Chapter 6 — Finance and banking

Finance remains AI’s largest commercial value pool, with deployments spanning algorithmic trading, credit scoring, fraud detection, robo-advisory, and increasingly central-bank supervisory technology. The sector is also where the “classical” statistical-ML wave still drives most of the value, even as generative AI is layered on top. This chapter develops the financial-AI literature with the depth a graduate course in econometrics, finance, or AI strategy requires, and connects each deployment pattern to its underlying methodological literature.

6.1 Chapter overview

This chapter is structured as follows. Section 6.2 develops algorithmic and quantitative trading from its 1980s origins through to LLM signal extraction, with attention to the econometric methodology that distinguishes successful quantitative funds. Section 6.3 covers generative AI deployment across the major global banks, with case detail on JPMorgan COiN and Morgan Stanley’s advisor assistant. Section 6.4 develops credit scoring as the field’s largest historical value pool, including the FICO methodology, alternative data, the fairness debate, and the April 2026 FICO 10T regulatory inflection. Section 6.5 covers Southeast Asian fintech credit, where alternative-data underwriting is a commercial necessity rather than an optimisation. Section 6.6 develops fraud detection from FICO Falcon’s 1992 neural network to modern graph machine learning at the Bank for International Settlements. Section 6.7 covers robo-advisory and the major-bank exit wave. Section 6.8 develops central-bank supervisory AI with the BIS Project Aurora as the canonical case. Section 6.9 covers Islamic finance and Shariah-compliant AI as a particularly instructive special case for emerging-market readers. Sections 6.10–6.11 develop the DBS and JPMorgan COiN anchor cases at chapter-length depth. Section 6.12 develops the trader-quant culture-clash failure mode. The chapter concludes with a graduate exercise set and further reading.

Reading note. This is the most quantitatively dense chapter in Part II. Where mathematical detail is given, it is given at the level needed to understand what was new about the deployment, not at the level needed to reproduce it. Students with strong econometric backgrounds should read the cited primary literature alongside the chapter; students from a strategy or general-management background should take the chapter’s empirical claims at face value and focus on the operational and strategic implications.

6.2 Quantitative and algorithmic trading

6.2.1 The historical lineage

The intellectual lineage of modern quantitative trading runs through three foundational research programmes:

Modern portfolio theory (Markowitz, 1952) established the mean-variance optimisation framework. The investor’s problem is to maximise expected return $E[R_p]$ subject to a variance constraint $\text{Var}(R_p) \leq \sigma^2_{\text{max}}$, where the portfolio return $R_p = \sum_i w_i R_i$ and $w_i$ are weights that sum to one. The closed-form solution under standard assumptions identifies the efficient frontier of optimal portfolios.
The capital asset pricing model (Sharpe, 1964; Lintner, 1965) introduced the single-factor model $E[R_i] - R_f = \beta_i (E[R_m] - R_f)$, where $\beta_i$ is the asset’s covariance with the market. CAPM gave traders a tractable benchmark for “alpha” — return above what the market beta predicts.
Multifactor models (Fama and French, 1993; Carhart, 1997) extended CAPM to capture additional systematic risk premia: size (small-minus-big), value (high-minus-low book-to-market), momentum (Carhart 1997), and later profitability and investment (Fama-French 5-factor, 2015). The factor-investing literature has grown to encompass dozens of “factor zoo” candidates; a famous critique (Harvey, Liu, and Zhu, 2016) argues that most are spurious, surviving only because of multiple-testing problems.

Quantitative funds that operationalised these frameworks emerged in the 1970s–1980s. James Simons founded Renaissance Technologies in 1982; the firm’s flagship Medallion Fund — closed to outside investors since 1993 — has reportedly produced annualised returns of approximately 39% net of fees over its history, the most successful sustained track record in any asset class. D. E. Shaw (founded 1988), Two Sigma (founded 2001), and DRW (founded 1992) operate at similar quantitative scale.

6.2.2 The architectural pattern as an AI factory

The architectural pattern at every successful quantitative fund is a textbook AI factory in the sense of Chapter 3:

Component	At Renaissance / Two Sigma / D. E. Shaw
Data pipeline	Tick data, level-2 market data, satellite imagery, credit-card panel data, weather, news sentiment, Twitter/X firehose, alternative data from over 1,000 third-party providers
Algorithm development	Classical ML (random forest, gradient boosting), deep learning (LSTM, transformer), and increasingly LLM-based signal extraction
Experimentation platform	The trading book itself — every trade is a real-money experiment, with rigorous attribution of P&L to signals
Software infrastructure	Co-located, sub-millisecond, on-premise (cloud is too slow for high-frequency execution); proprietary backtesting and order-management systems

The factory’s defining feature is the experimentation platform. A quantitative fund that cannot rigorously attribute P&L to signals — separating skill from luck, separating signal from execution — cannot improve. Funds that fail tend to fail at this attribution problem first; their researchers cannot tell which signals are working and which are not, and their portfolio drifts toward correlated factor exposure.

6.2.3 The LLM signal-extraction wave (2024–2026)

The 2024–2026 evolution adds LLM-driven signal extraction from earnings calls, regulatory filings, analyst reports, and management presentations. The pattern:

Open-source or fine-tuned LLMs read a corpus of textual sources at scale.
The LLM extracts structured signals: sentiment, topic distribution, deception probability (using stylistic cues), guidance change relative to prior calls, novel topics relative to prior calls.
The signals feed into a higher-level model that predicts forward returns conditional on the signal.
Position sizing and risk management apply standard portfolio construction.

Several quant firms have publicly disclosed material P&L attribution to LLM-derived signals in 2025. Two Sigma’s published research and Citadel’s hiring patterns (heavy emphasis on ML PhDs with NLP specialisation since 2023) are public-record indicators.

