2026-09-30 — Ping An Bank becomes the first listed Chinese lender to adopt formal AI rules

• Ping An Bank became the first listed Chinese lender to formally adopt rules governing its use of artificial intelligence, after its board approved AI management measures in a Wednesday filing The Shenzhen-listed bank is part of Ping An Group; SCMP analysts expect other mainland banks to follow as regulators tighten oversight of AI use, and the full rules have not been made public. (South China Morning Post)

2026-09-23 — Ping An’s CTO says daily AI token use is up tenfold in a year while cost per token fell 76%

• Ping An’s chief technology officer Ray Wang told Alibaba’s Apsara Conference that the group’s average daily AI token consumption rose from 30 billion at the end of last year to more than 300 billion, on roughly doubled computing capacity (800 to 1,500 petaflops), with cost per token down 76%. He said full-stack optimization raised token output per unit of computing power by 368%; that its AI-powered “Express Service” has handled over 100 million user interactions in four months with a 92% end-to-end resolution rate; that AI-generated code is about 80% of total code output; and that its AI Doctor’s hallucination rate is under 0.3%. These are company-reported figures from a vendor-conference talk. (Ping An press release via PR Newswire)

2026-08-21 — A critical read of the H1 results: new-business growth trails peers, and an unverifiable AI health-model claim surfaces

• Insurance Business Magazine’s analysis of the H1 2026 results found Ping An’s Q1 new-business-value (NBV) growth of 20.8% “trailed several domestic peers by a wide margin” — China Life posted 75.5%, with some peers exceeding 70%. This is the first outside-analyst pushback this thread has logged against a set of figures that, until now, have come almost entirely from Ping An’s own filings and press releases.
• The same analysis flags a new, unverifiable AI claim: Ping An says its “Medical LLM 3.5” achieved a top score on HealthBench Hard, but “public trackers…list no entry for a Ping An model,” meaning the claim can’t currently be checked against any public leaderboard — a different, not-yet-logged model from the PingAnGPT-Qwen3-32B/CNFinBench claim already on this thread, and a direct instance of the “flagged, not yet resolved” pattern this thread already carries (the AI Family Doctor user-count discrepancy). This is the MH-adjacent thread of Ping An’s AI story getting scrutinized rather than repeated — a health-specific model claim the company has not backed with checkable evidence. (Insurance Business)

2026-08-20 — The interim report refines the automation percentages, and adds two bigger numbers

• Ping An’s official 1H 2026 interim report restates the automation figures this thread is built on, slightly down on both: 59% of life claims settled through “quick claim” (against ~60% on file) and 94% of auto policies sold through the dealer channel intelligently issued in under a minute on average (against ~93%). The movement is noise; the point is that these are now company-filed rather than inferred. (Ping An 1H 2026 results)
• Two figures are genuinely new and larger than anything this thread had: smart fraud detection avoided RMB 7.11bn in claims payouts, up 10.4% YoY, and AI service agents handled ~939 million interactions — 81% of all customer-service volume — in H1. An 81% share moves this from “AI assists the service desk” to “the service desk is AI, with a human exception path”, which is the scale at which the NFRA framework’s AI-decision liability question stops being theoretical. (Ping An 1H 2026 results)

2026-08-20 — Group-level H1 results land, with an AI number of a different shape

• Ping An Insurance (Group) Company of China (HKEX:2318/SSE:601318, the parent) reported H1 2026 on its confirmed 08-20 date: revenue RMB 615.4bn (+12.6% YoY), net profit RMB 92.6bn (+36.1% YoY), interim dividend RMB 0.98/share. No group-level profit-share AI figure comparable to the subsidiary’s 4.6%, but real activity-volume metrics: average daily token consumption exceeded 120bn in June 2026 (up from 30bn in December 2025), AI service reps handled ~939mn interactions (81% of total customer-service volume), AI-assisted sales reached RMB 57,313mn in H1, fraud-detection AI saved RMB 7.11bn in claims (+10.4% YoY), and 88% of business scenarios are now AI-enabled. (PR Newswire, Manila Times)

