🔒 Facilitator Only

Welcome & Session Frame

"The most useful thing you can do in the next 75 minutes is say what you actually think."

2 minutes — then straight into Topic 1

Your opening — 2 minutes, no slides

Tone. Speak as a peer — not a host. Sit down if you can. The moment you stand at a lectern, this becomes a presentation. It isn't.

Pace. Slow. Two minutes is longer than it feels. Let silence do work.

1 Welcome & Who's in the Room

Good morning. thank you for being here today, Someone said to me, the world is being kidnapped by a F1 car, he was referring to how AI is changing our workplaces and our lives.

We have over 20 senior people in this room — most of you work in significant financial institutions buy-side, sell-side and insurance, most of you have to use AI on daily basis and most importnatly make decision about the use of AI in your respecitve firms right now.

The aim of this event is to share your throught and reflections on how AI impacts our industry in certain aspect as we cannot discuss them all in one session. We'll also see some actual use cases.

I'm asking you to express your opinions and views freely about what's working and what isn't. Please provide actual examples to enrich the conversation.

I am here to facilitate — My job is to ask questions, and make sure all people in this room participate and get the chance to express their views and disagreement.

We will cover what leading institutions are doing — but we won't pretend that headline numbers tell the full story. They usually don't.

2 House Rules

Let's start first with some house rules: "Everything discussed today is under Chatham House Rules — insights and themes may be taken out of this room; attribution may not."

Views expressed are of the Individuals stating them views and not the institutions they work for. Please keep your phones on silent. Coffee available throughout. We will finish shirtly after 10AM. There is a short demonstrations after the discussions — approximately 3-5 minutes each — before we close.

3 Frame the Landscape — 20 Seconds

We will cover three areas: operational efficiency and compliance, generating alpha, and what it actually takes to get AI into production at scale. We have 25 minutes per topic — and I will start each one with a context and a question, Let's dive in.

let's see some AI figures in financial services right now: the opportunity is real and large — $340 billion in addressable value by some estimates. The challenge is equally real. Most institutions are not automating faster enough and their governance and complaince are not adapting appropriately enough to be ready for this systemic shift. Also Most pilots are not reaching large scale production. And the regulatory environment is about to get significantly more demanding. We are going to talk about all of it.

"Let's start with something concrete: where in your organisation has AI had the most tangible operational impact in the last 12 months — and how are you measuring it?"
First Topic
Topic 1 — Operational Efficiency & Compliance
Automation without redesign is just faster bureaucracy.

Operational Efficiency & Compliance

"AI accelerates everything — including the things that can get you fined."

~25 minutes allocated — two parts

Operational Efficiency

"Automation without redesign is just faster bureaucracy."

The real debate in the room

Every institution is automating.
But automation without redesign is just faster bureaucracy.

The real question is not whether AI is reducing headcount or cost —
it is whether anyone has fundamentally changed how work gets done.

Efficiency Theatre
Deploying AI on top of broken processes and calling it transformation
Structural Redesign
Using AI as the catalyst to rethink workflows, roles, and operating models

Many banks are doing the former and reporting the latter. That's the honest truth worth surfacing.

If the group is speaking guardedly: "There's a version of this where AI is doing the equivalent of automating a fax machine. Are any of us guilty of that?"

Tier-1 examples — use as conversation starters, not benchmarks

JPMorgan Chase
COIN — Contract Intelligence
Processes 12,000 commercial credit agreements in seconds. Previously required 360,000 lawyer-hours annually.
JPMorgan Annual Report, 2023
"What's your equivalent of COIN? What process are you still doing the way JPMorgan used to process credit agreements?"
Goldman Sachs
LLM-Assisted Code Generation
Engineering teams use AI to generate and review code. Internally reported 20–40% productivity uplift in targeted workflows.
Goldman Sachs Investor Day, 2024
This stat draws scepticism — surface it. "Does that number feel right to you? If not, what's yours?"
HSBC
AI-Powered AML Monitoring
ML models reduced false positives in AML transaction screening by ~60%, freeing compliance analysts for higher-value review.
HSBC ESG & Operations Report, 2024
"The gain here came from reducing noise, not reducing headcount. Is that the pattern you're seeing?"

Do not present these as "this is what good looks like." Use them as prompts to surface the room's experience. If someone works at one of these institutions, ask them to add texture or correct the record.

