Erste Builds AI as a Governed Platform Inside Digital Banking

For Maurizio Poletto, the difference between the first wave of executive excitement about AI and the current moment is not that the board suddenly became interested. Erste Group’s board-level excitement, he said, was present “from day one.” What changed is that the excitement is now attached to operational problems: how to roll AI out to employees, how to get them to use it, how to make it legal and compliant, and how to persuade security teams to open up rather than shut the technology down.

That is the point at which Poletto distinguishes AI from earlier technology cycles. He compared the current moment with the blockchain enthusiasm of five or six years ago, when “every problem in the world” was briefly expected to be solved by blockchain. In his telling, blockchain enthusiasm did not produce the same concrete list of implementation challenges. AI has. The challenges are not incidental; they are the work required before value appears.

Poletto called the technology “not a hype” and “here to stay,” but the substance of his argument is less about conviction than about organizational consequence. Once the bank starts putting AI into employee tools and customer-facing systems, the board’s belief is tested against compliance, security, product quality, and customer trust.

Folley Ogundele framed the central constraint as the regulated environment in which Erste operates: governance, data residency, compliance, auditability, and related requirements shape the speed of adoption. Poletto’s boundary was more specific. In banking, the point where a team has to slow down is when AI starts accessing or processing customer data.

The reason is not only regulatory. Poletto described banking as “an industry built on trust.” Customer data is among the most sensitive information the bank holds, and losing trust is not merely expensive; it takes time to regain. That is why he rejects a bootstrap approach once customer data is involved. At that moment, the institution has to “do it properly.”

Erste made an unusually hard choice early. According to Poletto, when the group began engaging with AI about two and a half to three years ago, many banks chose to begin with data that did not include customer information: help-center material, general information, data on public portals, and other content away from customer data. Poletto called that “a smart approach.” Erste chose differently.

It started, he said, “the hard way,” connecting customer data from day one. He acknowledged “lots of mistakes” in the process, but said the decision came from the view that customer-data integration would eventually be necessary. If that was the destination, the bank preferred to confront the hard part immediately rather than build around it.

That decision added friction. It required compliance and security to be involved early, and Poletto said it “put a bit of gray hair” on the team’s head. But he argued that the accumulated knowledge now gives Erste more confidence to move faster. The hard work did not remove the constraints. It made the organization more capable of operating within them.

The group’s adoption strategy split AI into two tracks. For productivity tools, Poletto said Erste allowed countries significant freedom. Because he sits at the holding level rather than running a specific country, he did not want the group function to become a bottleneck for local teams building tools that help employees work with documents, working instructions, policies, and similar content. In that domain, the instruction was to move quickly, use the available tools, and learn as much as possible.

Customer-facing AI was governed differently. Poletto said the rule was clear: local teams were not to touch customer-facing AI. That work would be done centrally. The reason was tied to Erste’s existing architecture. The group has one digital platform across its countries, and any AI conversation, advisory, or support experience in the app using customer data would be built centrally.

The contrast is specific: a country team could move fast on an employee chatbot for policies or internal documents; it could not independently build an AI customer conversation inside the banking app. The local track was meant to maximize learning and avoid group-level bottlenecks. The central track was meant to keep control over customer data, product consistency, and customer-facing advisory and support.

The internal rollout of tools such as ChatGPT therefore did not simply introduce employees to AI. From Poletto’s description, it also created space for local experimentation while the group preserved stricter control over the AI systems customers would actually encounter.

Erste’s customer-facing work exposed a more subtle adoption problem: customers do not necessarily use conversational AI the way the bank’s product teams imagine they will. Poletto said the bank learned that “not all our customers ask the same question we would ask.” Many retail customers, he said, are not yet ready to have a conversation with the bank inside a digital channel, even if they are comfortable speaking to a person in a branch.

The distinction matters because George, the digital banking app that Ogundele described as used by millions of customers, is already used functionally. Customers use it as a tool or service. They check where they stand, make transfers, and complete tasks. Poletto’s point is that shifting from a tool to a “companion” requires customer learning. It is not enough to add an AI interface and wait for demand to appear.

Erste has therefore worked with both reactive and proactive AI patterns. Reactive AI means the customer opens the conversation and begins asking questions. Proactive AI means the bank pushes conversational nudges to customers and observes how they respond. Poletto said the stronger results came from the proactive side, because prompts give customers an idea of what kind of conversation they can have.

