Erste Group, UBS, and Deutsche Bank: When AI Decides on a Bank Before the Customer Makes Contact

Erste Group, UBS, and Deutsche Bank are investing massive sums and substantial organizational resources into artificial intelligence. This is not only transforming banking operations; it also raises a new question for corporate communications: What happens to a bank's perception when customers no longer research information themselves, but have it summarized by AI?

It is a question that has barely been raised in bank boardrooms until now: What does an artificial intelligence answer when a potential customer asks about your bank? Not about current interest rates or how the stock price has developed, but about something more fundamental: How innovative is this bank? How robust is its digital infrastructure? How seriously does it take artificial intelligence? What risks exist? And which bank would be the better choice for a specific customer?

This has not yet become an established metric in the banking sector. There is no generally recognized "AI Reputation Score" by which institutions can be compared. Nor can it be seriously claimed that ChatGPT, Gemini, Claude, or other systems have already replaced traditional sources of information for banking clients. The empirical foundation for such a claim is lacking.

However, the technical development that makes these questions relevant in the first place has long been a reality. Banks are incorporating generative AI into their own processes and customer interfaces, structuring knowledge for machines, migrating massive datasets, and developing systems that analyze documents, answer queries, or support employees in decision-making. At the same time, generative AI systems can consolidate information from disparate sources and generate answers from them. It is precisely at this intersection that a new form of digital corporate perception begins. EYECONIFAY refers to this emerging field of analysis as AI Perception Management.

The term is not an established regulatory or academic standard. It describes a perspective developed by EYECONIFAY on a trend that is becoming increasingly relevant for enterprises: It is no longer just humans, search engines, and media outlets that form an opinion about a company. AI systems also process information about companies and, in turn, generate answers for human users.

This may sound abstract at first. For banks, however, it is surprisingly concrete.

The Bank as a Dataset

A major bank no longer consists merely of branches, employees, accounts, and loans. It is also a massive information system. Annual reports, product descriptions, press releases, regulatory disclosures, investor relations documents, websites, historical information, and media coverage collectively form a digital image of the enterprise.

However, this image is not static. It changes every day.

A board member is replaced. A product is discontinued. A business unit is sold. Two companies merge. A subsidiary gets a new name. A bank acquires a competitor. A press release outlines a strategy that is revised just a few years later.

For a human, this is often a solvable problem. Anyone seriously examining a company can compare different sources, note publication dates, and contextualize contradictions.

For automated information systems, the situation is more complex. Generative AI is not a traditional database that automatically finds the single correct record for every query. Depending on the system, models may rely on trained knowledge, search functionality, external data sources, internal knowledge bases, or retrieval systems. A key technical method is Retrieval-Augmented Generation, or RAG. In RAG, relevant information is retrieved from a knowledge source in response to a query and provided to the generative model as context to formulate an answer.

The U.S. National Institute of Standards and Technology (NIST) now explicitly addresses the evaluation of such systems. Within its TREC program, NIST examines retrieval mechanisms, source attribution, and the quality of generated answers, among other factors. At the same time, NIST emphasizes that hallucinations - meaning invented or insufficiently supported statements - remain a relevant issue for large language models.

This brings to light a fact that was previously less critical for corporate communications: It is no longer necessarily enough for correct information to exist somewhere on the internet. What can be decisive is whether an information system finds it, categorizes it correctly, and appropriately considers it in its response.

This is not a guarantee that a well-maintained corporate website will automatically lead to an accurate AI response. Such a claim would be unprovable. But it is a compelling reason to examine a company's digital information architecture more closely.

Erste Group: 200 Years of Financial Knowledge for a Machine

An early example of this can be found at Erste Group.

As early as October 2023, Erste Bank introduced a generative AI prototype designed to make the bank's financial expertise accessible via a chat interface. The bank referred to the project as Austria's first financial AI. According to the institution, it utilized technology from OpenAI and ChatGPT to answer questions about finance and money in natural language. The answers were intended to be grounded in Erste Bank's more than 200 years of financial expertise.

