Lecture 3: Agentic AI & LLMs in Finance

From LLMs to agents · applications · failure modes · EU AI Act

Prof. Dr. Andre Guettler Director of the Institute
Helmholtzstraße 22, Room 205
andre.guettler@uni-ulm.de
+49 731 50 31 030

Oliver Padmaperuma Doctoral Candidate
Helmholtzstraße 22, Room 203
oliver.padmaperuma@uni-ulm.de
+49 731 50 31 036

3.1 Course objectives

• 3.1 Course objectives
• 3.2 Recap from Lecture 2
• 3.3 LLMs in finance: what they are
• 3.4 Agentic AI: from answer to action
• 3.5 Where LLMs are deployed today
• 3.6 Failure modes
• 3.7 Governance & EU AI Act
• 3.8 Conclusion of Lecture 3

• Welcome to
• Course Objective
• Course at a glance (1/3)
• Course at a glance (2/3)
• Course at a glance (3/3)
• Assignments / Exams

Welcome to Emerging Technology & Finance

• This is a flipped-classroom Bachelor course: every regular lecture (odd weeks) is followed by a flipped session (even weeks) where all groups present on the same topic. There is no exam to register for — sign up on the course Moodle page by 15 October 2026 so you receive announcements and the token-allocation quiz links.
• Form a group of 4 by the end of Week 1 (Moodle sign-up sheet). Stragglers will be allocated by the lecturers.
• Grading is 100% cumulative across the 6 flipped sessions: each session = 50% peer-allocated tokens + 50% lecturer evaluation. Each group gets 20 fresh tokens every flipped week to allocate to other groups via a Moodle quiz within 5 minutes of the session ending.
• Submission per session: upload your slide PDF to Moodle before each flipped session starts. Ask questions during or right after each session — that is the preferred channel.
• Admin / studies / exam-eligibility questions go to the registrar’s office (Studiensekretariat) at studiensekretariat@uni-ulm.de.
• Course-content questions outside class: email oliver.padmaperuma@uni-ulm.de, CC andre.guettler@uni-ulm.de.
• We also recommend the student advisory service.

Course Objective

Scope

We will:

• Survey six emerging-technology modules at the cutting edge of finance: agentic AI · blockchain & DeFi · fintech business models · RegTech & cybersecurity · CBDCs
• Pair every regular lecture with a flipped session in which every group presents their angle on the topic
• Train critical evaluation, presentation, and peer-judgment skills via a transparent token-based peer-grading mechanic
• Place the technologies in a real-world business and regulatory context (PSD2/3, MiCA, EU AI Act, post-quantum standards)

We will NOT:

• Build production-grade fintech systems or trade live capital
• Cover deep technical implementations (we treat code as supplement, not core)
• Run a separate written exam or final-pitch competition — the cumulative flipped-session grade is the entire grade

Approach

Flipped-classroom alternation (12 weeks)

• Odd weeks (W1, W3, W5, W7, W9, W11): regular lecture introducing the topic
• Even weeks (W2, W4, W6, W8, W10, W12): flipped session — all groups present and allocate tokens
• Groups of 4, formed by end of Week 1

Token mechanic (the grading vehicle)

• 20 tokens per group per flipped session
• Each group allocates them to other groups, weighing insight · originality · clarity · critical depth
• Cumulative across 6 sessions = 50% of final grade · lecturer evaluation = 50%

Course at a glance (1/3)

Foundations of Digital Disruption in Financial Services

Week 1

22.10.2026

What is ‘emerging tech in finance’, how did we get here, where is it going

• Three waves of digital disruption in finance
• Today’s actors: incumbents, challengers, Big Tech, infrastructure
• Regulatory backdrop: PSD2, MiCA, EU AI Act
• Why now: structural drivers
• What this course will cover

Flipped — Digital Disruption in Financial Services

Week 2

29.10.2026

Group presentations · token allocation · discussion

• Recap of the foundations lecture
• Group presentations on digital disruption
• Token allocation & next steps

Agentic AI & LLMs in Finance

Week 3

05.11.2026

From LLMs to agents · applications · failure modes · EU AI Act

• LLMs in finance: architecture, training, capabilities
• Agentic AI: from answers to actions
• Applications: RAG, robo-advisors, AML, trading agents
• Failure modes: hallucination, drift, prompt injection
• Governance: EU AI Act and high-risk obligations

