Session Introduction

Facilitator Brief — Open Peer Discussion

This is a peer discussion, not a presentation.
Every person in the room has seen AI deliver or disappoint in a real institution.
The goal of the next 75 minutes is to find out why — and how.

🎯 Topics 3 Discussions
⏱ Per Topic ~25 Minutes
💬 Format Discussion
🔒 Rules Chatham House
Topic 1 of 3

Operational Efficiency & Compliance

Automating inefficient processes — and the regulatory risk that comes with it.

~25 minutes allocated (Part 1: Operational Efficiency· Part 2: Compliance & Legal)
Market Context

What the data says right now

$340B
in annual value AI could unlock in banking alone, predominantly through productivity gains and cost reduction
McKinsey Global Institute, 2025
15–25%
cost savings reported by firms that scaled AI across operations, within 18 months of deployment in targeted functions
Deloitte AI in Financial Services Survey, 2025
72% vs 19%
say AI has changed how repetitive workflows run — but only 19% are measuring those gains rigorously
Gartner Financial Services AI Adoption Survey, 2025
What Is Failing & Why

Familiar Patterns:

01
Process automation on top of broken processes
Is AI being applied to the right workflows?
02
The "pilot cemetery"
Proof-of-concepts proliferate but how do they scale?
03
Measuring inputs, not outcomes
Institutions count AI deployments, not value delivered. Do we have the right KPIs metrics? Are we measuring what changed for the customer or the P&L?
04
Shadow AI undermining governance
Staff are using consumer AI tools (ChatGPT, Copilot) outside of sanctioned environments. How institutions are managing such a risk?
Up Next
Continued — Compliance & Legal
AI accelerates everything — including the things that can get you fined.
Part 2 — Compliance & Legal

Compliance & Legal

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

Market Context

What the data says right now

High Risk
EU AI Act classification for most AI in credit decisioning, insurance underwriting, and AML — requiring full documentation, human oversight, and conformity assessment before deployment
EU AI Act, Article 6 & Annex III, 2024
67%
of compliance officers say AI has increased their institution's regulatory risk exposure over the past 12 months — not reduced it
Kroll Global Fraud & Risk Report, 2025
$2.7B
in global AML fines in 2024 — regulators are scrutinising the AI and model-based systems behind compliance decisions, not just the outcomes
ACAMS Global AML Fines Report, 2024
What Is Failing & Why

Familiar Patterns:

01
Models in production without documentation trails
02
AI making decisions that cannot be explained to regulators
Can AI tools generate legally required explainable decisions?
03
Compliance teams excluded from AI design
AI is built by technology and data science teams, What is the compliance teams input in the AI model design and sign off?
04
Shadow AI creating ungoverned liability
Employees using consumer AI for legally sensitive tasks — drafting contracts, advising clients, interpreting regulations — It this liability risk exposure tracked?
Insight

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 a technology question and a governance question many institutions have not yet answered.

EU AI Act, Title III — High-Risk AI Systems, 2024
Closing Topic 1
Up Next
Topic 2 — Generating Alpha
You have the data and maybe the model but not everyone has the speed.
Topic 2 of 3

Generating Alpha

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

25 minutes allocated
Market Context

What the data says right now

65%
of hedge funds now incorporate alternative data into investment processes — satellite imagery, web scraping, NLP on filings — up from 35% in 2020
Man Group Annual Report
$143B
projected size of the global alternative data market by 2030 — growing at 54% CAGR as AI lowers the cost of signal extraction
Grand View Research
15–30% vs reality
AI signals show 15–30% higher Sharpe ratios in backtests — but significantly lower out-of-sample performance when the strategy becomes crowded
Two Sigma Research, 2024
What Is Failing & Why

Familiar Patterns

01
Backtest inflation and overfitting
Models optimised on historical data look brilliant in backtests, but how they perform/underperform live?
02
Alternative data with legal and ethical exposure
Some alternative data sources — scraped web data, location data, inferred personal data — carry regulatory risk under GDPR, FCA, and SEC rules. Are the quant teams assessing such risk?
03
Alpha decay from strategy overcrowding
When competing firms deploy similar models on the same data, How this impacts alpha?
04
LLM hallucination in investment research
Analyst teams using LLMs to synthesise earnings calls and filings are encountering confident, plausible, and factually wrong summaries — What mechanisms are there to detect them at scale.
Insight

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
Closing Topic 2
Up Next
Topic 3 — Implementation Challenges
The pilot worked. Now what?
Topic 3 of 3

Implementation Challenges

The pilot worked. Now what?

25 minutes allocated
Market Context

What the data says right now

18 months
average time from AI pilot to production in financial services — nearly three times longer than comparable technology deployments in fintech and tech-native firms
McKinsey Global AI Survey, 2025
60%
of AI proofs-of-concept in financial services never reach production. This is a documented organisational pattern.
Gartner Financial Services AI Report, 2025
70%
of UK bank IT budgets are spent maintaining legacy infrastructure — leaving 30% for everything else, including AI deployment, modernisation, and security
Bank of England Technology Survey, 2024
What Is Failing & Why

Familiar Patterns

01
Pilots built on curated data that doesn't reflect production reality
Production exposes the AI to the actual data estate — messy, incomplete, inconsistently labelled. How does this degrade the AI model?
02
AI teams operating independently of enterprise architecture
Are AI labs and innovation teams building solutions that cannot be integrated into core systems?
03
No clear production ownership
How is the model degradations - drifts, or fails- are measured and escalated? Who are the owners - in the business line and tech function?
04
Cost unpredictability of LLM APIs at production scale
How accurate business cases measure the cost transition from the pilot to production?
Insight

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
Closing the Session
Up Next
Section 2 — Live Demo
Let's see some production grade AI applications in practice.