industrial deployments landed on Aug 2: Foxconn's 2,000 edge-AI-box deployment cut defect rate from 0.8% to 0.15%; CIMC Ningbo (500+ robots + AI vision in container production) was named a National Excellent Smart Factory; Huaxiang Flange Factory and Taisteel Stainless Steel both deployed AI vision on production lines. Together with 14 Industrial Machinery cases this week (the second-largest industry slice), Chinese heavy-industry AI is now operating at line-rate scale, not pilot scale.
Generative AI in life sciences: first randomized Phase 2 efficacy from an AI-designed drug
Insilico Medicine's Rentosertib — the first AI-discovered target AND AI-designed compound — showed +98.4 mL FVC improvement vs -20.3 mL on placebo in the GENESIS-IPF Phase IIa trial published in Nature Medicine, with patients not on background antifibrotics reaching +187.8 mL. This is the first randomized clinical proof that an AI-discovered drug can deliver clinical efficacy, marking a category-defining milestone beyond deployment telemetry.
Anthropic Claude dominated enterprise model choice this week
Of the 42 provider mentions in this week's content, Anthropic Claude led with 14 (33%), including SK Telecom's national-scale customer support deployment (lifted in-call LLM quality by 34%, reached 89% of human-agent quality post-call), Bayer China's Dr. Bei training system (1,000+ users, 6,000–7,000 hours saved/year), Sun Finance's ID-verification pipeline (extraction accuracy 79.7%→90.8%), and Rakuten's Claude Code adoption. OpenAI came in second with 11 mentions.
US deployments still dominate, but China accelerated sharply
49 of 121 use cases (40.5%) were US-based, with the UK (8), France (6), India (6), Canada (5), Japan (4), Italy (4), South Korea (3), Ireland (3) rounding out the top 10. China contributed 13 — driven by industrial AI (Foxconn, CIMC, Huaxiang, Sinochem DeepSeek), insurance (Tencent WeBao, China Pacific, Ping An), and pharmaceuticals (Bayer China's Dr. Bei, Intsig Claim Agent). China's 13 is its highest single-week share since W22.
Under the Hood
Pipeline and operations behind the headline numbers—search tuning, data checks, internal notes, next steps, and linked case IDs.
Search Strategy
Query Performance
| Query | Hit | Notes |
|---|---|---|
| Insurance claims automation production deployment case study | High | Yielded 15 insurance cases this week — top industry. Groupama, China Pacific, Tencent WeBao, AXA all surfaced. |
| Industrial machinery AI quality inspection vision production line | High | 14 industrial cases including Foxconn 2000-edge-AI-box deployment, CIMC Ningbo, Huaxiang/Taisteel. |
| AI drug discovery Phase II Nature Medicine published clinical trial | Medium | Found Insilico GENESIS-IPF Phase IIa; broader pharma AI queries still mostly yield pipeline announcements. |
| Claude Amazon Bedrock enterprise customer support deployment | High | SK Telecom (Korea), Bayer China (Dr. Bei), Sun Finance (Latvia), NRI (Japan) — Anthropic Claude clearly dominates enterprise model choice. |
| Vertex AI demand forecasting retail inventory production | Medium | Cainz (Japan) is the standout; other Vertex AI cases were smaller-scale. |
| China industrial AI smart factory national designation | High | CIMC Ningbo (National Excellent Smart Factory), Foxconn, Huaxiang — Chinese government designation drives newsroom coverage. |
New Queries Added
Data Quality
Records with content <500 chars (first-pass quality gate failure)
Auto-flagged for manual review; not auto-archived since both have valid company linkage.
Use cases inserted in `status=pending` awaiting Ran review (carry-overs + new)
Per Pitfall #20 NO AUTOMATIC STATUS PATCHES — 5 carry-overs (Bank of Ireland, KPTCL, Rexel Canada, Tencent WeBao, Ifan Ifans) + 3 new (Ramp GPT-5.5, Foxconn, CIMC Ningbo, Huaxiang/Taisteel). Need explicit criteria for auto-flip.