Flipped — Agentic AI & LLMs in Finance

Week 4

12.11.2026

Group presentations · token allocation · discussion

• Recap of the agentic AI lecture
• Group presentations on real LLM and agent deployments
• Token allocation & next steps

Blockchain, Crypto, DeFi & Tokenisation

Week 5

19.11.2026

From distributed ledgers to MiCA-regulated markets

• Blockchain primer: ledgers, consensus, smart contracts
• Crypto markets: BTC, ETH, stablecoins
• DeFi primitives: AMMs, lending, derivatives
• Tokenisation of real-world assets
• MiCA framework and EU enforcement

Course at a glance (2/3)

Flipped — Blockchain, Crypto, DeFi & Tokenisation

Week 6

26.11.2026

Group presentations · token allocation · discussion

• Recap of the blockchain & DeFi lecture
• Group presentations on real protocols and deployments
• Token allocation & next steps

Fintech Business Models

Week 7

03.12.2026

Neobanks, embedded finance, BNPL, Open Banking, Big Tech in finance

• Neobanks: N26, Revolut, Monzo, Chime
• Embedded finance & BaaS
• BNPL: Klarna, Affirm, regulatory pushback
• Open Banking & PSD2 outcomes
• Big Tech in finance

Flipped — Fintech Business Models

Week 8

10.12.2026

Neobanks, embedded finance, BNPL · group presentations · token allocation

• Recap of the fintech business-models lecture
• Group presentations on real companies and unit economics
• Token allocation & next steps

RegTech, Cybersecurity & Privacy-Preserving Compute

Week 9

17.12.2026

Industrialising compliance · cyber-threat landscape · ZKPs, MPC, federated learning · post-quantum

• RegTech overview: industrialising compliance
• KYC/AML automation in production
• Cybersecurity threats in finance
• Privacy-preserving compute: ZKPs, MPC, federated learning
• Post-quantum cryptography & the migration

Flipped — RegTech, Cybersecurity & Privacy-Preserving Compute

Week 10

07.01.2027

Group presentations · token allocation · discussion

• Recap of the RegTech & security lecture
• Group presentations on vendors, incidents, and emerging tech
• Token allocation & next steps

Course at a glance (3/3)

CBDCs & the Future of Money

Week 11

14.01.2027

Wholesale vs retail design · Digital Euro · e-CNY · programmable money

• What’s a CBDC: wholesale vs retail
• The Digital Euro state of play
• China’s e-CNY and small-country implementations
• Programmable money: feature, threat, or both
• Stablecoins as private money: tension with CBDCs

Flipped — CBDCs & the Future of Money

Week 12

21.01.2027

Final session · group presentations · token allocation · course wrap-up

• Recap of the CBDCs lecture
• Group presentations on real CBDC projects
• Token allocation, final standings, and course retrospective

Assignments / Exams

Flipped-Classroom Presentation Series 100% of your grade

Six in-class group presentations across six emerging-tech topics, graded cumulatively. Each session: 50% peer-allocated tokens + 50% lecturer evaluation.

Group of up to 4.

Submit by emailing oliver.padmaperuma@uni-ulm.de, CC andre.guettler@uni-ulm.de. Subject pattern: Emerging Technology & Finance_assignment-1-flipped-classroom-presentations_surname1_surname2_…

21 January 2027

3.2 Recap from Lecture 2

• 3.1 Course objectives
• 3.2 Recap from Lecture 2
• 3.3 LLMs in finance: what they are
• 3.4 Agentic AI: from answer to action
• 3.5 Where LLMs are deployed today
• 3.6 Failure modes
• 3.7 Governance & EU AI Act
• 3.8 Conclusion of Lecture 3

• What the flipped session surfaced

What the flipped session surfaced

• The “third wave” candidate the room collectively bet on most strongly — usually agentic AI, sometimes programmable money.
• A regulatory pattern: every disruption story had a regulator either enabling, gating, or trailing it.
• A failure-pattern across Wirecard / FTX / Greensill: governance failed before technology did.