What was interesting about this was less the chatbot itself than the underlying question: How can a company's internal knowledge be structured and made usable for a generative system?

Erste Bank explained at the time that content from its websites, publications, and educational materials had been transferred into a digital knowledge base. According to the bank, this database contained over 3,400 entries. During the testing phase, Erste reported conducting more than 3,300 chats with financial experts and collecting over 1,700 pieces of feedback. The bank also noted that 457 knowledge gaps were identified and closed during this phase. Simultaneously, it explicitly pointed out that generative AI could generate incorrect answers and that hallucinations could not be ruled out.

This highlights a point that is crucial for the further evolution of AI Perception: The bank was not merely attempting to use a generic AI. It was attempting to link its proprietary knowledge directly with generative AI.

Since then, development has continued. In its strategy for 2025, Erste Group identifies modern technologies, particularly artificial intelligence, as central to its goal of making financial advice more broadly accessible. In its 2025 Annual Report, the group describes further advancements in digital banking and within the George ecosystem.

However, this does not mean that Erste possesses a fully controlled "AI Reputation" today. Such a conclusion cannot be drawn from publicly available information. Nor is there a reliable public basis for a statement such as "Erste achieves an AI accuracy rate of 95 percent" across its entire corporate digital footprint.

What can be proven, however, is something else: Erste has been working for several years to combine its financial knowledge with artificial intelligence and increasingly integrate AI into its digital banking offerings.

For EYECONIFAY, this is precisely where the interest lies. The more corporate knowledge is processed by machines, the more critical the question becomes: What knowledge is actually available to a machine, and how is that knowledge structured?

UBS: The Past Must Fit into the New Digital Bank

At UBS, the challenge presents itself differently. There, it is not only about building new AI systems, but simultaneously managing the integration of one of the largest banking mergers of recent decades.

UBS acquired Credit Suisse in 2023. This required integrating not only customers, employees, and business units, but also systematically merging technical systems, applications, accounts, and infrastructure.

In March 2026, UBS reported a major milestone: Following the completion of the migration of Swiss-booked clients, all former Credit Suisse clients worldwide had been successfully migrated to the UBS infrastructure. According to the bank, approximately 1.2 million clients globally were transferred. For the Swiss migration alone, UBS stated that more than 80,000 tests and over 132,000 hours of migration-related staff training were required. Daily payment volumes on the UBS platform increased by 25 percent to nearly 3.1 million transactions following the migration.

While client migration is largely complete, the integration of the technical infrastructure is not yet fully finished. In March 2026, UBS stated that the final phase includes the decommissioning of legacy IT infrastructure, with substantial completion of the integration planned by the end of 2026.

For the question of AI Perception, this development is particularly intriguing because a merger does not just combine two organizations; it brings together two distinct informational histories.

Credit Suisse possesses decades of corporate history. UBS possesses its own. Legacy annual reports, press coverage, product documentations, and corporate disclosures do not automatically disappear from the digital space simply because a company is acquired.

This does not explicitly mean that UBS currently suffers from a problematic AI perception. There is no empirical study to support that, and such an assertion would be unjustified.

Rather, it illustrates a structural phenomenon that EYECONIFAY defines as Digital Cannibalism: A vast volume of historical and contemporary information fragments can make it fundamentally harder for algorithmic systems to map a company's current identity unambiguously.

This term, too, is a proprietary EYECONIFAY concept rather than an established academic term.

The question is therefore not whether old information on the internet is "bad." Rather, it is: In the context of a complex corporate history, how can an organization ensure that current and historical information are clearly differentiated by automated systems?

Concurrently, UBS Elevates AI to a Group-Wide Priority

While driving the Credit Suisse integration forward, UBS is massively scaling its artificial intelligence capabilities.

In the fourth quarter of 2025, UBS reported having more than 380 live AI use cases deployed across the group, with approximately 780 additional AI applications in development. Furthermore, UBS was executing nine major group-wide AI transformation initiatives.

These figures demonstrate one thing above all: At UBS, artificial intelligence is no longer confined to an isolated innovation lab.

