Key points
On September 17, Moonshot AI released a financial-industry AI solution built around three pillars: 10+ authoritative data sources, 9 packaged financial skills, and 5 compliance/security measures. First-wave institutional users include ICBC, CSC Financial (CITIC Construction Investment Securities), CICC, E Fund, Ant Group, and Meituan; on the primary-market side, Sequoia China, ZhenFund, and IDG Capital.
Nine skills map to nine job types
| Skill | Target role | Published efficiency data | |---|---|---| | Institutional financial modeling | IB / PE | 5–7 person-days → 0.5–1 | | Institutional research reports | Sell-side / secondary research | 10–20 days → ~2 days | | Institutional PPT | Insurance / IC decks | 7–10 person-days → ~2 | | Financial dynamic charts | Research marketing | CAGR and delta annotations in-chart | | Earnings commentary | Earnings-season output | Internal benchmark score +8–9 points | | Consensus map | Sell-side expectation tracking | — | | Portfolio review | Fund managers | — | | Holdings morning note | Morning meetings | — | | HK IPO lens | HKEX filings tracking | Scans filing, hearing, listing status |
The three skills with public numbers all land in the "cut labor by more than half" range; financial modeling shows the largest compression. Deliverables are framed as "adjustable" Excel models and comparable-company valuation tables — the tool replaces the drafting stage, not the judgment stage. Internal benchmarks report earnings-commentary scores improving 8–9 points and financial modeling rising from 69 to 77 (vs. Opus 5's 75); scoring criteria and sample size were not disclosed.
Data sources wired to MCP, with traceable numbers
Data list: Wind, Eastmoney, S&P Global, Caixin,.cls (Cailianshe), GilData, Tianyancha, Dun & Bradstreet, Crunchbase, THS iFinD; macro side connects IMF, World Bank, and the Fed's FRED. Data enters research tasks via MCP; Scholar full-text search, Word/Excel/PPT output tools, browser extensions, and Computer Use round out the stack. The key claim: every critical figure can be traced back to its original source — answering the finance industry's first question about AI output, "where did this number come from?"
The CSC Securities case: two months to three working days
CSC co-built a "risk assessment gateway" with Kimi, first validated on trustee report generation covering 50+ issuers, 100+ outstanding corporate bonds, and 60+ reports. Integrating existing business systems took 3 working days (planned: 2 months); per-report production fell from ~30 to ~10 minutes; manual effort dropped 67%. Architecture: 1 main agent plus 13 specialized sub-agents. CSC CIO Xiao Gang noted an internal evaluation system where Kimi K3 performed well on long-context, multi-document risk attribution tasks. Human review and decision responsibility is explicitly preserved — drafts do not directly form project conclusions.
Five compliance gates
1. Data classification and tiering (outbound boundaries by sensitivity) 2. Personal-information protection (minimum necessity, filtering/masking/interception) 3. Data-source and tool authorization down to each call 4. Generated-content verification (source annotation, flagged unverified items) 5. Audit and accountability traceability (task–tool-call linkage retention)
The real value of gate 5 combined with the gateway: new scenarios can reuse the same compliance checks, making the second deployment much cheaper than the first.
Quant funds moved first
DeepSeek, Weidu, Niankong, Mingshi, Ubiquant and other top quant funds set up AI Labs around 2025. Traditional manual factor pipelines take 90–180 days; agent-based closed loops at top firms take 7. A strong researcher's weekly high-quality factor output rose from 1–2 to 4 with AI; 90%+ of researchers hired in the past 5 years have AI backgrounds.
Factors are no longer scarce — the bottleneck moved
Hu Chonghai (Guotai Haitong Asset Management) describes the bottleneck shift: LLMs can generate thousands of candidate factors from one hypothesis, so the industry's main contradiction has moved from "lacking factors" to "lacking factor vetting." Their answer is four acceptance gates — economic logic, new/extreme samples, turnover and impact cost, single-style exposure — followed by routine inspection and elimination. The pipeline itself is conventional; what changed is the denominator: acceptance-gate throughput is now the constraint.
25% of Anthropic's ARR
Anthropic released 10 financial agents (pitchbooks, financial modeling, earnings review, valuation, audit, ledger reconciliation, month-end close, KYC screening, credit memos, market research) and reported finance as its second-largest enterprise revenue vertical after tech; ~40% of its top 50 customers are financial institutions. At a ~$65B ARR (late July), finance contributes roughly $16B; Anthropic's internal view is finance may overtake coding as its largest vertical by 2027. An industry survey puts AI usage mix at: research/fundamental analysis 35%, real-time macro/sentiment 15–20%, quant model building 10–15%, compliance document review 10%, credit risk 8–10% — closely matching Kimi's nine skills.
Job requirements changed first
Bank postings already demand agent-era skills: Ningbo Bank requires LLM research experience and AI content review; China Merchants Bank's Shanghai branch hires "digital finance (AI application)" roles covering post-training, prompt engineering, RAG, MCP tool calling, and multi-agent orchestration; ICBC's 2027 campus recruiting adds a tech-elite track.
What to watch
- Whether the efficiency numbers can be independently reproduced by clients
- Whether the gateway is equally cheap for a second scenario
- Whether human-review headcount scales with upstream output (the only human-in-the-loop step; no one has published this number)
- Data-source renewal costs and stability
- Whether Anthropic's "finance overtakes coding" call lands on schedule
Who pockets the efficiency gains?
Tools compress hours, institutions compress cycle times, job descriptions rewrite skill requirements — but beneficiaries differ. If an analyst cuts model building from 5 person-days to 1, does the freed time go to deeper judgment, or to producing triple the report volume? The answer isn't in the tool; it's in the incentive system.
References
1. Moonshot AI Kimi financial solution official release, 2026-09-17 — https://www.kimi.com/ 2. Xinhua Finance via PANews, "Kimi releases financial AI solution, already connected to major institutions", 2026-09-17 — https://news.qq.com/rain/a/20260917A0CZL200 3. IT Home, "Moonshot releases Kimi financial AI solution, first adopters include ICBC and CSC", 2026-09-17 — https://field.10jqka.com.cn/20260917/c680021150.shtml 4..cls, "When factors are no longer scarce, where does excess return come from?", 2026-09-17 — https://fund.eastmoney.com/a/202609173877273783.html 5. Guandian, "Moonshot Kimi releases financial solution with CICC, Ant Group as clients", 2026-09-17 — https://news.qq.com/rain/a/20260917A0D7HO00