Unique Research · Industry research · July 9, 2024
In financial digitization, large models and generative AI are becoming a new engine of change — remaking how institutions operate and raising the bar on customer experience. Banks, insurers, and brokerages are all exploring these technologies. This piece maps impact, operating models, application scenarios and cases, representative Chinese AI finance vendors, seven capabilities needed to land GenAI, and the hard challenges ahead.
How generative AI hits finance
1. A large efficiency prize
McKinsey Global Institute estimates generative AI could create $2.6–4.4 trillion of annual value across industries. Banking alone may see $200–340 billion a year — 2.8–4.7% of industry revenue, or 9–15% of operating profit — driven mainly by productivity. Leaders have already begun: Citi’s risk and compliance team used GenAI to assess the impact of new federal capital rules; the Wall Street Journal reported Goldman Sachs planned to roll out its first code-generation GenAI tool to thousands of developers by end of June. Firms that succeed are designing operating models that fit GenAI’s risks — not just bolting models onto old processes.
2. Four organizational shapes for landing GenAI
McKinsey surveyed GenAI use at 16 of the largest financial institutions in Europe and the U.S. (combined assets near $26 trillion). More than 50% had adopted more centralized structures to land general-purpose AI — a pattern that may loosen as use matures and business units prioritize their own GenAI work. Four main forms appear:
- Highly centralized: a core team owns design and delivery of GenAI solutions, sits independently, and can move fast with decision autonomy.
- Leadership-central decisions, business execution: smoother leadership–GenAI-team communication and integration, but every project may wait on leadership review.
- Business-led, leadership support: bottom-up buy-in unlocks fast support from lines and functions; maturity can vary sharply across units.
- Highly decentralized: cross-functional teams communicate well and ship insights quickly, but siloed projects can miss shared knowledge and best practice — and under-exploit AI’s potential for breakthrough innovation.
Main scenarios and cases
1. Scenario map
Shenqing Tech (深擎科技) argues AI helps retail finance move toward both intensive and fine-grained service: intensive coverage of long-tail customers at lower cost with “good enough” (~70-point) quality; fine-grained personalization from asset size, trading patterns, needs, and household profiles — including life-cycle and event-triggered content that used to be capacity-constrained.
A South China top brokerage executive’s main LLM / GenAI use cases: centralized Q&A; advisory consultation (expanding stock diagnostics beyond VIP desks via reports and prospectuses); stock diagnosis (faster hotspot capture and plain-language views for RMs); fund diagnosis; and account diagnosis (portfolio risk/return reports). Broader GenAI patterns include:
- Fraud detection & prevention — synthetic “anomaly” patterns to keep detectors ahead of new attacks; less supervision, more automation.
- Personalized products & support — personalization can add up to ~10% annual revenue; GenAI turns transaction history, spend, and savings goals into tailored offers, virtual assistants, and automated form flows.
- Risk & credit scoring — synthetic data enriched with real data for more flexible, less manual scoring.
- Compliance & regulation — high-quality synthetic data and continuous monitoring for consistent, low-error reporting and breach alerts.
- Market & investment analysis — pattern discovery across history; trade parameters (timing, stops, sizing).
- Document processing / report generation — multi-source heterogeneous reports (legal disputes, financials, ownership, charters) for credit and risk; poor report quality raises NPL risk.
2. Representative cases
AlphaSense Assistant — GenAI chat over millions of business and financial documents, powered by AlphaSense’s market-intelligence LLM (ASLLM), with auditable answers back to sources; plus enterprise intelligence over proprietary knowledge.
FeatureSpace TallierLTM™ — claimed first large transaction model (LTM); self-supervised pretraining on cross-jurisdiction behavior; up to 71% better fraud-detection accuracy vs industry-standard models at a typical 5:1 false-positive ratio; APIs turn histories into privacy-preserving “behavior barcodes.”
