Erste Group's AI Transformation in Practice: How a European Bank Balances Compliance and Innovation

Introduction: From Excitement to Execution

At an OpenAI enterprise event in London, Maurizio Poletto, Chief Platform Officer and COO of Erste Group — one of Europe's leading banking groups — shared the group's real-world experience in AI transformation. Erste Group is one of the largest financial services providers in Central and Eastern Europe, headquartered in Vienna, Austria, with operations spanning seven countries: Austria, Czech Republic, Slovakia, Romania, Hungary, Croatia, and Serbia, serving approximately 16 million customers. The group operates one of Europe's largest digital banking platforms — George — which was first launched in Austria in 2015 and has since expanded across the group's markets. Built around a user-experience-first design philosophy, George offers comprehensive digital banking services including account management, payments, investments, and loans, with over 8 million monthly active users. This cross-border unified platform architecture means that any AI feature deployment must account for the complexity of multiple languages, regulatory frameworks, and cultural differences across markets.

Maurizio's talk revealed how a regulated financial institution navigates the balance between compliance and innovation, including the mistakes they've made and the lessons they've learned along the way.

AI Is Not Blockchain: This Time It's Really Different

Maurizio admitted that board-level excitement about AI existed from day one, but what changed was the nature of that excitement — initially it was limited to personal experiences and PowerPoint demos, whereas now it comes with a long list of practical problems to solve: how to roll it out to employees, how to ensure compliance, and how to convince security teams to open up rather than lock down.

He specifically contrasted AI with the blockchain hype from five or six years ago: "Blockchain got everyone excited too — as if every problem in the world could be solved by blockchain. But it never generated a series of real challenges that needed solving. AI is completely different — this isn't hype, it's here to stay."

This comparison carries deep industry context. Between 2017 and 2019, the global banking industry experienced a blockchain frenzy. The R3 consortium (comprising dozens of banks including Goldman Sachs and JPMorgan) developed the Corda platform, and major banks established blockchain labs to explore use cases in cross-border payments, trade finance, and KYC sharing. However, most projects stalled at the proof-of-concept stage and never reached large-scale production. Core reasons included: distributed ledger technology lacked a clear value proposition within existing centralized clearing systems, regulatory frameworks were unclear, and trust issues around data sharing persisted among participants. This history has made banking executives wary of new technology waves, making Maurizio's comparison all the more compelling — AI is different because it immediately generates real operational problems that need solving, rather than remaining theoretical.

This observation is remarkably precise. The hallmark of truly valuable technological change isn't the absence of problems, but the quality and urgency of those problems. When an organization starts seriously discussing data governance, compliance frameworks, and security architecture, that's precisely the signal that the technology has entered a genuine implementation phase.

Choosing "Hard Mode": Connecting Customer Data from Day One

Why They Didn't Take the Easy Route

Many banks chose the "smart" path for AI deployment — starting with scenarios that don't involve customer data, such as help centers and public information portals. But Erste Group made a bold decision: when they first started working with AI two and a half to three years ago, they connected customer data right away.

Maurizio explained the logic behind this decision: "This has to happen eventually, so why not do it now?" Of course, they made plenty of mistakes along the way, and the team earned "quite a few more gray hairs." But he believes it was worth it — the team accumulated enough knowledge and confidence to move forward faster now.

Managing Speed in a Regulated Environment

In banking, the moment you start accessing customer data, you have to slow down. Maurizio emphasized: "We're an industry built on trust. If you lose trust, beyond being expensive, it takes a very long time to rebuild."

European banks face multiple regulatory constraints when handling customer data. GDPR (General Data Protection Regulation) sets strict rules for the collection, processing, and storage of personal data, with fines for violations reaching up to 4% of global annual revenue. Additionally, the European Banking Authority's (EBA) ICT risk management guidelines, PSD2 payment services directive provisions on data access, and DORA (Digital Operational Resilience Act), which began implementation in 2024, all impose specific requirements on banks' technology architectures. In the AI domain, the EU AI Act classifies AI systems in financial services as high-risk applications, requiring conformity assessments, human oversight, and transparency disclosures. This explains why Erste Group's choice to connect customer data from day one, while difficult, ultimately avoids the risk of large-scale compliance retrofitting later on.

Their strategy: compliance, security, and proper processes must be in place before moving forward. This isn't conservatism — it's respect for the fundamental nature of the industry.

A Dual-Track Strategy: Internal vs. External

Productivity Tools: Let Country Teams Move Fast

Erste Group adopted a clear layered strategy. For internal productivity tools (such as chatbots that help employees learn documents, work guidelines, and policies), country teams are free to move quickly. Maurizio didn't want to be a bottleneck for his colleagues across countries — "These tools — get them live as fast as possible, learn as much as possible."

Customer-Facing: Centralized Control

But on the customer side, the rules are entirely different. Since the group operates a unified digital platform across all countries, customer-facing AI conversations, advisory services, and support must be developed centrally. This ensures consistency of experience and controllability of quality.