The methodological innovation that matters most is the move from sparse hand-engineered text features (a sentiment dictionary count, a topic-model loading) to dense LLM embeddings that capture semantic and pragmatic content the dictionary approach missed. The trade-off is interpretability: the LLM-derived signals are harder to explain to a risk committee than the dictionary-based signals they replaced.

6.2.4 High-frequency trading and the regulatory regime

High-frequency trading (HFT) emerged in the late 1990s and dominated by the late 2000s. The 2010 Flash Crash (6 May 2010, Dow Jones falling roughly 9% within minutes before recovering) was the moment regulators recognised that algorithmic execution had outpaced market microstructure. The post-2010 regulatory response:

US: SEC Rule 15c3-5 (Market Access Rule) requiring pre-trade risk controls; Regulation SCI (Systems Compliance and Integrity) since 2014.
EU: MiFID II and MiFIR since 2018, with explicit algorithmic-trading testing and reporting requirements.
Singapore: MAS guidelines for algorithmic trading (2022 update) requiring senior-management accountability for algorithm risk.
Bank Negara Malaysia: equivalent guidelines for participants in the regulated FX markets.

The graduate-level reading of the HFT regulatory environment is that the regulatory infrastructure has caught up with HFT but lags AI-driven trading. Algorithmic-execution rules cover speed and pre-trade risk; they do not cover the integrity of the underlying signal or the systemic-risk implications of correlated-AI strategies. This gap is one of the most-discussed open problems at IOSCO and the BIS Innovation Hub.

6.3 Generative AI in major banks

The 2023–2025 wave brought generative AI into the operating cores of every major global bank. The deployment pattern is consistent: a permissioned in-house LLM platform (often built on Azure OpenAI or AWS Bedrock with private inference), fronted by role-specific applications (research drafting, code generation, customer-support assist, compliance review), with rigorous data-leakage controls and content moderation.

6.3.1 The deployment table

Bank	System	Reach	Reported impact
JPMorgan Chase	LLM Suite	~230,000 employees by 2025	3–6 hours saved per employee per week; estimated $1–1.5B annual value
Morgan Stanley	AI @ Morgan Stanley Assistant	98% of advisor teams	Across $5.5T in client assets
Bank of America	Erica	Retail customers	3 billion lifetime client interactions; deterministic NLP, not generative
Goldman Sachs	GS AI Platform	11,000+ developers and bankers	+20% engineering productivity
Citi	Citi Stylus, Citi Assist	140,000 employees by 2025	Document drafting, research, trading-floor productivity
Deutsche Bank	Beacon (NLU + RAG)	Investment-banking research desks	Time-to-research drafts halved
BBVA	BBVA AI Factory	~120,000 employees	Cited as one of the most-mature European deployments
HSBC	AI at HSBC programme	~210,000 employees	Productivity gains reported but not quantified publicly

6.3.2 The JPMorgan COiN baseline

JPMorgan’s COiN system, deployed in 2017 and now mature, eliminates 360,000 lawyer-hours per year by reading 12,000 commercial credit agreements in seconds. The original architecture was an ML pipeline (rule-based extraction with ML clause classification); the 2023+ evolution added LLM-based natural-language understanding on top. President Daniel Pinto estimated AI’s annual value at the bank at $1–1.5 billion against an $18 billion 2025 tech budget — roughly an 8% productivity dividend on the technology budget alone, ignoring revenue uplift.

6.3.3 The Morgan Stanley advisor pattern

Morgan Stanley’s AI @ Morgan Stanley Assistant is operationally distinct from the COiN pattern. It is a knowledge surface for human advisors rather than an autonomous decision system. The economics:

The marginal cost of advisor time has fallen by approximately 40% (each advisor can now serve ~1.5–2× the prior client load).
The marginal cost of compliance review has fallen because the assistant proactively flags potentially problematic communications.
The marginal revenue per advisor has risen because advisors spend more time on client relationships and less time on document retrieval.

The pattern is distinctively the augmentation case in the Acemoglu-Restrepo automation/augmentation taxonomy (Chapter 15): the assistant complements human advisors rather than replacing them. The deployment economics work because the marginal cost of advisor time has fallen faster than the marginal revenue per client, allowing the bank to grow its advisor productivity without growing its advisor headcount proportionally.

6.3.4 The deployment template

The emerging deployment template across major banks has six layers:

Inference infrastructure — typically Azure OpenAI for closed-frontier models with private inference, plus AWS Bedrock or self-hosted Llama / DeepSeek for open-weight workloads where data sensitivity matters.
Permission and data-leakage controls — every model query is logged; the model’s accessible context is restricted by user identity; output is scanned for PII and confidential-data leakage before display.
RAG layer — the bank’s policies, product documentation, regulatory references, and (where appropriate) anonymised case histories are indexed for retrieval.
Role-specific applications — research-drafting tools for analysts, code-generation tools for engineers, customer-support assist for relationship managers, compliance-review tools for the second line of defence.
Adoption and training — every role with an application has an associated training programme, KPI alignment, and ongoing measurement.
Governance — model risk management committee, bias and fairness audits, regulatory reporting, and incident response.

A graduate student asked to specify a generative-AI deployment for a regional bank should specify all six layers, not just the first.

6.4 Credit scoring

Credit scoring is the financial-AI sector’s largest historical value pool. The 1989 introduction of the FICO score is plausibly the single most successful pre-ML statistical risk model ever deployed.