2026-08-18 — H1 results land on the due date, with an AI number attached

• Ping An Healthcare and Technology / “Ping An Good Doctor” (1833.HK) reported H1 2026 on its confirmed board-meeting date. Total revenue RMB 2.484bn; net profit attributable to shareholders RMB 219mn, +63.5% YoY; adjusted net profit RMB 227mn, +37.7%. B-end corporate health-management revenue RMB 714mn, +65.1% YoY, now 28.7% of total — the enterprise channel is where the growth is. (Sina Finance)
• The disclosure this thread’s ledger entry was explicitly holding for exists: AI contributed ~4.6% of gross profit (AI业务贡献毛利占比约4.6%), alongside 9.7mn+ cumulative AI-doctor users and AI-medical-pay network coverage at 149,000 of 245,000 partner pharmacies. A rare case of a health company putting a hard number on AI’s contribution to the P&L rather than describing a strategy. (Sina Finance)
• This is the SUBSIDIARY only. The group parent (HK:2318 / SH:601318) reports separately on 2026-08-20, now pinned by the company’s own voluntary HKEX/SSE announcement filed 08-13 — closing the “group-level date unconfirmed” gap this thread carried since 08-04. Watch for a group-level AI-contribution figure to set beside the 4.6%.
• ⚠️ Trap logged, and it is a nasty one for this issuer specifically: English-language wire mirrors (PRNewswire, Barchart) served 2026-dated URLs and headlines ("…reports strong interim results…AI fuels new growth momentum") whose BODY was the August 19 2025 H1-2025 release, prior-year numbers and all — RMB2.5bn revenue, RMB134mn profit, dateline “HONG KONG, Aug. 19, 2025.” The genuine 2026 figures were reachable only via primary Chinese-language financial press. Do not accept an English wire mirror for Ping An without checking the dateline inside the body.

2026-08-04 — Thread opened + crawled

• Opened following a board-pass audit that found the entire insurance pocket with zero live thread coverage — Ping An was flagged in its own board gloss as “the strongest insurtech-AI angle (240M-customer ecosystem)” among the pocket’s seven kingdom-rank insurers, and this crawl confirms that flag with hard, dated, primary-sourced numbers.
• Ping An’s WAIC 2026 press release (2026-07-21, fetched and confirmed directly): the Intelligent Underwriting System cut average initial underwriting review time to ~1.5 hours, doubled daily processing volume per underwriter, and raised the risk-interception rate by 16%; an “Intelligent Policy Issuance Robot” now auto-processes 93% of new vehicle-insurance policies (turnaround 6 minutes -> 1.2 minutes); Ping An P&C claims 100% AI coverage across core claims scenarios with an 80% efficiency gain. (Ping An)
• Insurance Journal (2026-03-31): nearly 60% of accident/health claims are now automated (up from almost none five years ago), some settling in 51 seconds; AI handled 70% of Ping An Bank’s RMB 500B in 2025 loan recoveries; the auto-insurance expense ratio fell 1.7 points over nine years (~RMB 5B in added underwriting profit); headcount fell 118,000+ (~30% from its 2018 peak). CTO Ray Wang: “The AI era successfully opened the window for reshaping services. The returns on investment are tangible, highly visible, and unequivocally compelling.” (Insurance Journal)
• PingAnGPT-Qwen3-32B, Ping An’s self-developed financial LLM, ranked #1 on the CNFinBench leaderboard — ahead of GPT-4o, Claude Sonnet 4, and DeepSeek-R1 (671B) — on financial reasoning, Q&A, and compliance/risk- control metrics; deployed across 97 real business scenarios including auto-insurance claims. (Ping An, 2026-03-13)
• China’s NFRA issued sector-wide AI guidance for banking/insurance, dated 2026-06-18 (金发〔2026〕8号): explicitly covers AI in underwriting, claims, and pricing; establishes that liability for AI-assisted decisions rests with the insurer, not the algorithm; requires human- oversight mechanisms for high-risk applications and board-level AI governance. No enforcement action naming Ping An specifically has surfaced — this is the framework Ping An now operates under, the rule of the road, not evidence of scrutiny. (Sina Finance)
• Flagged, not yet resolved: Ping An’s “AI Family Doctor” user count is reported inconsistently across sources (90M monthly active users in the WAIC release vs. 12M annual users elsewhere) — worth reconciling against a single primary figure before it’s cited as a headline number.

← Backstory