AI Atlas Weekly Report — 2026 Week 35
Use Case Highlights
Trustly, the Swedish pay-by-bank payment processor serving 110 million consumers in 30+ countries, migrated from a legacy BI tool to Amazon Redshift + Amazon QuickSight + Amazon Q in QuickSight. The new analytics stack scaled users from 80 to 300, generates insights 4x faster, and costs 50% less to maintain. Over 6 months users across risk, data science, executive, product, and sales teams submitted 15,000 natural-language questions via Amazon Q with 89% accuracy.
Guangdong Provincial Health Commission launched '粤医智影' (Yueyi Zhizhi), an AI-assisted imaging reading system based on the provincial telemedicine platform, covering 2,093 institutions. 7 exam types (CT lung nodules, pneumonia, rib fractures, DR TB, extremity fractures, coronary calcification). CT re
AIG deployed multi-agent AI underwriting (AIG Assist) across 8 lines of business, achieving 30% quoting lift, 55% time-to-quote reduction, and 40% binding improvement, with underwriting timelines compressed from 3-4 weeks to under 1 day.
At Owens Corning's Tessenderlo plant in Belgium, Waites' AI-driven predictive maintenance system detected a critical temperature spike on the non-drive-end main bearing of the 40-year-old Ball Mill #5 on February 26, 2024, preventing an unplanned shutdown that would have faced a 17-week lead time for a replacement end shaft. The company avoided over $11.24 million in potential production losses, repairs, and downtime — a 21,600% ROI — and has since expanded wireless vibration and temperature mon
Ingredion Adopts AI to Detect Failures and Boost Machine Uptime CASE STUDY # Ingredion Adopts AI to Detect Failures and Boost Machine Uptime How real-time condition monitoring helped Ingredion avoid critical downtime and save over $1M at just one plant. ## Ingredion Adopts AI to Detect Failures
Trends
Health care at scale — provincial rollouts beat single-site pilots
Health Care Providers & Services led the week with 18 new use cases. The standout is Guangdong Provincial Health Commission's 'Yueyi Zhizhi' AI imaging reading system rolled out across 2,093 medical institutions — a single deployment covering more sites than the combined total of every other healthcare insert this week. Provincial and national-scale healthcare AI is replacing single-hospital pilots as the dominant deployment pattern.
Banking and insurance cross 19 deployments — agentic underwriting goes production
Banks (10) and Insurance (9) together contributed 19 of the week's 130 use cases. Highlights: Trustly 4x faster insights via Amazon QuickSight+Redshift, AIG 40% binding lift with multi-agent AI underwriting across 8 lines, and SIGNAL IDUNA's German cloud migration. Agentic underwriting has moved from POC to production-grade at top-tier carriers.
Agentic + generative continues to dominate content signals
16 use cases mention 'agentic', 15 'generative', and 11 'autonomous' — the highest cross-week concentration since the W31 report. Production deployments now routinely ship multi-agent workflows (Fluz payments, Betsson France 75% auto-resolution, Deutsche Telekom Merlin AI copywriter) instead of single-LLM chatbots. The vocabulary shift is no longer marketing — it's reflected in the operational details of the inserts.
Predictive maintenance breaks out as the top industrial pattern
Owens Corning's $11M loss prevention at the Belgian Tessenderlo plant via Waites AI predictive maintenance, Ingredion's AI failure detection for machine uptime, Foshan Power Supply Bureau saving 96,000 hours/year, Arnotts cutting unplanned downtime with Factory AI, and Unilever Indaiatuba cutting maintenance costs 45% — 5 predictive-maintenance anchors this week versus 2-3 in prior weeks. Industrial AI is the clearest 'reactive → predictive' shift in the current data set.
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 |
|---|---|---|
| AI deployment case study 'in production' -tutorial -course | High | Returned Trustly, Owens Corning, Ingredion — strong metric-bearing hits this week |
| generative AI 'reduced by' OR 'cut by' enterprise case study | High | Anchored AIG, Banorte, DPD Switzerland 60%+ time reduction stories |
| AI agent underwriting insurance production | High | Drove AIG, SIGNAL IDUNA, Tokio Marine fraud detection inserts |
| predictive maintenance AI plant factory | Medium | 5 hits this week vs 2-3 prior weeks — break-out pattern |
| China hospital AI imaging national provincial rollout | High | Captured Guangdong 2,093 institutions, Zhengzhou 350K+ exams, DAMO GRAPE, Tiantan CT |
New Queries Added
Data Quality
Records with content <500 chars (insufficient for publication-grade detail)
Auto-flagged; require manual expansion before promotion to published (per Pitfall #20, no auto-archive by cron).