6.4.1 The FICO methodology

The classical FICO score (FICO 8 and earlier) is a logistic regression over five canonical factor groups, with weights established by Fair, Isaac & Co. and largely held stable across decades:

Factor	Weight	Description
Payment history	35%	Whether the consumer has paid past credit obligations on time
Amounts owed	30%	Credit utilisation (current balances relative to credit limits); especially the revolving utilisation ratio
Length of credit history	15%	Average age of accounts, age of oldest account, age of most recent account
Credit mix	10%	Diversity of credit types (revolving, instalment, mortgage)
New credit	10%	Recent inquiries and recently opened accounts

The model’s stability across decades is a feature, not a bug — lenders, regulators, and securitisation markets all rely on the score’s predictive consistency. Changes are made through major version updates (FICO 8 → FICO 9 → FICO 10 → FICO 10T) with multi-year transition windows.

6.4.2 The FICO 10T regulatory inflection

The watershed event in mainstream credit scoring came on 22 April 2026, when FICO Score 10T received GSE approval for use in conforming mortgage underwriting. The 10T variant — the “T” stands for “trended” — uses 24 months of trended balance and payment data rather than point-in-time snapshots. The trend signal materially outperforms the snapshot signal in predicting default, particularly for borrowers near regulatory thresholds.

The economic significance of the GSE approval: the conforming mortgage market — Fannie Mae and Freddie Mac purchases — comprises the bulk of US residential mortgage origination. Approval of FICO 10T for this market means that trended-data and machine-learning-derived scores now flow through the largest US mortgage market, with downstream effects on mortgage pricing, securitisation, and consumer access to credit.

6.4.3 Alternative-data underwriting

The 2010s+ wave of fintech credit firms (Upstart, Affirm, Zest AI, LendingClub) deployed ML models trained on alternative data — bank-account transaction history, employment patterns, education, mobile-phone metadata, e-commerce behaviour. The headline performance claims:

Upstart approves 101% more applicants than traditional FICO models, while achieving comparable default rates (Upstart 2024 reported figures).
Zest AI customers report 25% higher approvals at 15% lower default rates in published case studies.

These figures should be read with a methodological caveat: the comparison set (traditional FICO underwriting) and the alternative-data set (Upstart’s ML underwriting) are not perfect controls. Upstart’s customer base is selected by acceptance into Upstart’s funnel, which is itself a function of marketing reach and partnership patterns. Properly identified randomised comparisons are rare in the alternative-data literature; the percentage-point improvements should be treated as upper bounds on the true treatment effect.

6.4.4 The fairness debate

Alternative-data underwriting has been the subject of sustained fairness debate since the mid-2010s. Three concerns recur:

Disparate impact. ML models trained on historical lending decisions inherit historical bias. If lenders historically approved white applicants at higher rates than equivalent Black applicants, the model trained on those decisions will continue the pattern, even if race is not a model input. The Buolamwini-Gebru pattern (Chapter 14) generalises beyond face recognition.
Proxies for protected characteristics. Zip code, school attended, employer name, and similar features can serve as proxies for race, religion, or national origin. Removing race as a model input does not address the proxy problem.
Explainability and recourse. Under US fair-lending regulation (Equal Credit Opportunity Act, Regulation B), denied applicants are entitled to specific reasons for the denial. ML models with hundreds of features make this difficult; SHAP and LIME-based explanation methods are standard practice but their fidelity to the underlying model is incomplete.

The 2024 CFPB guidance, the 2025 Fed SR 11-7 update, and the EU AI Act high-risk category for credit scoring (Chapter 14) all formalise the regulatory expectation that fairness, explainability, and recourse are not optional features in production credit-scoring systems.

6.4.5 The Hemachandran chapter on emerging-market banking

Hemachandran and Rodriguez (eds., 2024) — chapter authors Jafar, Alam, and El-Chaarani — emphasise three ML applications that scale particularly well in emerging-market banking, where bureau data is thin or non-existent:

SME credit scoring using alternative data — mobile-phone usage, geolocation, utility-payment history, e-commerce behaviour. The deployments work because emerging-market borrowers are typically thin-file in the conventional sense but have rich digital footprints.
Cash-flow forecasting for working-capital lending, using bank-feed data and integrated accounting platforms.
Portfolio-construction tools for retail wealth management at HNW thresholds previously uneconomic for conventional advisor-led service.

The chapter’s central observation is that emerging-market banking is leapfrogging the developed-world architectural pattern: instead of building bureau-data-centric underwriting and then extending to alternative data, emerging-market lenders are building alternative-data-centric underwriting from the start.

6.5 The fintech-credit landscape in Southeast Asia

The Southeast Asian fintech-credit landscape is particularly illustrative because alternative-data underwriting is a commercial necessity rather than an optimisation. The bureau-data infrastructure that supports developed-market lending is patchy: Indonesia’s Sistem Layanan Informasi Keuangan (SLIK) is comprehensive for bank borrowers but does not cover unbanked populations; the Philippines and Vietnam have similar coverage gaps; Malaysia’s Central Credit Reference Information System (CCRIS) is mature but focused on bank credit.

6.5.1 The major firms

Kredivo (Indonesia). Founded in 2016 by Akshay Garg, FinAccel’s Kredivo product offers buy-now-pay-later and personal lending. Underwrites on mobile-phone metadata, e-commerce behaviour, and bill-payment history. As of 2025, Kredivo serves over 8 million users with originations exceeding USD 5 billion cumulative. The firm acquired a regional bank licence (Bank Bisnis Internasional) in 2022, becoming the first Indonesian fintech to operate as a fully licensed bank.

Akulaku (Indonesia, Philippines, Malaysia, Vietnam). Founded by Li Wenbo in 2016 (based in Singapore). Consumer credit at scale across four ASEAN markets. Uses smartphone metadata, utility-payment data, and social-graph signals. Acquired a banking licence in Indonesia (Bank Neo Commerce) and operates as a regional super-app.