Tier-1 ghost companies (active but ALL UCs archived — anonymized/template-fill pattern)
6 of 36 archived this week (DreamzTech Client, ASTOLA Customer, BinaryBits Customer, CX Platform User, UK Logistics SME, National Telecom). 30 awaiting Ran decision — recommend batch archive.
Missing website_url (carry-over from W30)
Ifan Ifans (Welsh dairy farm 3d2a1ec2) — url=''. Carried forward to next week.
Records with zero coordinates (0,0) — geographic signal lost
Tracked but not blocking — most are real deployments whose source pages don't list HQ coords. Lower priority than content/coverage fixes.
Observations
Pipeline recovery from 22-day outage (Jul 5 → Jul 26) is complete: 13 successful runs this week, 121 use cases ingested, 93 new companies. Avg rejection rate improved to 54.98% (from 80%+ plateau in W23-W26 and 68.22% in the partial recovery week W30).
Anthropic Claude dominated enterprise model choice this week with 14 mentions (33% of provider mentions) — SK Telecom (national-scale customer support, +34% LLM quality), Bayer China (Dr. Bei training system, 1,000+ users), Sun Finance (Latvia, ID extraction 79.7%→90.8%), Rakuten (Claude Code), NRI (Japan).
Insilico Medicine's Rentosertib Phase IIa publication in Nature Medicine is a category-defining milestone: first AI-discovered target + AI-designed compound to deliver positive randomized Phase 2 efficacy (+98.4 mL FVC vs -20.3 mL placebo). Strongest signal yet that generative-AI drug discovery has graduated from pipeline to clinical validation.
Pending UC backlog grew from 4 (Aug 1) to 8 (Aug 2): 5 carry-overs plus 3 new (Ramp GPT-5.5, Foxconn, CIMC/Huaxiang). Per #20 no auto-flip, but auto-flip criteria should be documented to avoid queue stagnation. Last week's search totals are not comparable (only 2 runs due to pipeline recovery).
Next Steps
- 1highskills/ai-atlas-data-quality-check/SKILL.md (Tier-1 ghost company audit, new subsection)
36 Tier-1 ghost companies flagged on Aug 2 — companies `status=active` but ALL their use cases are archived (anonymized/template-fill pattern: "UK Logistics SME (Anonymous)", "ASTOLA Customer", "DreamzTech Client", etc.). 6 of 36 already archived this week; 30 awaiting Ran decision.
Add to ai-atlas-data-quality-check/SKILL.md: (1) a Tier-1 ghost rule that counts `archived` vs `published` UCs per company; (2) if ALL UCs are archived AND company name matches anonymized/client/customer regex (or contains "(Anonymous)"), auto-flag for archive; (3) include a `qc_tier1_ghost_audit.py` script. Reduces dashboard noise and prevents the ghost companies from inflating active-company counts.
- 2highskills/ai-atlas-data-quality-check/SKILL.md (pre-insert validation Step 3.0)
Carry-over from W24/W25/W26. Pre-insert validation still not enforced: 2026-07-04 incident showed 5 pending UCs (from 2026-07-03 23:37 run) had 7+ data quality issues at insertion (Atlantic Health case alone had 5 broken fields). 8 pending UCs now sit in the queue with similar issues.
Implement the pre-insert rules proposed in W24: (1) reject if `content` null/empty/<500 chars; (2) reject if `latitude==0 AND longitude==0`; (3) reject if `industry=='Unclassified' AND description.length<500`; (4) require non-empty `summary` for new UCs. Add `qc_pre_insert_block` Python helper. Carry-over unresolved for the 4th consecutive week — needs hard commitment.
- 3mediumskills/daily-ai-push-v2/SKILL.md, Step 1 (search layer) — Ollama health-check
Ollama produced 0 searches in 5 of 13 runs this week despite being the prior default. tavily-remote MCP and Exa MCP carried the workload (131 of 202 = 65% of searches). This may indicate Ollama's experimental web_search endpoint is degraded or rate-limited.
Add a Step 0 health-check to daily-ai-push-v2/SKILL.md: smoke-test Ollama with 1 query before the run starts; if it returns 0 results OR errors, log a warning and rebalance layer weights toward Exa/Tavily. Prevents silent dependency on a degraded local tool.