Records with Unknown country (geographic coverage gap)
Heuristic re-extraction attempted during Step 3; 18 of 25 still unresolved and carry forward to next week's QC.
Records with Unknown continent (downstream from country gap)
Resolved automatically once country is filled in; will re-run after next QC pass.
Aug 28 pipeline run missed entirely (zero inserts)
Root cause not yet diagnosed; flagged in systemHealth. Likely cron executor issue, not search-pipeline issue.
Pipeline metrics JSONL logging gap (4 days)
DB has inserts for Aug 27, 29, 30 but no metrics entries; flagged as warning in systemHealth.
Observations
Pipeline metrics JSONL observability regressed: only 3 of 7 days this week have metrics entries (Aug 24/25/26), yet DB inserts appear on Aug 24/25/26/27/29/30. Aug 28 had zero inserts — full pipeline miss. The 'thisWeek' totals above undercount true activity by ~30-50%.
Average rejection rate climbed to 72.5% (vs 68.8% last week), driven by fetch_fail becoming the top rejection reason (19 hits). Ollama was the sole active search engine this week (30 of 32 searches); the rotation to Tavily/Exa that rescued the W34 quota did not recur.
Pipeline output per logged run improved: avg validated per run rose to 11.7 (from 9.5 last week) and avg new use cases per run rose to 8.0 (from 7.3). Manual curation (18 hits) remains the second-largest rejection category — the carry-over KPI-only rule is still not enforced.
Industry diversity held: 45 distinct industries across 130 inserts, with Health Care Providers & Services (18) and Software & Services (14) leading but no single industry exceeding 14% share. 6 continents represented. Coverage profile matches the W34/W33 mix.
Next Steps
- 1highskills/ai-atlas-data-quality-check/SKILL.md
8th consecutive carry-over (W24→W35): pre-insert validation still not enforced. This week's 5 short-content records and 25 Unknown-country records would have been caught by the W24-proposed Step 3.0 rules.
Promote the W24 pre-insert rules into ai-atlas-data-quality-check/SKILL.md Step 3.0: (1) reject if content is null/empty/<500 chars; (2) reject if country is Unknown AND continent is Unknown; (3) reject if industry==Unclassified AND description.length<500; (4) require non-empty summary for new UCs; (5) default published_at to created_at when null; (6) reject if content lacks business-process OR technical-architecture detail (per 2026-08-23 ERRORS.md).
- 2highskills/ai-atlas-data-quality-check/SKILL.md
5th consecutive carry-over (W31→W35): Tier-1 ghost company audit still not implemented. The 16 template-fill companies carried in 8/18 batch (Pitfall #42) and 4 ghost-co archived on 8/24 confirm the audit is needed.
Add a Tier-1 ghost audit rule to ai-atlas-data-quality-check/SKILL.md: for each company, count archived vs published UCs; if all UCs are archived AND the company name matches an anonymized/client/customer pattern from prior batches (e.g., '(Anonymous)', 'Customer', 'Client', 'SME', 'Consortium', '(with X)', '(formerly Known as)'), auto-archive the company and surface in ERRORS.md.
- 3mediumskills/daily-ai-push-v2/SKILL.md
NEW for W35: pipeline_metrics.jsonl logging has a 4-day gap this week (Aug 27, 28, 29, 30 missing or partial). Aug 28 had zero inserts. The metrics append step is silently dropping writes — likely stdout buffering or exception path.
Add a try/except wrapper around the metrics append in daily-ai-push-v2/SKILL.md with explicit fsync and a fallback jsonl-append to /tmp/pipeline_metrics.jsonl AND a mirror copy to /Users/clawclaw/.openclaw/workspace-ai-atlas/tmp/. Add a cron-start heartbeat file written before search begins, so missing-run days are detectable even when the run crashes mid-execution.