Atome (Singapore, regional). Advance Intelligence Group’s BNPL product, founded 2019. Regional spread across nine markets including Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam, Hong Kong, Taiwan, and Mainland China. Underwrites on transaction-frequency data and partnerships with major retailers.

Aspire (Singapore). Founded 2018, focused on SME banking. Cash-flow-based underwriting on integrated bank-feed and accounting data. Distinctive for its B2B focus in a market dominated by consumer fintechs. Series C raised USD 100M in 2023 at a USD 600M+ valuation.

Funding Societies (Singapore, regional). Founded 2015 by Reynold Wijaya and Kelvin Teo, focused on SME peer-to-peer lending. ML-mediated risk segmentation across thin-file SMEs. The largest digital SME-financing platform in Southeast Asia by origination volume.

Validus (Singapore, regional). SME lending across Singapore, Indonesia, Vietnam, and Thailand. The most explicit example of the “alternative data plus supply-chain financing” pattern: many of Validus’s loans are anchor-buyer-guaranteed receivables financing, where the alternative data is the borrower’s transaction history with a known anchor buyer.

6.5.2 The regulatory landscape

The regional regulatory landscape varies substantially:

Singapore: MAS has the most-developed digital-banking framework in the region. The 2020 Digital Bank licences (issued to GXS, Trust Bank, ANEXT, and others) explicitly accommodate alternative-data underwriting under the same prudential framework as traditional banks.
Indonesia: Otoritas Jasa Keuangan (OJK) has progressively tightened P2P-lending regulation since 2018, with major reforms in 2022 and 2024 raising minimum-capital and operational requirements.
Philippines: Bangko Sentral ng Pilipinas (BSP) Digital Bank framework (2020) and the 2025 amendments to the Credit Information Corporation Act expand the scope of credit information available to lenders.
Malaysia: Bank Negara Malaysia issued the digital bank framework in 2020 and licensed five digital banks in 2022 (Boost-RHB, GXBank, KAF Investment Bank, AEON-MoneyLion, YTL-Sea). The framework explicitly accommodates alternative-data underwriting.
Vietnam: State Bank of Vietnam’s regulatory framework has been slower to formalise, with Decree 13/2023 establishing a sandbox framework for fintech but without comprehensive primary legislation.

6.5.3 The structural pattern

The structural pattern across Southeast Asian fintech credit is that the alternative-data approach generalises better in emerging markets than in developed markets — partly because the bureau data alternative is weaker, partly because the digital footprint per consumer is denser (smartphone penetration is high; cash transactions are giving way rapidly to digital wallets and BNPL), and partly because regulatory frameworks have been built around the new methodology rather than retrofitted.

A graduate student studying emerging-market fintech should read Tan and Khor (2024) for the comprehensive regulatory survey, and the BIS WP 1145 paper (Auer, Frost, and Vidal Pastor, 2024) for the macroeconomic implications.

6.6 Fraud detection

Fraud detection is the financial-AI sector’s most consistently profitable category. The 1992 deployment of FICO Falcon Fraud Manager — a neural network running in production on most of the world’s payment-card transactions — predates almost every other commercial AI deployment of comparable scale.

6.6.1 The classical pattern

Modern card-not-present fraud detection uses a stack of methods that have evolved over four decades:

Rule-based filters for the most obvious patterns (transactions from sanctioned jurisdictions, transactions exceeding velocity thresholds, transactions with mismatched billing/shipping addresses).
Classical ML scoring (logistic regression, gradient boosting) on transaction features (amount, merchant category, geographic distance from prior transaction, time since prior transaction, device fingerprint, IP geolocation).
Deep learning sequence models (LSTM, transformer) on the user’s recent transaction history, capturing patterns that classical features miss.
Graph neural networks on the transaction graph, capturing patterns where the suspicious behaviour is the relationship between transactions rather than any single transaction.

Each layer adds incremental signal. The current architectural pattern at the major payment networks (Visa, Mastercard, American Express) is a layered ensemble where the upstream rule-based filters handle the obvious cases and the downstream graph-ML models handle the subtler ones.

6.6.2 The reported performance figures

Mastercard Decision Intelligence Pro reports up to 300% improvement in detecting at-risk merchants and 200% reduction in false positives versus the prior generation of detection systems. Visa Advanced Authorization is credited with preventing over USD 40 billion in fraud annually. These numbers should be read as Mastercard’s and Visa’s own claims, not as audited third-party figures, but the order of magnitude is consistent across both networks and across independent academic studies of the same time period.

6.6.3 The graph-ML frontier

The 2024–2026 frontier is cross-institution fraud graphs — graph ML applied to transaction networks that span multiple banks and payment processors via privacy-preserving compute (federated learning, homomorphic encryption, secure enclaves).

The privacy-preserving requirement is binding because banks cannot share raw transaction data across institutions under most regulatory regimes. Federated learning lets each bank train its share of a graph-ML model on local data, sharing only the model parameters; homomorphic encryption lets the consortium compute on encrypted data without decrypting it; secure enclaves (Intel SGX, AWS Nitro, Azure Confidential Computing) provide hardware-rooted trust for joint computation.

6.6.4 BIS Project Aurora

The Bank for International Settlements’ Project Aurora (BIS Innovation Hub Nordic Centre, 2024) is the canonical demonstration of cross-institution fraud detection. The project trained a graph neural network on synthetic transaction data simulating multi-bank money-laundering networks, with results published in a 2024 BIS report. The headline finding: graph machine learning can detect up to 3× more money-laundering networks at 80% lower false positives than rules-based approaches.