Customers Aren't Ready for "Conversational Banking"

Proactive Push vs. Passive Response

Maurizio shared an important finding: many retail customers aren't ready to have conversations with their bank. They're accustomed to using digital banking apps as tools — transferring money, checking balances — not as a partner they can talk to.

As a result, Erste Group adopted a two-pronged approach:

• Reactive AI: The customer initiates the conversation
• Proactive AI: The system pushes suggestions, guiding customers to discover conversations they can have

Proactive AI in banking scenarios typically relies on event-driven architecture and real-time data analytics pipelines. The system continuously monitors customer transaction patterns, account status changes, and lifecycle events (such as salary deposits, large expenditures, fixed deposit maturities, etc.), using pre-trained models or rule engines to identify actionable moments, then generating personalized push recommendations. The key challenge lies in balancing push frequency and relevance — too many pushes lead to user fatigue and notification dismissal, while too few fail to deliver AI's value.

The results showed that proactive pushing performed far better than passive waiting. The reason is simple: "Many customers love AI, but they simply don't know what to ask."

Skepticism About Alexa-Style Conversational Banking

Maurizio holds an open but cautious view on the future of "conversational banking." He recalled when Alexa first appeared and consultants predicted people would stop using apps and start asking questions by voice. "I take public transport to work every day, and I've yet to see anyone ask their phone 'What's my balance?'"

He believes that if the interface is well-designed, tapping a button to complete a transaction is still more efficient than voice commands. But he doesn't rule out future possibilities.

The Real Mission: Using AI to Serve the 80% of Neglected Customers

This was the most insightful part of the entire conversation. Maurizio pointed out that currently, Erste Group provides genuine financial advice to only about 20% of its customers — those who choose to walk into a branch, who are typically wealthier, older, and have the time.

The other 80% use the tools purely functionally and have never received financial advice. Ironically, "most of that 80% are precisely the people who need advice the most."

Behind this phenomenon lies a natural economic constraint of the traditional banking model: a financial advisor's annual salary plus operational costs can range from €80,000 to €150,000, meaning only customers with sufficient assets under management (AUM) can receive one-on-one service, typically with a threshold of €100,000 or more. This creates the so-called "advice gap" — low- and middle-income customers, despite needing financial guidance the most (such as debt management, emergency savings, and introductory retirement planning), are excluded from service because it's economically unviable. The UK's FCA (Financial Conduct Authority) estimated in a 2020 study that approximately 20 million adults in the UK fall within this "advice gap." In the Central and Eastern European markets served by Erste Group, this proportion is likely even higher. AI's near-zero marginal cost of service theoretically breaks this constraint, making inclusive financial advice possible.

Erste Group's vision is to use AI to provide financial care for all customers, not just the wealthy ones. "We're looking after the second or third most important thing in people's lives — their money. Our ambition is to serve all customers, not just the affluent ones."

OpenAI Set the Standard: Challenges and Opportunities

Maurizio raised an interesting "complaint" about OpenAI: as a successful consumer product, ChatGPT has set the bar for user experience. "Nobody is going to say 'Your AI conversation is a bit worse than OpenAI's, but since you're a bank I'll accept it.' They'll just say 'This is terrible, OpenAI does it better.'"

This means that banks, despite all their industry constraints (risk management, security, compliance), must still meet the experience standards of consumer-grade AI products. It's an extremely high bar, but it's also what drives the industry forward.

Core Advice: Embrace Failure, Iterate Fast, and Rebuild

Maurizio's final advice was direct and pragmatic: accept that you'll make mistakes. Erste Group's AI platform is already on its second version — the first was completely rebuilt. He's even already planning for a potential third rebuild in a year and a half.

"In some organizations, this might be seen as wasting money. In our organization, we believe this is the right way to do it."

This attitude has deep theoretical support in software engineering. Google's classic 2015 paper Hidden Technical Debt in Machine Learning Systems pointed out that the actual machine learning code in ML systems represents only a small fraction — the surrounding data pipelines, feature stores, monitoring systems, and serving infrastructure are the primary sources of complexity. In the AI platform space, technical debt accumulates especially fast because underlying model capabilities undergo generational shifts every 6-12 months — from GPT-3.5 to GPT-4 to GPT-4o, each leap can invalidate previous architectural assumptions. Erste Group's choice to embrace rebuilding rather than patching reflects a mature engineering culture — acknowledging that in a rapidly evolving technology landscape, over-engineering and reluctance to rewrite are equally wasteful.

The key isn't avoiding rebuilds, but making each rebuild faster. This mirrors the growth trajectory of startups: start with a monolithic application, discover it doesn't scale, then redesign. The only difference is that with AI coding assistants (such as GitHub Copilot, Cursor, etc.), the speed of refactoring keeps increasing, and the human cost of code rewrites is dropping significantly.

Conclusion

Erste Group's AI transformation story teaches us this: there are no shortcuts to advancing AI in regulated industries, but there is a right attitude — confront the hardest problems head-on, accept the inevitability of failure, maintain reverence for customer trust, and always remember that the ultimate purpose of technology is to serve more people.

Key Takeaways