The Week-2 flipped session was deliberately designed to surface the live tensions in fintech: which “third wave” candidate the cohort thinks lands, and why. As you read these recap bullets, look for the threads that connect: regulatory framing turns up in every story; governance failures precede technology failures in every collapse; and the most credible disruption claims are the ones backed by an operational deployment rather than a slide deck. Today’s lecture takes the strongest of those candidates — agentic AI — and dissects what it actually is and where it actually works.

3.3 LLMs in finance: what they are

• 3.1 Course objectives
• 3.2 Recap from Lecture 2
• 3.3 LLMs in finance: what they are
• 3.4 Agentic AI: from answer to action
• 3.5 Where LLMs are deployed today
• 3.6 Failure modes
• 3.7 Governance & EU AI Act
• 3.8 Conclusion of Lecture 3

• A short architecture primer
• What LLMs are actually good at
• Finance-specific LLMs

A short architecture primer

• Transformers — neural networks that process text via “attention” over a sliding context window.
• Tokens — sub-word units (~3–4 characters); a model “thinks” in tokens, not words.
• Pre-training — predict the next token over trillions of tokens of internet text (GPT-4 ≈ 13T tokens of training).
• Fine-tuning + RLHF — calibrate the model to follow instructions and refuse harmful requests.

You don’t need to understand the maths to read fintech AI claims critically — but you must understand the failure-mode shape that the architecture creates.

The transformer architecture turns text generation into an ultra-fast pattern-matching exercise: given the preceding tokens, predict the most likely next one. This is not the same as reasoning, and the gap between “produces plausible text” and “reasons correctly” is the source of every LLM failure mode we’ll discuss today. Pre-training builds the world model; fine-tuning shapes the behaviour; RLHF (Reinforcement Learning from Human Feedback) sands off the rough edges. The order matters — a well-pre-trained model with weak fine-tuning is dangerous; a poorly-pre-trained model with aggressive fine-tuning is just unhelpful (Chen, Yang, and Liu 2023).

What LLMs are actually good at

• Summarisation — earnings transcripts, 10-K filings, research notes.
• Extraction — names, dates, amounts, clauses out of unstructured text.
• Classification — sentiment, topic, anomaly tagging.
• Code generation — boilerplate data plumbing, SQL, simple analytics.

• Numerical precision — six-figure arithmetic is not what they do.
• Causal reasoning — counterfactual thinking under uncertainty.
• Novel-domain planning — anything outside the training distribution.
• Calibration — confidence is decoupled from correctness.

The strong-side bullets are tasks where patterns are the answer; the weak-side bullets are tasks where exactness is the answer. A useful test before deploying an LLM in finance: ask whether the task you want to automate is more like “find the relevant sentence in this filing” (strong) or “compute the precise return of this trade” (weak). The most common deployment failure is to use an LLM for a weak-side task because it sounds like a strong-side task. Note especially the calibration problem — an LLM’s confident-sounding wrong answer is more harmful than an obvious miss, because humans defer to confident text (Acemoglu and Restrepo 2020).

Finance-specific LLMs

• BloombergGPT (2023) — 50-billion-parameter model trained on Bloomberg’s proprietary financial corpus + general data. Closed-source.
• FinGPT — open-source family, instruction-tuned on financial tasks (Chen, Yang, and Liu 2023). Lets researchers replicate and audit.
• J.P. Morgan IndexGPT — internal product for index-methodology generation; commercial offering announced 2024.
• Generic frontier models (GPT-4-class, Claude-class) often beat finance-specific models on general benchmarks — what’s the trade-off?

The finance-specific vs general-purpose LLM debate has shifted since BloombergGPT’s 2023 launch. Early evidence suggested domain-specific training helped on finance tasks; later evaluations on benchmarks like FLUE and FinBen showed frontier general models often matched or exceeded specialised ones — because their general training corpus already included substantial finance content. The trade-off most banks now navigate is control: a specialised model running on-premise can satisfy data-residency and governance requirements that a frontier API call cannot. So the question is rarely “which is more accurate?” — it’s “which fits our governance constraints?” Cite Chen, Yang, and Liu (2023) for the open-source FinGPT family.