The project’s findings have been institutionalised in the BIS Innovation Hub’s roadmap and influence supervisory AI design at multiple G20 central banks. The Singapore counterpart (Project Mojito, MAS Innovation Lab) and the EU counterpart (Project Atrium, ECB) develop the same architectural pattern in different regulatory contexts.

A graduate student should read the BIS Project Aurora technical report alongside Ron and Shamir (2013) on the original cryptocurrency-network analysis methodology — the underlying graph-analysis techniques generalise from public-blockchain analysis to private-banking-network analysis once the privacy-preserving compute layer is in place.

6.7 Robo-advisory

Robo-advisory crossed USD 1.2 trillion in AUM by year-end 2024, with the major players’ AUM as follows:

Provider	AUM (USD, 2024 year-end)	Origin
Vanguard Personal Advisor Services	365B	Hybrid (advisor + robo)
Empower (formerly Personal Capital)	200B	Hybrid
Schwab Intelligent Portfolios	89.5B	Pure robo
Betterment	56.4B	Pure robo
Wealthfront	50B	Pure robo
Goldman Sachs Marcus Invest	(closed 2024)	Pure robo
JPMorgan Automated Investing	(closed 2024)	Pure robo
Ellevest	(acquired 2024)	Pure robo

6.7.1 The exit wave

The 2023–2024 exit of Goldman Sachs Marcus Invest, JPMorgan Automated Investing, and Ellevest from the robo-advisory market is one of the most-studied strategic patterns of the period. The proximate causes:

Customer acquisition cost (CAC) was higher than the long-term value (LTV) of the average robo-advisor customer. Vanguard, Schwab, and Betterment had lower CAC because of pre-existing customer relationships and brand; new entrants without these had to acquire through paid marketing at uneconomic rates.
Asset-management margin compression. The fee for robo-advisory has been compressed to 25–35 bps for most providers. Goldman and JPMorgan, accustomed to higher fee margins in their traditional wealth-management businesses, faced cannibalisation pressure that smaller pure-robo firms did not.
Operational cost asymmetry. The marginal cost of running a robo-advisor is close to zero for an existing wealth-management firm with the data and compliance infrastructure; for a pure-robo firm with a small AUM, the fixed-cost overhead per dollar of AUM is much higher.

The teaching point is that scale is not the same as competitive advantage. The robo-advisor as commoditised distribution channel has won; the robo-advisor as proprietary moat has not. Firms that succeeded were those whose robo-advisor was a feature of a broader wealth-management platform; firms that failed were those whose robo-advisor was the platform itself.

6.7.2 The 2024–2026 evolution: LLM-driven advice

The 2024–2026 evolution adds LLM-driven advice surfaces on top of the robo-advisor stack. Morgan Stanley’s AI @ Morgan Stanley Assistant is the most-deployed example: not a robo-advisor but a knowledge surface that lets human advisors handle 2–3× the prior client load. The unit economics work because the marginal cost of advisor time has fallen, not because the distribution has been automated.

The pattern suggests that the long-run wealth-management equilibrium is hybrid: pure-robo for the mass-affluent segment, advisor-plus-AI for the HNW and UHNW segments, with the AI assistant compressing the cost asymmetry between them. The 2026 industry data is consistent with this trajectory.

6.8 Central banking and supervisory AI

The 2024–2026 wave of central-bank AI deployment is one of the period’s quieter but most economically consequential developments.

6.8.1 The deployment landscape

Central bank	System	Application
BIS Innovation Hub	Project Aurora	Cross-institution money-laundering detection
BIS Innovation Hub	Project Atlas	Cross-border payments analysis
Bank of England	Prudential Regulation Authority ML	Stress-testing scenario generation
European Central Bank	NLP analytics	Supervisory dialogue and bank disclosures analysis
Monetary Authority of Singapore	Project Mojito	Suspicious-transaction monitoring
Monetary Authority of Singapore	Veritas	Fairness-and-explainability framework for FS AI
Federal Reserve	SR 11-7 model risk management	Updated guidance for ML model governance
Bank Negara Malaysia	Sandbox-tested supervisory analytics	On the Open API data layer
People’s Bank of China	DC/EP analytics	CBDC transaction monitoring
Reserve Bank of India	RegTech sandbox	Fintech supervisory analytics

6.8.2 The graduate-level reading

Two important threads run through the supervisory-AI literature:

The asymmetry of supervisory information. Central banks supervise large numbers of financial institutions but observe each only periodically. ML on transaction data, regulatory submissions, and market data lets supervisors identify patterns at scale that periodic on-site inspection cannot. This is the central motivation for Project Aurora and similar initiatives.
The limits of supervisory AI. ML systems are themselves prone to bias, drift, and explainability failures. A central bank that uses ML to allocate supervisory attention can systematically under-supervise institutions whose data patterns the model treats as benign — and the institutions can learn to game the model. The MAS Veritas framework is the most-developed regulatory response to this concern: it prescribes how supervisory ML systems should be tested, audited, and governed.

The Singapore counterpart (Project Mojito, MAS Innovation Lab) and the EU counterpart (Project Atrium, ECB) develop the same architectural pattern in different regulatory contexts. The convergence across G20 central banks is striking: by 2026, every major central bank has at least pilot-stage supervisory ML deployments, and the BIS Innovation Hub is the de facto coordinating institution.

6.9 Islamic finance and Shariah-compliant AI

Islamic finance AI is a particularly instructive special case for emerging-market readers, illustrating both the scale of opportunity in alternative regulatory environments and the operational challenges of deploying generative AI in domains where verification matters more than throughput.

