From 10 Days to 2: When Financial Workflows Are Broken into Nine Moves
> A financial analyst's day usually goes like this: scan overnight announcements on Bloomberg or Wind, switch to exchange websites to download PDF financial reports, type revenue, gross margin, and operating cash flow into Excel templates one by one, then flip through the latest research reports from three or four comparable companies to verify methodologies. By the time this is done, two or three hours are gone — and the real analysis hasn't started. > > On September 17, Moonshot AI released an AI solution for the financial industry: 10+ authoritative data sources, 9 financial skills, 5 compliance and security measures, plus institutional-grade data modeling and report delivery capabilities. The first batch of adopting institutions includes ICBC, CITIC Construction Investment Securities, CICC, E Fund, Ant Group, and Meituan; on the primary-market side, Sequoia China, ZhenFund, and IDG Capital.
The industry's productivity has long been stuck on two things: the speed of data processing and the breadth of information gathering. The former is limited by how fast a person can type; the latter by how many documents one person can read. Meanwhile, demands from the market keep rising — research conclusions that used to have a day of lead time now require minute-level responses, and shifts in global market structure need to be absorbed instantly.
🧩 Nine Moves for Nine Types of Work
The core design of this solution packages the key stages of financial work into 9 skills, each corresponding to a category of work with stable analytical steps and clear deliverable standards.
| Skill | Job type | Published efficiency data | |---|---|---| | Institutional financial modeling | Investment banking, PE | 5–7 person-days → 0.5–1 person-day | | Institutional research reports | Sell-side and secondary research | 10–20 days → ~2 days | | Institutional PPT | Insurance, investment committee reporting | 7–10 person-days → ~2 person-days | | Financial dynamic charts | Research marketing | Charts can annotate CAGR and deltas | | Earnings reviews | High-frequency output during earnings season | Internal benchmark score up 8–9 points | | Consensus map | Sell-side expectation tracking | — | | Portfolio review | Fund managers' daily work | — | | Holdings morning report | Investment research morning meetings | — | | HK IPO Lens | HKEX listing-application tracking | Scans filing, hearing, and listing status |
The way to read this table is the last column. Three job types have public quantitative figures; the other six list only names. All three with numbers fall in the "cutting manpower by more than half" range; financial modeling shows the largest compression, from 5–7 person-days down to 0.5–1.
A concrete example shows where the boundary lies. Using the "institutional financial modeling" skill to build a financial model for a commercial aerospace company based on management accounts, audit reports, and due diligence materials, the output is an Excel model with adjustable assumptions plus a comparable-company valuation table. The keyword is "adjustable." The deliverable cannot be a sealed conclusion — it must be a working file an analyst can continue to modify. This determines that it replaces the drafting stage, not the judgment stage.
In the solution's built-in internal benchmark tests, earnings-review scores improved by 8–9 points after users applied the financial skills, and financial modeling rose from 69 to 77 points, exceeding Opus 5's 75 points. The specific scoring criteria and sample sizes were not disclosed.
🔗 Data Sources via MCP, Key Figures Traceable
The data-side list deserves a separate look: Wind, Eastmoney, S&P Global, Caixin Data, Cailianshe, Gotosun (Hithink Yujian), Tianyancha, Dun & Bradstreet, Crunchbase, and iFinD; on the macro side, IMF, the World Bank, and the Fed's FRED database.
These data feeds connect directly into research tasks via MCP. Key figures can be checked back to their original sources; provenance and methodology are traceable. It also connects Scholar academic full-text search, Word/Excel/PPT artifact generation tools, plus browser extensions and Computer Use, allowing internal systems without APIs to join the workflow.
"Key figures can be checked back to their original source" matters more than the number of data sources. The first question finance always asks of AI output is "where did this number come from." Traceable methodology means an analytical conclusion can be audited item by item by a reviewer, rather than accepted or rejected only as a whole.
🏦 The CITIC Construction Investment Case: Two Months Down to Three Working Days
The most solid public case in this solution comes from CITIC Construction Investment Securities (CSC).
The two parties co-built a "risk assessment gateway," with "interim trustee report generation" as the first validated scenario. Scope: 50+ issuers, 100+ outstanding corporate bonds, and production of 60+ interim trustee affairs reports.
CSC's architecture is 1 main Agent plus 13 dedicated sub-Agents, handling periodic repetitive work like morning-report compilation and regular reviews, plus complex research covering multiple instruments simultaneously. CSC's Chief Information Officer Xiao Gang noted that the company first built an evaluation system combining objective and subjective measures to assess multiple models' end-to-end delivery on real financial tasks; Kimi K3 performed excellently, with more stable identification and attribution of risk clues in long-horizon analysis tasks involving multiple documents.
One easily overlooked detail: after report first drafts are generated, business staff's review and decision responsibilities are retained — preliminary materials do not directly form approval conclusions. Humans remain in the loop; their position simply shifts from "production" to "review."
🔐 Five Gates That Turn Compliance Requirements into Executable Checks
Finance's peculiarity is that no matter how capable a model is, it can't enter business processes without passing compliance review. The risk assessment gateway CSC and Kimi co-built implements security and compliance requirements as five concrete measures:
1. Data classification and grading — outbound boundaries set by sensitivity level 2. Personal information protection — minimal necessity, filtering/masking/interception 3. Data source and tool authorization — enforced on every call 4. Generated content verification — sources and unverified items annotated 5. Audit and accountability tracing — tasks and tool calls logged with linkage
The fifth item, read together with the last step, reveals the gateway's real value. It saves not just first-time integration time but the review cost of every new scenario afterward: new scenarios can reuse the same compliance check mechanism for integration and go-live review. For a brokerage with many business lines, this determines how much cheaper the second scenario is than the first.