Three questions — safe to provocative

Q1 Opening — Safe
"Where in your organisation has AI had the most tangible operational impact in the last 12 months — and how are you measuring it?"
Gets people speaking in specifics. The measurement tail often reveals the most.
Q2 Probing
"When you look at the processes you've automated, how many were redesigned before the automation was applied — versus taken as-is?"
Pivots from "what did you do" to "how did you think." Most groups pause here. That pause is valuable.
Q3 Provocative — Use with care
"Is there a version of your AI operational programme that is genuinely transformative — or are you optimising a business model that AI will eventually make obsolete anyway?"
Deploy only if the conversation has been confident and self-congratulatory.

Reframe: "Let me ask it differently — where has AI disappointed you operationally? What did you build that didn't deliver?"

Negative examples almost always unlock more honest responses than positive ones in a room of peers.

The banks saving the most from AI are not the ones with the most AI.

They are the ones who stopped doing things.

The productivity gain is not in the tool.
It is in the decision to eliminate the process entirely.

Synthesised from BCG Financial Services AI Value Study, 2024 & McKinsey AI adoption research

Deploy when the conversation has circled around "efficiency" without anyone questioning what they're being efficient at.

Follow with: "Has anyone in this room actually killed a process — not automated it, killed it — because AI gave you the confidence to stop doing it?"

That question generates either a very good story or a very honest silence. Both are useful.

Synthesise 1–2 themes that emerged. Name them explicitly before moving on.

"What I'm hearing is [theme]. We'll come back to the governance dimension of that when we hit compliance. Let's go there now."
Continuing
Part 2 — Compliance & Legal
"AI accelerates everything — including the things that can get you fined."

Compliance & Legal

"AI accelerates everything — including the things that can get you fined."

The real debate in the room

The most powerful models are often the least explainable.
Regulators want both: accuracy and transparency.

The question is not whether your AI is right
it is whether you can prove to a regulator that it is right, and fair, and auditable.

Black-Box Performance
High accuracy. Low interpretability. Regulators cannot audit it. Lawyers cannot defend it.
Governed, Auditable AI
Explainable decisions. Model risk documentation. Human oversight built in. Slightly less performant.

Most institutions are quietly choosing performance and hoping the regulator doesn't ask. The explainability trade-off is real — and most risk committees haven't been shown it explicitly.

Frame it directly: "How many of you are running AI models in production that you couldn't fully explain to your model risk committee if they asked today?"

Most people won't answer directly. The silence is the answer.

Tier-1 examples — use as conversation starters, not benchmarks

Citi
Explainability Layer for Credit AI
Built model-agnostic explanation layers on top of credit decisioning AI, translating outputs into plain-language adverse action notices that satisfy fair lending requirements.
Citi Technology & Operations Report, 2024
"If a customer challenges a credit decision made by your AI, what does the adverse action explanation actually say — and who wrote it?"
Deutsche Bank
AI Governance Charter
Published a formal AI governance charter following EU AI Act passage — establishing risk tiers, mandatory human review thresholds, and cross-functional AI oversight committees.
Deutsche Bank AI Governance Framework, 2024
"Have you published your AI governance commitments externally? If not — what would it take, and what are you not ready to commit to yet?"
HSBC
SR 11-7 Aligned Model Risk for AI
Extended the SR 11-7 model risk management framework to cover ML models, with annual validation cycles and explicit documentation of model purpose, limitations, and override thresholds.
HSBC Model Risk Management Framework, 2023
"Are your AI models in the same risk management framework as your traditional models — or in a parallel track your CRO hasn't fully reviewed?"

These are structural governance choices, not technology choices. The question is whether the institution has made AI governance a first-class function. Deutsche Bank's external publication is a strong prompt — most firms have internal commitments they'd never publish. Ask why.

Three questions — safe to provocative

Q1 Opening — Safe
"How many AI systems in your institution are currently operating in production without a completed formal model risk review?"
Most people won't know the exact number. That uncertainty is itself the signal. Let it sit.
Q2 Probing
"If your regulator asked you today to explain the last hundred decisions made by your most critical AI model — how confident are you in that explanation?"
Forces a concrete test of explainability readiness. Most rooms go quiet. Let the silence sit for a full five seconds before moving on.
Q3 Provocative — Use with care
"Is your AI governance framework actually governing AI — or governing the AI you know about? Because those are not the same thing."
The shadow AI problem in disguise. Deploy when the room has been confident about its governance posture.

Reframe: "Has your compliance team ever discovered an AI system in production that they didn't know existed? What happened?"