6.9.1 The scale and the players

The global Islamic finance industry exceeded USD 4 trillion in assets by 2024, with Malaysia, the GCC, Indonesia, and Pakistan as the largest markets. The major firms operating Shariah-compliance ML systems:

Zoya — US-based, AAOIFI-certified Shariah-compliance screening for individual investors. ~40,000 instruments screened in real time.
Wahed Invest — US/UK-based Islamic robo-advisor with assets under management exceeding USD 1B.
Musaffa — Cross-jurisdiction Shariah-compliance platform with brokerage and screening services.
IFG.vc — UK-based Islamic-finance media and advisory firm operating an institutional screening service.
HalalScreener — Malaysian-origin Shariah-compliance service with regional reach.

The firms automate AAOIFI Standard No. 21 screening across roughly 40,000 instruments, applying Shariah filters on income from impermissible activities (alcohol, tobacco, conventional banking, etc.) and on financial ratios (debt to market capitalisation, interest-bearing securities to total assets, accounts receivable to market capitalisation).

6.9.2 The methodological constraint

The deployments remain primarily rule-based rather than LLM-grounded — the regulatory tolerance for hallucination in fatwa-adjacent domains is essentially zero. Generative AI is increasingly used for customer-facing explanation rather than core screening logic.

The structural challenge is verification. A retrieval-augmented LLM can summarise an AAOIFI ruling, but the user (or the Shariah supervisory board) must be able to verify the citation chain back to authoritative sources. This is a generalisable pattern: in regulated domains, AI’s value comes from compressing the time to a verifiable answer rather than from delivering a fully autonomous one.

6.9.3 The implications for emerging-market AI design

The Islamic-finance case generalises to a broader emerging-market AI design principle: in domains where verification is costly and errors are highly consequential, the right architectural pattern is “AI for retrieval, humans for verification”. The graduate-level reading is that the autonomous-AI paradigm imported from Silicon Valley does not fit emerging-market regulatory realities, and that the right deployments are those that amplify human verification capacity rather than replace it.

6.10 Anchor case — DBS revisited (banking-specific)

The DBS transformation discussed in Chapter 4 produced specifically banking outcomes that deserve treatment in their own right.

6.10.1 The banking metrics

Anchor case — DBS banking outcomes. - Credit-card origination time fell from 21 days to 4 days — a four-fold improvement driven by journey redesign, not just ML. - 50,000 personalised daily nudges delivered to consumer banking customers. - End-to-end AML surveillance combining rules, network link analysis, and ML on internal and external data — the bank’s most operationally valuable AI deployment. - SGD 150M additional revenue + SGD 25M from loss prevention attributed to AI in a single recent year. - Lowest staff turnover in Singapore (10% vs 15–20% industry average) — using ML to predict employee attrition risk and intervene early. - 3× engineering productivity with internal AI assistants by 2025.

6.10.2 The credit-card origination journey

The 21 → 4 day improvement is worth dissecting. The pre-transformation journey had nine sequential steps: application, manual KYC, credit-bureau pull, underwriting review, fraud check, credit-limit assignment, card production, postal delivery, and activation. Each step had its own queue, its own SLA, and its own handoff to the next.

The post-transformation journey collapsed seven of the nine steps into a software-mediated workflow. KYC is now real-time via MyInfo (Singapore’s central identity service); credit-bureau and fraud checks run in parallel rather than sequentially; underwriting is ML-mediated with human review only for edge cases; the card is digitally provisioned to the customer’s wallet immediately, with the physical card following by post.

The point is that the four-day credit-card origination is not an ML achievement; it is a workflow-redesign achievement enabled by ML at three specific steps. A bank that deployed the same ML models without redesigning the workflow would not see the 21 → 4 day improvement.

6.10.3 The personalised-nudges programme

The 50,000 daily nudges programme is operationally distinct from typical banking marketing. The nudges are individualised, action-prompting messages — “you have an unused $X in your account; consider moving it to a fixed deposit”, “your monthly subscription to Y has changed price”, “your balance is approaching your typical low point; here’s how to avoid an overdraft”. The system is built on a behavioural-economics framing (Thaler-Sunstein nudge theory) rather than a marketing framing.

The economic significance: the nudges produce a measurable lift in customer-balance behaviour and a measurable reduction in customer-complaint volume. The bank’s published case studies attribute roughly SGD 30–40M in annual revenue impact to the nudges programme, with comparable cost savings from reduced complaint handling.

6.11 Anchor case — JPMorgan COiN: the $5B-tech-budget anchor

Case study — JPMorgan COiN. Deployed in 2017, COiN (Contract Intelligence) reads commercial credit agreements at scale. The often-cited “360,000 lawyer-hours saved” figure derives from the bank’s own 2017 disclosure. By 2025, the LLM Suite extends the same template across ~230,000 employees, with ChatGPT-class capability layered on top of the bank’s proprietary models.

6.11.1 The technical architecture

COiN’s original 2017 architecture was a hybrid pipeline: rule-based extraction for boilerplate clauses, supervised ML clause classification for non-boilerplate clauses, and a layered review interface that surfaced the model’s confidence and let lawyers adjudicate uncertain cases. The system processed approximately 12,000 commercial credit agreements that previously required 360,000 lawyer-hours per year.

The 2023+ LLM Suite extension layers generative AI on top: contract summarisation, anomaly detection (clauses unusual relative to the bank’s historical agreements), counterparty-risk assessment from filings and news, and drafting assistance for new agreements. The suite is built on Azure OpenAI with private inference, with rigorous data-leakage controls and content moderation.

6.11.2 The value at scale

President Daniel Pinto estimated AI’s annual value at the bank at USD 1–1.5 billion against an USD 18 billion 2025 tech budget — roughly an 8% productivity dividend on the technology budget alone, ignoring revenue uplift. The case is instructive because it shows that AI value at scale comes from integration, not heroic individual deployments.