3.4 Agentic AI: from answer to action

• 3.1 Course objectives
• 3.2 Recap from Lecture 2
• 3.3 LLMs in finance: what they are
• 3.4 Agentic AI: from answer to action
• 3.5 Where LLMs are deployed today
• 3.6 Failure modes
• 3.7 Governance & EU AI Act
• 3.8 Conclusion of Lecture 3

• What makes a system
• Why agentic ≠ smarter

What makes a system “agentic”

• Tool use — the model can call external functions (search, calculator, database query, API).
• Multi-step planning — the model decides what to do, observes the result, adjusts its plan.
• Memory — state carries across turns; the agent remembers what it tried.
• Autonomy — the agent executes actions without per-step human approval.

Agency is a spectrum, not a binary. A robo-advisor that rebalances quarterly with human sign-off sits low on the spectrum; an LLM that autonomously executes trades sits high.

The agentic shift in 2023–25 was made possible by two changes: model quality crossed the threshold where multi-step plans actually work most of the time, and infrastructure (function-calling APIs, vector databases, agent frameworks like LangChain and LangGraph) made it cheap to wire models into action loops. The implication for finance is consequential: an agent that books transactions or sends customer communications is a fundamentally different governance problem from a chatbot that answers questions. The shift from “AI as a feature” to “AI as a worker” changes who is liable, who is supervised, and what fails when the model is wrong (Agrawal, Gans, and Goldfarb 2022).

Why agentic ≠ smarter

• Agentic systems take the same prediction quality and apply more autonomy to its consequences.
• A wrong prediction made autonomously is far more expensive than a wrong prediction surfaced to a human reviewer.
• The hard question is rarely “is the model accurate enough?” — it is “how much autonomy should this accuracy buy?”
• Many production “AI” systems improve dramatically by being made less agentic, not more.

Consider a 95-percent-accurate fraud detector. In a non-agentic deployment, it flags transactions for an investigator who confirms or dismisses each one; the 5 percent false positives cost time but not customers. In an agentic deployment where the same model autonomously freezes flagged accounts, that same 5 percent false-positive rate produces thousands of wrongly frozen accounts per month and a customer-service crisis. The technology improvement isn’t in the model — it’s in choosing where on the autonomy spectrum to deploy it. This is the central insight of Prediction Machines (Agrawal, Gans, and Goldfarb 2022): AI changes the cost of prediction, but humans must still set the policy around what to do with predictions.

3.5 Where LLMs are deployed today

• 3.1 Course objectives
• 3.2 Recap from Lecture 2
• 3.3 LLMs in finance: what they are
• 3.4 Agentic AI: from answer to action
• 3.5 Where LLMs are deployed today
• 3.6 Failure modes
• 3.7 Governance & EU AI Act
• 3.8 Conclusion of Lecture 3

• A taxonomy of deployed use cases
• Robo-advisors in detail

A taxonomy of deployed use cases

• Earnings-call summarisation
• Equity-research note generation
• ESG-report extraction
• Reading-list curation

• Triage / Tier-1 chatbots
• Internal knowledge-base Q&A
• Onboarding form-filling assistants

• AML transaction-narrative drafting
• KYC document parsing
• Sanctions-screening pre-filter

• Anomaly explanation
• Stress-test scenario generation
• Internal model documentation

• Goal-based portfolio recommendations
• Tax-loss-harvesting assistants
• Plain-language explainers

• News-sentiment signal pipelines
• Strategy-idea brainstorming for quants
• Order-flow-narrative agents (experimental)

The card-grid lays out where LLMs demonstrably operate in finance today. Notice the asymmetry: the Customer service, Research, and Compliance cards have widespread deployment with measurable productivity gains; the Trading card is mostly claims and very thin evidence. When your group researches deployments for Week 4’s flipped session, this asymmetry is the most useful filter — if a vendor pitches “AI trading” with no public performance data, the realistic deployment is in their marketing department, not their trading desk. The honest cases of LLMs adding value sit in the unglamorous middle: summarising filings, drafting AML narratives, parsing KYC documents.

Robo-advisors in detail

• 15-year history; mostly pre-LLM algorithmic portfolio construction (mean-variance with constraints).
• Recent generation adds LLM-driven explanation and onboarding layers; underlying portfolio logic mostly unchanged.
• Combined AUM globally ≈ $1.5T (Betterment, Wealthfront, Vanguard PAS, Schwab Intelligent Portfolios, plus European peers) — small vs incumbent wealth managers’ ~$120T.
• Why is the gap so large? Three reasons: trust, complexity of high-net-worth needs, and customer acquisition costs that scale poorly past a certain size.