AI is being embedded deeply across various operational areas of the bank. UBS describes AI applications serving clients, employees, and internal processes. In its annual reporting, the bank also highlights AI-powered chatbots assisting approximately 1.2 million mobile banking users.

The organizational significance of AI is equally evident. UBS established the position of Chief Artificial Intelligence Officer, appointing Daniele Magazzeni to the role in 2025.

Here again, however, it would be incorrect to infer the quality of the bank's public AI perception directly from the sheer number of its AI applications. Deploying 380 live applications does not automatically guarantee a better or worse reputation within an AI system.

What it does demonstrate is how deeply AI has penetrated the organization.

And with that, the role of data shifts.

When artificial intelligence is deployed across an increasing number of business processes, issues of data quality, sourcing, recency, governance, and traceability become paramount. This is not a claim unique to UBS, but a general consequence of the expanding adoption of such systems.

Deutsche Bank: When the Machine Processes Documents Instead of Just Answering

Deutsche Bank pursues yet another focus. It utilizes artificial intelligence not merely for interaction, but increasingly for processing complex information.

An example is dbTextract. According to Deutsche Bank, dbTextract is a generative AI solution that extracts and classifies information from complex documents such as forms, emails, and handwritten notes. The bank reports that the system is now deployed across multiple departments, achieves an accuracy rate of 97 percent, and reduces average processing time by approximately 40 percent. The underlying shared technology platform is capable of processing up to 4,000 pages per minute.

These metrics are impressive, but they require precise interpretation.

The 97 percent figure is not a general accuracy rating for Deutsche Bank's overall AI operations; it refers specifically to the dbTextract application as described by the bank.

The same applies to the 40 percent time savings - it is a metric specific to that application and cannot be generalized across all AI systems within the group.

This distinction is crucial when enterprises communicate their AI activities. A single successful use case is not the same as an aggregate key performance indicator for a corporation's entire AI architecture.

Deutsche Bank also works with AI assistants such as dbLumina and deploys generative AI in customer service. On its technology platform, the bank outlines AI applications spanning document processing, knowledge work, and client interaction.

Particularly noteworthy is the aspect of human oversight. In an AI use case described by the bank, the system generates draft proposals that are subsequently reviewed, modified, or rejected by trained staff. The final decision remains with the human operator.

This provides an important indicator of how major financial institutions handle one of the core challenges of generative AI: While machines can accelerate information processing, it does not follow that machines should make the final call.

The Real Shift Occurs Between the Systems

At first glance, the three banks appear to share little in common.

Erste Group uses artificial intelligence primarily to make financial expertise and advisory services digitally accessible. UBS combines AI with an immense technological integration mandate while simultaneously scaling hundreds of AI use cases. Deutsche Bank deploys generative AI for document processing, workforce augmentation, and customer service.

Yet from the perspective of EYECONIFAY, a common evolution unites them:

Corporate knowledge is becoming machine-readable and machine-processable.

This is fundamentally different from traditional digitization.

A PDF of an annual report is digital. A structured knowledge base is also digital. But an AI system processes the two in vastly different ways.

The decisive question is therefore no longer simply whether information is available online.

The question is: Is it current? Is it unambiguous? Is it consistent? Is the source verifiable? Can a system accurately attribute the information to the correct entity? And can the generated answer be audited?

This is precisely where the concept of AI Perception Management intervenes.

From SEO to a New Information Logic

The corporate world has encountered this trajectory before in a different form.

Search Engine Optimization (SEO) emerged around the question of how websites become visible in search engines. With Answer Engine Optimization (AEO), the focus shifted: How is information ingested and utilized in direct answers?

Generative AI takes this a step further. The user no longer needs to open ten search results. They can ask a question and receive a single, synthesized response.

This fundamentally alters the role of the original information source.

The source does not disappear. It simply becomes less visible.

For instance, a user might ask an AI about a bank and receive an answer synthesized from multiple disparate sources. The user no longer necessarily sees the underlying research process that produced that response.

This represents a structural shift in information distribution.