The hard work was the document-engineering pipeline that makes the model output usable to lawyers and analysts: workflow integration with the bank’s existing case-management systems, audit trails for regulatory examination, role-based access controls, and the long-cycle adoption work to bring the bank’s lawyers and analysts into productive use of the tooling.

6.11.3 The contrast with Watson Health

The COiN case is a useful contrast with the Watson Health case (Chapter 7). Both are IBM-era projects (COiN was originally developed in partnership with IBM consulting services) of comparable initial ambition. COiN’s success came from narrow problem framing (reading commercial credit agreements) and tight workflow integration with the bank’s lawyers. Watson Health’s failure came from broad problem framing (any cancer treatment recommendation) and weak workflow integration with the participating clinical sites.

The lesson generalises: the binding constraint in AI deployment is rarely model capability; it is the narrowness of the problem and the depth of the workflow integration. The Iansiti-Lakhani framework’s emphasis on workflow redesign is precisely this point.

6.12 The trader-quant culture clash failure mode

One under-discussed pattern in financial-services AI is the cultural friction between the front-office trading culture (intuition-driven, individual P&L, short time horizon) and the AI factory’s culture (data-driven, team P&L, longer time horizon).

6.12.1 The pattern

The most consistent failure mode in bank AI deployments is not technical; it is cultural. Specifically:

The front office refuses to adopt model output that contradicts trader intuition. The model’s win rate is judged on its agreement with the trader’s prior judgement rather than on its forward predictive accuracy.
The front office refuses to surface the data needed to train models in the first place. Trader notes, prior decisions, and rationale are considered proprietary to the individual trader, not to the firm.
The compensation system rewards individual P&L attribution rather than systematic-strategy contribution. Traders are incentivised to claim winning trades as their judgement and losing trades as model error.

6.12.2 The successful counter-pattern

The successful AI deployments at major banks (Goldman’s GS AI Platform, JPMorgan’s LLM Suite, DBS’s GANDALF) have invested heavily in cultural alignment alongside technical infrastructure:

Compensation realignment. Trader bonuses are tied partly to systematic-strategy P&L, with the bank capturing the systematic component and the trader capturing the discretionary component.
Data ownership policies. Trader notes and decision rationale are treated as firm assets, with the bank investing in tooling that captures them as a byproduct of the trader’s normal workflow.
Model adoption metrics. Adoption is measured at the trader-cohort level, with explicit interventions for cohorts whose adoption is below threshold.

The graduate-level reading is that the binding constraint in financial-services AI is rarely the model; it is the front-office culture. A bank that solves the cultural problem captures the productivity gains; a bank that solves the technical problem without addressing the culture does not.

6.13 Connection to financial econometrics

Financial econometrics has a substantial intellectual tradition that intersects with the financial-AI literature in important ways.

6.13.1 Volatility modelling

The ARCH model (Engle, 1982) and GARCH (Bollerslev, 1986) family established conditional volatility modelling as the foundation of modern risk management. The current frontier is realised-volatility modelling (Andersen, Bollerslev, Diebold, and Labys, 2003) using high-frequency data, with deep-learning extensions (LSTM-GARCH, transformer-based volatility models) emerging in the 2020s.

A graduate student should read the Engle 1982 Econometrica paper and the Bollerslev 1986 Journal of Econometrics paper as the foundational texts; the Andersen et al. (2003) paper is the realised-volatility canonical reference.

6.13.2 Spillover and connectedness

The Diebold-Yilmaz (2009; 2012) spillover index methodology has become the standard tool for measuring volatility connectedness across assets, sectors, and countries. The current frontier extends the methodology to time-varying parameters (Diebold, Yilmaz, and Demirer, 2018) and to higher-moment connectedness (Bouri et al., 2021).

The financial-AI implication: connectedness measures can be inputs to ML models predicting forward returns or to graph-ML models on cross-asset networks. The spillover framework is the right entry point for graduate students who want to bridge financial econometrics and financial AI.

6.13.3 Term-structure and yield-curve modelling

The Nelson-Siegel-Svensson family of yield-curve models, the affine term-structure literature (Duffie and Kan, 1996; Dai and Singleton, 2000), and the dynamic term-structure literature (Diebold and Li, 2006) provide the methodological substrate for fixed-income AI. Modern deep-learning term-structure models (Bianchi, Büchner, and Tamoni, 2021) extend the framework with greater predictive accuracy at the cost of interpretability.

6.13.4 The methodological synthesis

The graduate-level synthesis is that financial-AI methodology builds on, rather than replaces, financial econometrics. The ML models that perform best in production are those whose design is informed by the econometric literature — feature engineering reflects factor models, evaluation reflects forecasting-error metrics, and risk management reflects volatility modelling. The “AI is replacing econometrics” framing that occasionally appears in the trade press is empirically false in the funds and banks where the work is actually done.