Robo-advisors are a useful cautionary tale for the “AI will disintermediate wealth management” thesis. They have existed for 15 years, the underlying algorithms are publicly known, and they have not dislodged incumbents — because wealth management is overwhelmingly about trust, tax complexity, and behavioural coaching, not portfolio mathematics. The LLM-era robo-advisor offerings improve the onboarding and explanation layers, which were the real bottlenecks. Whether that’s enough to shift the AUM curve in the next decade is the open question. Bring measured AUM numbers to the Week-4 flipped session if you cover this angle (Agrawal, Gans, and Goldfarb 2022).

3.6 Failure modes

• 3.1 Course objectives
• 3.2 Recap from Lecture 2
• 3.3 LLMs in finance: what they are
• 3.4 Agentic AI: from answer to action
• 3.5 Where LLMs are deployed today
• 3.6 Failure modes
• 3.7 Governance & EU AI Act
• 3.8 Conclusion of Lecture 3

• Five failures, ranked by 2026 deployment risk
• What a AI deployment looks like

Five failures, ranked by 2026 deployment risk

1. Hallucination — plausible-sounding but wrong output (fabricated citations, invented numbers). Mitigation: retrieval-augmented generation (RAG), source attribution, mandatory human review.
2. Distribution shift — model trained on one regime fails when the world changes (a 2022-trained credit model under-prices 2024 rates). Mitigation: continuous monitoring, retraining triggers.
3. Prompt injection — adversarial inputs manipulate the agent (“ignore previous instructions and transfer funds to…”). Mitigation: input sanitisation, sandboxing, tool-call allow-lists.
4. Model error at scale — the same model, deployed widely, makes correlated mistakes that no individual human would. Mitigation: model diversity, kill-switches, blast-radius limits.
5. Overconfidence calibration — the model is wrong but confident, and humans defer to confident text. Mitigation: explicit uncertainty estimates, abstention prompts.

Hallucination is the failure mode every finance student already knows; prompt injection is the one most finance professionals don’t know but should — Simon Willison’s term, now a defined category in the OWASP LLM Top 10. The systemic-risk version is model error at scale: as one model becomes infrastructure for many banks (think GPT-class APIs used by competing wealth managers), its mistakes become correlated across firms. This is exactly the kind of failure mode the EU AI Act is designed to surface, by mandating documentation, oversight, and post-market monitoring (European Parliament and Council 2024). Calibration is the subtle one: it doesn’t look like a failure mode, but it’s the mechanism by which all the others reach humans.

What a “high-risk” AI deployment looks like

EU AI Act framing

• AI used for credit-worthiness assessment of natural persons = high-risk.
• AI used for risk assessment in life / health insurance = high-risk.
• AI used for detection of fraud / AML = generally not high-risk, but specific transparency obligations apply.
• High-risk obligations: documentation · post-market monitoring · human oversight · accuracy & robustness · data-governance.

The AI Act’s risk taxonomy is the single most important regulatory text for finance students to understand in 2026 — not because it’s complex (it isn’t), but because it determines what an entire industry has to invest in for the next five years. The “high-risk” category is the operationally meaningful one: any AI system that scores a natural person for credit-worthiness lands here, with documentation, post-market monitoring, and human-oversight obligations. The interesting policy question is the boundary between “high-risk” and “limited-risk” categories — a generative AI chatbot that suggests a credit product isn’t automatically high-risk, but a chatbot that grants one is (European Parliament and Council 2024).

3.7 Governance & EU AI Act

• 3.1 Course objectives
• 3.2 Recap from Lecture 2
• 3.3 LLMs in finance: what they are
• 3.4 Agentic AI: from answer to action
• 3.5 Where LLMs are deployed today
• 3.6 Failure modes
• 3.7 Governance & EU AI Act
• 3.8 Conclusion of Lecture 3

• What an AI-governance programme looks like in a bank
• What changes in 2026 vs 2027

What an AI-governance programme looks like in a bank

• Inventory — catalogue every AI model in production with version, owner, business purpose.
• Risk-tier mapping — apply AI Act tiers + institute-specific tiers above the legal baseline.
• Documentation — for each model: training data lineage, intended use, performance, known failure modes.
• Monitoring — drift detection, performance metrics tracked in production, alert thresholds.
• Human oversight — defined decision rights, escalation paths, who has kill-switch authority.
• Third-party governance — when the bank uses a vendor LLM (Microsoft, OpenAI, Anthropic via API), contractual obligations replace direct inspection.