Almost every institution has a version of this story. Getting it out unlocks the rest of the conversation.

The EU AI Act does not ask whether your AI works.

It asks whether you can prove it works safely, fairly, and accountably.

That is not a technology question.
It is a governance question many institutions have not yet answered.

EU AI Act, Title III — High-Risk AI Systems, 2024

Use when the conversation has stayed in the abstract — people discussing regulation in principle but not in practice.

Follow with: "Who in this room has mapped their AI systems against the EU AI Act's high-risk classification list? And what did you find?"

The answers — and the silences — will drive the last five minutes of this topic.

Synthesise 1–2 themes that emerged. Name them explicitly before moving on.

"What I'm hearing is that governance is lagging deployment — and the gap is widening. Let's go to alpha generation — where the pressure to deploy fast is nowhere higher."
Up Next
Topic 2 — Generating Alpha
"Everyone has the data. Not everyone knows what to do with it."

Generating Alpha

"Everyone has the data. Not everyone knows what to do with it."

~25 minutes allocated

The real debate in the room

AI can find patterns no human analyst would spot.
But when enough firms deploy the same AI on the same data, the pattern disappears.

The edge is not in the algorithm.
It is in what you feed it that no one else has.

First-Mover Alpha
Novel signal + fast deployment = genuine edge, before competitors find the same pattern
Commoditised Alpha
The same models on the same data = crowded trades and rapid alpha decay

The crowding problem is the most honest conversation you can have in this space. Most firms using "alternative" data are using the same vendors. The data stopped being alternative two years ago.

Provocation: "If your alpha model uses satellite data and NLP on earnings calls — how many of your direct competitors are running the same signals on the same data right now?"

Tier-1 examples — use as conversation starters, not benchmarks

Two Sigma
Data as Manufacturing
Treats data sourcing and signal extraction as a manufacturing process — with rigorous pipeline engineering, decay monitoring, and signal retirement protocols when alpha erodes.
Two Sigma Research, 2023
"Do you have a process for retiring AI signals when they stop working — or do you keep running them and wonder why performance is down?"
Man Group / AHL
Dimension Fund AI Allocation
AI dynamically allocates capital across asset classes and strategies within the Dimension Fund, with continuous learning on live data rather than static backtested parameters.
Man Group Annual Report, 2024
"Is your AI trained on live performance data — or is it still running parameters optimised on historical data from a different market regime?"
JPMorgan
ThemAtic Investment AI
AI-driven thematic investing platform identifying structural economic shifts from unstructured data sources — patent filings, regulatory documents, supply chain data.
JPMorgan Research, 2024
"Where is your firm finding signal that isn't already in the Bloomberg terminal — and how defensible is that source?"

Two Sigma's signal retirement point is particularly useful — it surfaces the question of whether firms are measuring alpha decay at all. Most aren't systematically.

If anyone in the room is from an asset manager, the Man Group AHL example tends to prompt honest comparison of their own live vs. backtested performance gap.

Three questions — safe to provocative

Q1 Opening — Safe
"Where in your investment process has AI delivered a demonstrably repeatable edge — and how are you measuring that it is still working?"
The measurement tail is key. Most can describe where they've deployed AI; few can show live performance attribution that isolates the AI contribution.
Q2 Probing
"How much of the alpha you're attributing to AI is genuinely novel signal — versus faster execution of a thesis a human analyst already had?"
This is the question most quant teams avoid asking themselves. Speed is valuable, but it is not alpha generation. The distinction matters for how you invest in the capability.
Q3 Provocative — Use with care
"If your AI edge depends on alternative data that three of your largest competitors also subscribe to — are you generating alpha, or just running faster toward the same cliff?"
Deploy when the room has been confident about its data sources. Crowding is the most underacknowledged risk in quant AI.

Reframe: "Where has AI disappointed you in the investment process? What did you build that looked great in backtest and underperformed live?"

This almost always generates a real story. And real stories unlock the room.

Most AI in investment management is not generating new alpha.

It is helping humans express existing convictions with less friction.

That is genuinely valuable.
But it is not the same thing as finding signal no one else has found.

Synthesised from Two Sigma Research, 2023 & BCG Asset Management AI Survey, 2024

Deploy when the conversation has been about speed and efficiency rather than genuine signal discovery. It reframes what the room is actually debating.

Follow with: "If your AI is helping you execute faster on views your PMs already had — is that what you told your investors you were doing?"

That question either generates great candour or a very revealing non-answer.