Visa Account Attack Intelligence (VAAI) score (May rollout to U.S. FIs) — GenAI tool against enumeration attacks; risk score in ~4 ms; trained on 15+ billion transactions; 85% fewer false positives vs prior risk models.
Lanma Tech due-diligence report Agent — for a bank’s inclusive SME lending where RMs spend heavy time gathering materials and writing uneven due-diligence reports. Lanma’s LLM agent auto-drafts analysis and reports, cutting writing time by about 80% and helping catch omissions. (Lanma Tech is a Unique Partner.)
3. Representative Chinese AI finance vendors
Shenqing Tech (深擎) — founded 2018; NLP, personalization, LLMs; core team from IBM China AI Lab (Watson). Content tech and intelligent marketing for brokers and banks; AI Agent–based products; PMF claimed with coverage of ~80% of large/mid brokers and ~50% of large banks. Advantages cited: financial pretrained L1 model near GPT-4 on accuracy benchmarks; function-call accuracy 93%+; multi-turn semantic accuracy/recall 95% / 90%+; AIGC content pass rate 95%+; content products claimed 100× production efficiency in core marketing flows; deep co-builds with top-10 brokers.
Tianxin Tech (甜新) — founded 2016, Shanghai; AIGC + 5G video; investors include Sequoia China, GSR, Cowin. VCRM video-marketing stack for retail, banks, insurance, consumer finance, gaming. Multi-channel reach (AI interactive video, video notify, AI voice outbound, video/text SMS); digital twins from ~3-minute source video; personalized video at scale.
Lanma Tech (澜码) — enterprise LLM Agent platform (AskXBOT); team from Google, IBM, Tencent, ByteDance, Alibaba, Yitu, etc. Tens of millions RMB Series A from IDG, Lianxin, Atom Capital. Advantages: model-neutral Agent atomic-capability data across scenes; expert knowledge as the ceiling for enterprise Agents (vs model hallucination); packaged skills for finance/office workflows; integration with existing org, permissions, and infra.
Seven capabilities to land GenAI in finance
Tianxin Tech frames seven dimensions (privacy, tech/resource barriers, explainability, and compliance remain hard — but paths can be more direct than classic AI programs):
- Clear transformation goals & strategy — leadership vision, consistency, commitment; business-unit accountability; clear scenes and an operating plan.
- Pick the right scenes — pain points / value pockets: risk, credit, robo-advisory, service, etc.
- Data & tech accumulation — warehouses, mining, ML/DL; integrate models with systems, workflows, apps, and sources. McKinsey’s architecture pieces: context/cache, policy, model hub, prompt library, MLOps, risk engine, LLM ops. Data quality and security under shifting regulation.
- Efficient org design — dedicated AI teams, clear ownership, tight business partnership; equip existing teams with GenAI skills rather than only launching side programs.
- Talent & partnerships — finance+AI hybrids; refresh hiring; career paths; partner with leading AI firms/labs.
- Compliance & security risk — redesign model governance before scaling GenAI; explainability and fairness first.
- Continuous optimization & innovation — feedback, better algorithms, new scenes.
Three hard challenges
- Data quality & security — decades of fragmented mainframe estates; anonymization and protection of sensitive customer data.
- Financial regulation — credit, surveillance, transparency, ongoing performance monitoring; cloud flexibility vs data residency; IT integration.
- Ethics & bias — models can amplify historical lending bias; regulators and norms keep moving.
Outlook
How to put GenAI into production is still unsettled. Lanma Tech CEO Zhou Jian notes accuracy is not yet good enough in many finance scenes, and matching scene to stack remains the core problem — selecting or building models that beat humans on the specific task. Two cognitive traps: assuming one LLM can do every task (when staff cannot even specify expert knowledge precisely — better human–machine interaction is needed); and forcing users to adapt to tech instead of the reverse. Landing GenAI also needs cultural change, not only tech upgrades.
References (CN): McKinsey scaling GenAI in banking; DataCamp AI in finance; Visa VAAI press release. Author: qiuping. 11 CN images; attach on ingest.
— END —