The cheap rebranding of AI governance is “we have an AI committee that meets quarterly.” The expensive version is the inventory: most large banks have between 200 and 2,000 AI models in production across credit, fraud, marketing, operations, and customer service — and many do not know exactly how many. Counting them is the first step every bank has to take before the AI Act’s 2027 obligations bite. If your group’s Week-4 presentation covers a bank’s AI deployment, the most useful question to bring is “how did they inventory their existing models before this deployment?” The answer reveals the maturity of their governance (European Parliament and Council 2024; Arner, Barberis, and Buckley 2017).

What changes in 2026 vs 2027

• AI Act prohibited-practice rules in force (e.g. social scoring by public authorities).
• GPAI transparency obligations on model providers (training-data summaries, copyright compliance).
• Member states designate national supervisory authorities.
• Penalty framework operational.

• High-risk AI obligations in force — the heavy lift for finance.
• Conformity assessments required before market entry.
• Post-market monitoring obligations live.
• Notified bodies active for third-party audits.

The two-stage AI Act rollout is the practical reason finance firms are investing in governance now: the 2026 obligations are mostly disclosure and ban-list compliance, but the 2027 high-risk obligations are operationally substantial. Retroactive compliance is expensive — a bank that hasn’t inventoried its credit-scoring models by mid-2026 will struggle to meet the 2027 deadline. Cite European Parliament and Council (2024) for the exact dates and category definitions. The interesting parallel is that the AI Act mirrors GDPR’s structure (extraterritorial reach, percentage-of-turnover penalties, national authorities) which suggests its enforcement curve will follow GDPR’s: slow at first, then assertive.

3.8 Conclusion of Lecture 3

• 3.1 Course objectives
• 3.2 Recap from Lecture 2
• 3.3 LLMs in finance: what they are
• 3.4 Agentic AI: from answer to action
• 3.5 Where LLMs are deployed today
• 3.6 Failure modes
• 3.7 Governance & EU AI Act
• 3.8 Conclusion of Lecture 3

• Course at a glance (1/3)
• Course at a glance (2/3)
• Course at a glance (3/3)
• Further reading
• Prepare before next flipped session (Week 4)
• See you next time
• References

Course at a glance (1/3)

Foundations of Digital Disruption in Financial Services

Week 1

22.10.2026

What is ‘emerging tech in finance’, how did we get here, where is it going

• Three waves of digital disruption in finance
• Today’s actors: incumbents, challengers, Big Tech, infrastructure
• Regulatory backdrop: PSD2, MiCA, EU AI Act
• Why now: structural drivers
• What this course will cover

Flipped — Digital Disruption in Financial Services

Week 2

29.10.2026

Group presentations · token allocation · discussion

• Recap of the foundations lecture
• Group presentations on digital disruption
• Token allocation & next steps

Agentic AI & LLMs in Finance

Week 3

05.11.2026

From LLMs to agents · applications · failure modes · EU AI Act

• LLMs in finance: architecture, training, capabilities
• Agentic AI: from answers to actions
• Applications: RAG, robo-advisors, AML, trading agents
• Failure modes: hallucination, drift, prompt injection
• Governance: EU AI Act and high-risk obligations

Flipped — Agentic AI & LLMs in Finance

Week 4

12.11.2026

Group presentations · token allocation · discussion

• Recap of the agentic AI lecture
• Group presentations on real LLM and agent deployments
• Token allocation & next steps

Blockchain, Crypto, DeFi & Tokenisation

Week 5

19.11.2026

From distributed ledgers to MiCA-regulated markets

• Blockchain primer: ledgers, consensus, smart contracts
• Crypto markets: BTC, ETH, stablecoins
• DeFi primitives: AMMs, lending, derivatives
• Tokenisation of real-world assets
• MiCA framework and EU enforcement

Course at a glance (2/3)

Flipped — Blockchain, Crypto, DeFi & Tokenisation

Week 6

26.11.2026

Group presentations · token allocation · discussion

• Recap of the blockchain & DeFi lecture
• Group presentations on real protocols and deployments
• Token allocation & next steps