Exercises 6.1

Quantitative-fund architecture. For a quantitative fund of your choice (Renaissance, Two Sigma, D. E. Shaw, AQR, Citadel), construct the four-component AI factory specification. Identify which component is the moat. What evidence supports your identification?
The factor zoo critique. Harvey, Liu, and Zhu (2016) argue that most published return-predicting factors are spurious survivors of multiple-testing. (a) Construct a Bayesian framework for evaluating new factor claims. (b) Apply the framework to one specific factor of your choice. (c) Discuss the implications for ML-derived signals, where the multiple-testing problem is even more severe.
LLM signal extraction methodology. Design an LLM-based signal extraction pipeline for earnings-call sentiment. Specify: the input corpus, the LLM choice, the extracted signals, the evaluation framework, the integration with portfolio construction. Identify three failure modes and the controls that address them.
The COiN economics. JPMorgan COiN saved 360,000 lawyer-hours/year on 12,000 contracts. (a) Estimate the dollar value of those hours. (b) Estimate the build and ongoing cost of COiN. (c) Compute the ROI. (d) What does the high ROI suggest about why other banks have not built equivalent systems?
The robo-advisory exit. Goldman Sachs Marcus Invest, JPMorgan Automated Investing, and Ellevest exited the robo-advisory market in 2023–2024 while Vanguard, Schwab, and Betterment did not. Construct a strategic-management analysis distinguishing the two groups. What does the divergence tell you about competitive advantage in commoditised distribution channels?
The Upstart fairness audit. Upstart approves 101% more applicants than traditional FICO models. (a) Identify three sources of bias that this comparison may obscure. (b) Design a randomised controlled audit that would identify whether the additional approvals reflect genuine model superiority or selection bias. (c) Discuss the regulatory implications under the Equal Credit Opportunity Act.
The Southeast Asian fintech credit landscape. Choose two of the firms in §6.5.1 (Kredivo, Akulaku, Atome, Aspire, Funding Societies, Validus). For each, construct the four-component AI factory specification. Identify the binding constraint. Compare and contrast.
Project Aurora replication. Read the BIS Project Aurora 2024 report. (a) Identify the three most-important methodological choices the project made. (b) For each, construct an alternative choice and discuss how it would change the headline results. (c) Discuss the implications for whether the 3× detection improvement generalises to live deployments.
The trader-quant culture clash. A regional bank is deploying its first systematic-trading ML system. The front-office trading desk is sceptical and uncooperative. (a) Identify the three behavioural patterns most likely to undermine the deployment. (b) For each, design an organisational intervention. (c) What changes to compensation structure would help, and why?
The 22 April 2026 inflection. FICO 10T’s GSE approval changes the architecture of US mortgage underwriting. (a) Estimate the static effect on the conforming mortgage market (approval rates, default rates, mortgage rates). (b) Identify the dynamic effects over a 5-year horizon. (c) Discuss the implications for non-US mortgage markets that may follow the FICO 10T pattern.
The DBS playbook for a regional bank. Apply the DBS playbook from §6.10 to a regional bank in your country. (a) Identify the priority domains. (b) Identify the binding capability constraints. (c) Construct a 5-year transformation plan with annual milestones.
The financial-econometrics integration. Choose a quantitative finance application (return prediction, volatility forecasting, credit risk modelling, portfolio optimisation). Construct an integrated workflow that uses both classical econometric tools and modern ML. Identify the comparative advantages of each.

References for this chapter

Core

Iansiti and Lakhani (2020). Competing in the Age of AI. Harvard Business Review Press. (Chapters 1, 6, 8.)
Lamarre, Smaje, and Zemmel (2023). Rewired. Wiley. (Chapter on banking transformation.)
Hemachandran and Rodriguez, eds. (2024). Artificial Intelligence for Business. Routledge. (Jafar, Alam, El-Chaarani chapter on banking; Choudhury chapter on supply chain finance.)

Quantitative trading and financial econometrics

Markowitz (1952). Portfolio selection. Journal of Finance 7(1): 77–91.
Sharpe (1964). Capital asset prices. Journal of Finance 19(3): 425–442.
Fama and French (1993). Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33: 3–56.
Fama and French (2015). A five-factor asset pricing model. Journal of Financial Economics 116(1): 1–22.
Carhart (1997). On persistence in mutual fund performance. Journal of Finance 52(1): 57–82.
Engle (1982). Autoregressive conditional heteroskedasticity. Econometrica 50(4): 987–1007.
Bollerslev (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics 31(3): 307–327.
Andersen, Bollerslev, Diebold, and Labys (2003). Modeling and forecasting realized volatility. Econometrica 71(2): 579–625.
Diebold and Yilmaz (2012). Better to give than to receive: Predictive directional measurement of volatility spillovers. International Journal of Forecasting 28(1): 57–66.
Harvey, Liu, and Zhu (2016). … and the cross-section of expected returns. Review of Financial Studies 29(1): 5–68.

AI and financial deployment

BIS Innovation Hub (2024). Project Aurora: The power of data, technology and collaboration to combat money laundering across institutions and borders.
Auer, Frost, and Vidal Pastor (2024). The fintech credit landscape in Asia. BIS Working Paper 1145.
McKinsey & Company (2024). The state of AI in early 2024: Gen AI adoption spikes and starts to generate value.
McKinsey & Company (2026). The AI transformation manifesto: 12 themes driving growth.
Stanford HAI (2025). AI Index Report 2025.

Cases

Robertson and Hjuler (2009). Innovating a turnaround at LEGO. Harvard Business Review September.
DBS Bank Annual Reports (2017–2025), particularly the technology and digital transformation sections.
JPMorgan Chase Annual Reports (2017–2025), particularly the technology investment and AI disclosures.
Klarna AB (2024). One year of Klarna AI assistant. Press release, 28 February 2024.
Klarna AB (2025). CEO interview, May 2025: reversal of full-AI customer service strategy.

Regulatory and policy

European Commission (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council (Artificial Intelligence Act). Official Journal of the European Union.
National Institute of Standards and Technology (2023). AI Risk Management Framework (AI RMF 1.0).
ISO/IEC (2023). ISO/IEC 42001:2023 Information Technology — Artificial Intelligence — Management System.
US Federal Reserve (2025). SR 11-7 Updated Guidance on Model Risk Management.
Monetary Authority of Singapore (2022). Veritas Initiative: Fairness, Ethics, Accountability and Transparency in the Use of AI in Financial Services.