📈 Quant Firms Moved Even Earlier Than Brokerages
The earliest to actually change their working modes were quantitative funds. Billion-yuan quant funds including DeepSeek, Widebread, Niankong, Mingshi, and Jiukun successively established AI Labs around 2025, using AI to rebuild their quant research platforms — extending LLM capabilities from structured price-volume and financial data to unstructured text like research reports, announcements, and news, on top of factor mining, backtesting, and execution optimization that already ran on machines.
- Traditional manual research pipeline: 90–180 days; top institutions with agent closed loops: 7 days
- A high-level factor researcher: from 1–2 high-quality factors per week to 4 with AI
- Of researchers hired by such funds in the past 5 years, 90%+ have AI research backgrounds
- Whether the efficiency figures can be independently reproduced by customers
- Whether the gateway is equally cheap for the second scenario
- Whether human review steps get squeezed
- Renewal costs and stability of data source licensing
- When the "finance overtakes coding" prediction materializes
⚖️ Factors Are No Longer Scarce; the Bottleneck Moved
Hu Chonghai, head of quant investment at Guotai Haitong Asset Management, describes it as fitting a saddle on AI: AI is a powerful wild horse; the saddle is the capability to harness it. His team put the saddle on two stages: factor governance and portfolio management.
Where a researcher once took weeks to go from logic to a valid factor, an LLM now generates thousands of candidate factors from a single investment hypothesis. After the quantity explosion, most are statistical coincidences. The quant industry's main contradiction has shifted from "lacking factors" to "lacking factor discrimination." In this year's extreme markets, many factors with low surface correlation were highly homogeneous in economic logic, and resonated into drawdowns together when the market turned.
Their solution is four acceptance gates — logic (economic meaning), sample (new data and extreme states), trading (turnover and impact cost), and risk (single-style overexposure). Only factors passing all four enter the strategy library, which currently holds 2,000+ valid factors, with ongoing inspection and elimination mechanisms. Nothing in this process is AI-specific — it's all old methods. What changed is the denominator: candidates went from dozens to thousands, and the acceptance gates' throughput became the bottleneck.
💵 25% of ARR on the Table
Zooming out to model vendors reveals harder economics. Anthropic released 10 financial agents at a New York finance event, covering pitchbook building, financial modeling, earnings review, valuation analysis, statement audit, general ledger reconciliation, month-end close, KYC screening, credit memo writing, and comprehensive market research. At the same event, Anthropic said finance is already its second-largest enterprise revenue industry after tech; roughly 40% of its top 50 customers are financial institutions.
Based on the $6.5 billion ARR disclosed at end of July, finance contributes roughly $1.6 billion. The same material notes Anthropic's internal judgment: finance could overtake coding by 2027 as its largest vertical. If that holds, model vendors' product roadmaps will shift, and the "unsexy" parts of finance AI — compliance, audit, human review — will get more investment, because they are the ticket into this market, not add-ons.
Another industry survey gives the structure of financial AI usage: investment research and fundamental analysis 35%, real-time macro and sentiment monitoring 15–20%, quant model building 10–15%, compliance document review 10%, credit risk control 8–10%. Kimi's nine skills align closely with this structure.
🎓 Job Requirements Changed First
The change is transmitting faster than expected. In August, Bank of Ningbo's wealth management department posted a job listing for a strategy analyst role requiring LLM research experience, agent mechanisms, and AI content review. China Merchants Bank's Shanghai branch 2027 campus recruitment includes a "Digital Finance Role (AI Application track)" covering LLM post-training, prompt engineering, RAG, AI Agent development, MCP-based tool calling, Skill development, and multi-agent orchestration. ICBC's 2027 campus recruitment directly set up Tech Elite positions training along R&D manager and product manager tracks.
Once 80% of a sell-side analyst's daily work (reading announcements, organizing data, writing reviews, building models) is handed off, what remains for humans is defining problems, making judgments, and offering differentiated views. All three cannot be standardized into skill packages — and precisely because they can't be standardized, they become more valuable in the short term.
🔭 What to Watch Next
❓ Who Pocketed the Efficiency Gains?
Put the three threads side by side: the tool side compresses man-hours, institutions compress cycle times, job descriptions rewrite capability requirements. All point the same direction — but their beneficiaries differ.
If a sell-side analyst uses AI to cut model building from 5 person-days to 1, will the saved time go to more coverage and deeper judgment, or to producing three times as many reports? The answer isn't in the tools; it's in the incentive system.
📚 References
1. Moonshot AI Kimi Financial Industry Solution official release, 2026-09-17 — https://www.kimi.com/ 2. Xinhua Finance via PANews, "Kimi Releases Financial AI Solution, Connected to Multiple Leading Financial Institutions," 2026-09-17 — https://news.qq.com/rain/a/20260917A0CZL200 3. IT Home, "Moonshot AI Releases Kimi Financial AI Solution, First Adopters Include ICBC and CSC," 2026-09-17 — https://field.10jqka.com.cn/20260917/c680021150.shtml 4. Cailianshe, "When Factors Are No Longer Scarce, Where Does Alpha Come From?" 2026-09-17 — https://fund.eastmoney.com/a/202609173877273783.html 5. Guandian, "Moonshot AI Kimi Releases Financial Industry Solution, CICC and Ant Group Among Clients," 2026-09-17 — https://news.qq.com/rain/a/20260917A0D7HO00