Fintech Business Models

Week 7

03.12.2026

Neobanks, embedded finance, BNPL, Open Banking, Big Tech in finance

• Neobanks: N26, Revolut, Monzo, Chime
• Embedded finance & BaaS
• BNPL: Klarna, Affirm, regulatory pushback
• Open Banking & PSD2 outcomes
• Big Tech in finance

Flipped — Fintech Business Models

Week 8

10.12.2026

Neobanks, embedded finance, BNPL · group presentations · token allocation

• Recap of the fintech business-models lecture
• Group presentations on real companies and unit economics
• Token allocation & next steps

RegTech, Cybersecurity & Privacy-Preserving Compute

Week 9

17.12.2026

Industrialising compliance · cyber-threat landscape · ZKPs, MPC, federated learning · post-quantum

• RegTech overview: industrialising compliance
• KYC/AML automation in production
• Cybersecurity threats in finance
• Privacy-preserving compute: ZKPs, MPC, federated learning
• Post-quantum cryptography & the migration

Flipped — RegTech, Cybersecurity & Privacy-Preserving Compute

Week 10

07.01.2027

Group presentations · token allocation · discussion

• Recap of the RegTech & security lecture
• Group presentations on vendors, incidents, and emerging tech
• Token allocation & next steps

Course at a glance (3/3)

CBDCs & the Future of Money

Week 11

14.01.2027

Wholesale vs retail design · Digital Euro · e-CNY · programmable money

• What’s a CBDC: wholesale vs retail
• The Digital Euro state of play
• China’s e-CNY and small-country implementations
• Programmable money: feature, threat, or both
• Stablecoins as private money: tension with CBDCs

Flipped — CBDCs & the Future of Money

Week 12

21.01.2027

Final session · group presentations · token allocation · course wrap-up

• Recap of the CBDCs lecture
• Group presentations on real CBDC projects
• Token allocation, final standings, and course retrospective

Further reading

• Agrawal, Gans, and Goldfarb (2022) — Prediction Machines (updated ed.) — the economics framing.
• Chen, Yang, and Liu (2023) — FinGPT and the financial-LLM landscape.
• Acemoglu and Restrepo (2020) — sober assessment of when AI helps and when it doesn’t.
• European Parliament and Council (2024) — the AI Act text itself; focus on the high-risk-systems annex.

Prepare before next flipped session (Week 4)

1. Pick your group’s angle from the Week-4 presentation series brief.
2. Find a real deployment, not a press release. Bring measurable evidence (model name, task, false-positive rate, AUM, regulator action).
3. Anticipate the strongest counterargument to your thesis — peers will probe it in Q&A.
4. Upload your slide PDF to Moodle before 14:00 next Thursday.

See you next time

Reminder

• Next session: Lecture 4 — Flipped: Agentic AI & LLMs in Finance on 12 November 2026.
• Each group presents (6 min + 2 min Q&A).
• Slides due on Moodle before 14:00.

References

Acemoglu, Daron, and Pascual Restrepo. 2020. “The Wrong Kind of AI? Artificial Intelligence and the Future of Labour Demand.” Cambridge Journal of Regions, Economy and Society 13 (1): 25–35. https://doi.org/10.1093/cjres/rsz022.

Agrawal, Ajay, Joshua Gans, and Avi Goldfarb. 2022. Prediction Machines: The Simple Economics of Artificial Intelligence. Updated. Boston, MA: Harvard Business Review Press.

Arner, Douglas W., János Barberis, and Ross P. Buckley. 2017. “FinTech, RegTech, and the Reconceptualization of Financial Regulation.” Northwestern Journal of International Law & Business 37 (3): 371–413.

Chen, Hongyang, Hongyang Yang, and Xiao-Yang Liu. 2023. “FinGPT: Open-Source Financial Large Language Models.” arXiv preprint arXiv:2306.06031. https://arxiv.org/abs/2306.06031.

European Parliament and Council. 2024. “Regulation (EU) 2024/1689 Laying down Harmonised Rules on Artificial Intelligence (AI Act).” Official Journal of the European Union, L 2024/1689. https://eur-lex.europa.eu/eli/reg/2024/1689/oj.