Synthesise 1–2 themes that emerged. Name them explicitly before moving on.

"What I'm hearing is that the edge is real but fragile — and that most of us are better at generating alpha on paper than in practice. The final topic brings it all together: what does it actually take to move from pilot to production at scale?"
Up Next
Topic 3 — Implementation Challenges
"The pilot worked. Now what?"

Implementation Challenges

"The pilot worked. Now what?"

~25 minutes allocated

The real debate in the room

Modern AI requires clean, accessible, well-governed data
and flexible, cloud-native infrastructure.

Most financial institutions have
the exact opposite of both.

Cloud-Native Aspiration
AI strategy built for flexible, scalable infrastructure. Modern data pipelines. Rapid iteration. Continuous deployment.
Legacy Reality
30-year-old core banking systems. Siloed, dirty data. Change management cycles measured in quarters. IT budget already spoken for.

The honest version of this tension: most banks are building AI on a foundation that was never designed to support it. The AI strategy is more advanced than the infrastructure it's supposed to run on.

Provocation: "When you look at your AI ambitions and your actual infrastructure position — are those two things compatible? And if not, which one changes?"

Tier-1 examples — use as conversation starters, not benchmarks

ING
Cloud-First Infrastructure Transformation
Migrated core workloads to cloud-native infrastructure as a prerequisite for AI at scale — with explicit board commitment that legacy migration was an AI enablement investment, not a cost project.
ING Annual Report, 2024
"Has your board been told explicitly that legacy infrastructure is an AI bottleneck — and what was the response?"
JPMorgan
Unified Data Platform (Fusion)
Built a single internal data platform — Fusion — as the foundational layer for all enterprise AI. Standardised data access, governance, and lineage across every business line before scaling AI deployment.
JPMorgan Technology & Data Report, 2024
"What is the single source of truth for data in your AI systems — and when was the last time someone independently validated its quality?"
Starling Bank
Born-Cloud Advantage
With no legacy infrastructure to maintain, Starling deploys AI features in weeks rather than quarters — demonstrating the compound competitive advantage of cloud-native foundations over time.
Starling Bank Annual Report, 2024
"What would your AI deployment timeline look like if you had Starling's infrastructure? And what does the gap tell you about your current investment priorities?"

The Starling comparison is the most useful provocation — not because traditional banks should become fintechs, but because the deployment speed difference is a quantifiable measure of infrastructure debt.

JPMorgan's Fusion platform is worth dwelling on: the data foundation came before the AI. Most institutions are trying to do it the other way round.

Three questions — safe to provocative

Q1 Opening — Safe
"What is the single biggest infrastructure or data constraint that has slowed your AI deployment — and is it actively being addressed or just being worked around?"
"Being worked around" is the most common honest answer. Surface what that means in practice — it usually reveals a structural problem that won't resolve itself.
Q2 Probing
"When you've moved an AI from pilot to production — what was the thing that nearly killed it, that you didn't anticipate in the pilot phase?"
This request for a specific story almost always generates candid responses. The gap between pilot and production is where institutional reality reasserts itself.
Q3 Provocative — Use with care
"If you had to build your current AI infrastructure from scratch today — knowing everything you know now — what would you do completely differently?"
Deploy when the room has been candid. The answers reveal what people actually believe about their current architecture — versus what they say in public.

Reframe: "Tell me about an AI pilot that never made it to production. What killed it — and was it the technology, the data, the organisation, or the politics?"

That last list — technology / data / organisation / politics — tends to generate immediate, specific, and very honest answers.

The firms furthest ahead on AI implementation are almost never the ones with the best AI strategy.

They are the ones with the cleanest data.

Your AI programme is only as fast
as your data estate will allow.

Synthesised from McKinsey Global AI Survey, 2024 & JPMorgan Technology & Data Report, 2024

Deploy when the conversation has focused on models, tools, and vendors — without anyone naming data quality as the real constraint.

Follow with: "On a scale of one to ten — how would you honestly rate the quality and accessibility of the data your AI currently trains on? And what is it costing you that it's not higher?"

No one says ten. The gap between their number and ten is your implementation constraint.

Synthesise the thread across all three topics. What was the recurring theme? Name it explicitly.

"What I'm hearing across all three conversations is [theme]. That's the thing worth taking back from today. In a moment we'll shift from discussion to demonstration — and you can judge for yourselves whether the technology is ahead of or behind the conversation we just had."
Up Next
Section 2 — Live Demo
"Let's see what this actually looks like in practice."