That observation avoids two common assumptions: that AI adoption is blocked only by skepticism, or that interest naturally converts into use. Poletto’s claim is different. Customers may be positively disposed toward AI and still lack a mental model for how it applies to banking. The bank’s current strategy, he said, is to balance proactive and reactive approaches so that these conversations become familiar at the customer level.

When the question turned to what George should become — self-service, a financial coach, or something else — Poletto resisted a simple answer. He said he did not know whether banking will become conversational. His skepticism comes from earlier predictions around voice assistants. When Alexa arrived, he recalled consultants and companies telling banks that people would stop using apps and start asking questions instead.

Poletto’s test was mundane: he commutes by public transportation, and he said that when he first sees someone ask their phone, “What’s my balance?” he may change his mind. He has not seen that behavior yet. He also described a promotional scenario in which a customer asks Alexa for the last transaction on their account, only for Alexa to instruct them to open the banking app and sign in before the answer can be provided. To Poletto, that exposed something wrong with the imagined experience.

His point is not that conversational interfaces have no role. He explicitly did not rule out a future in which a customer can say, “Send the money somewhere,” and the system does it. But he argued that if an interface is well built, some transactions may still be easier through buttons and numbers than through speech or chat. Conversational capability is not automatically an improvement.

The more important ambition, in Poletto’s view, is not replacing every interaction with conversation. It is expanding access to financial advice. Erste currently provides financial advice to roughly 20% of its customer base, he said: the customers who choose to come into branches. The remaining 80% use the bank’s tools functionally but do not receive the same advisory support.

Poletto described that missing 80% as likely including the people who need advice most. The customers who come to branches are often wealthier, older, and able to make time for in-person service. The bank serves them well, he said. The larger challenge is serving everyone else in a way that can affect their lives.

That is where AI becomes strategically useful. Poletto compared the bank’s role to a doctor taking care of one of the most important things in a person’s life: money. He joked about whether money ranks second after health or behind love, but the underlying point was serious. The bank’s ambition is to serve all customers, “not just the wealthy one,” and AI is useful if it helps make that possible.

Folley Ogundele asked what role a partner such as OpenAI should play in a bank’s AI transformation: technology provider, co-builder, strategic adviser, or all three. Poletto turned the question back on OpenAI’s consumer success.

He said OpenAI had put banks “in a very bad situation” because ChatGPT is a successful consumer product that has set a customer-experience standard. Poletto described himself as a pro user and called it his “best $23” per month. The problem for a regulated bank is that customers will compare any AI experience inside the bank’s app with the standard set by consumer AI products, not with the bank’s constraints.

No customer, in his view, will excuse a worse AI conversation because the provider is a bank operating under risk, security, and regulatory requirements. They will simply say OpenAI can do it better. That raises the bar for what Erste must deliver, even when its operating environment is more restricted.

Ogundele accepted the challenge and compared the shift in expectations to Amazon’s effect on delivery. Once consumers learned that delivery could happen in one day, everyone else had to catch up, even if that standard had initially been costly. Poletto agreed with the analogy. The bar is high, he said, and there is no room for a “partially good experience.” Either the experience is good or it is not.

The partner role, as Poletto framed it, is not just supplying models or tools. OpenAI has helped establish the standard customers now bring into other digital experiences. Banks then have to meet that standard while still working through risk departments, security departments, regulatory requirements, and the limits of customer-data handling.

Maurizio Poletto’s main advice to organizations starting in regulated financial services was to accept mistakes. He said Erste is building AI as “a platform within the platform”: the group already has a digital banking platform, and AI is becoming a platform inside it.

That platform is already on its second version after roughly two years. Poletto emphasized that this was not a minor iteration. The first version had to be redone completely. He is not ruling out that a third version, in a couple of years, may also need to be completely different.

He compared the process to startup architecture. A small company often begins with a monolithic application; once it reaches scale, it discovers that the architecture is not scalable and must be reconsidered. Poletto argued that the same willingness to rebuild is necessary for AI in a bank. Some organizations would treat that as wasted money. Erste, he said, treats it as “the way to do it.”

The important operational question becomes how quickly the institution can redo the work. Poletto said he is already planning for the space to rebuild again in perhaps a year and a half, and he hopes coding assistants and related tools will make full rewrites faster. The goal is not to avoid getting the first version wrong. It is to learn quickly enough that rebuilding becomes part of the operating model.

In Poletto’s account, regulated-bank AI adoption is neither a free-for-all nor a static program. It requires separating fast local experimentation from centrally governed customer systems, confronting customer data if that is where the value will be, teaching customers what AI can do rather than assuming they know, and meeting consumer-grade expectations under banking-grade constraints.