It does not mean that AI automatically becomes the decision-maker. Nor is there currently any reliable empirical basis to claim that customers select banks primarily based on AI responses.

However, a new touchpoint is emerging between companies and their prospective clients. And this touchpoint is mediated by machines.

Invisible Churn Is Currently a Hypothesis - but a Compelling One

This is where another EYECONIFAY concept applies: Invisible Churn.

The term describes the scenario where a company is eliminated from a prospect's evaluation set during the early information-gathering phase - without ever generating a traditional lead.

In traditional sales, the process follows a familiar path: A potential customer researches, visits a website, makes contact, becomes a lead, and receives a proposal.

In an increasingly AI-mediated discovery process, the workflow changes: A user poses a query, receives a synthesized answer, and may decide on the spot not to evaluate a particular provider any further.

When this occurs, no lost lead appears in the CRM. In fact, no lead ever existed.

This is a compelling strategic hypothesis. However, it would be scientifically unsound to treat it as an established economic fact today.

EYECONIFAY therefore does not claim that banks are losing a specific percentage of revenue to Invisible Churn today. Such a statement would require data that is currently not publicly available.

This distinguishes the concept intentionally from many marketing claims surrounding AI. Invisible Churn remains a hypothesis that requires further empirical testing.

What Companies Can Actually Control

This leads to the practical question: What can a bank or any enterprise actually do?

The answer is less sensational than what some "AI Reputation" vendors promise.

A company cannot control how a third-party AI model responds. Nor can it guarantee that every AI system will deliver the exact same response every time. And it cannot simply "write" a desired narrative directly into an LLM.

However, an enterprise can influence the quality of its underlying data layer.

This includes providing clear and up-to-date corporate information, maintaining consistent entity descriptions, using unambiguous corporate and brand names, structuring documents cleanly, and offering traceable primary sources.

This may sound like conventional corporate communications. The difference lies in the perspective.

Historically, the core question was: Does the human understand our message?

In the age of generative AI, a second question becomes imperative: Can an automated information system accurately interpret and reproduce our corporate identity?

This is the core of AI Perception Management.

A New Audit Framework for Enterprises

From this perspective, EYECONIFAY proposes an internal audit model. This is explicitly an EYECONIFAY methodology and not an established industry standard.

It centers around seven core questions:

1. First: Are the primary information sources authoritative and unambiguous?
2. Second: Is core corporate data consistent across different external sources?
3. Third: Is the information up to date?
4. Fourth: Can companies, brands, subsidiaries, products, and executives be clearly differentiated from one another?
5. Fifth: Is relevant data technically structured so that digital systems can ingest and process it effectively?
6. Sixth: What do various AI systems actually output when presented with standardized prompts?
7. Seventh: Which of these outputs are commercially material to the business?

This final point is critical. Not every AI hallucination or error constitutes a business risk.

If a system misspells the name of a minor product, that is fundamentally different from generating false statements regarding the executive board, ownership structure, capital adequacy, or regulatory compliance.

AI Perception must therefore always be aligned with the economic materiality of the information.

No AI Reputation Standard Exists Yet

It is precisely here that the industry must exercise caution.

It would be tempting to assign companies a score between zero and one hundred today, claiming that one bank has an "AI Reputation of 94" while another scores only 72.

However, without a transparent, reproducible, and standardized measurement methodology, such a number is merely marketing hype.

A robust AI Perception Benchmark would first need to establish a verified baseline of reference facts. Standardized prompts would then need to be submitted across multiple AI models, with their outputs evaluated against that ground truth.

Such an evaluation would need to measure not only factual accuracy, but also recency, completeness, source attribution, and entity mapping.

Only after such a methodology has been repeatedly applied and validated could companies be meaningfully benchmarked against one another.

This is precisely why EYECONIFAY refrains from publishing fabricated AI Reputation Scores for Erste Group, UBS, or Deutsche Bank in this article. Publicly available data is simply insufficient for that purpose.

This is not a weakness of the analysis. It is a necessary boundary of rigorous corporate research.

The Bank of the Future May Have Two Digital Identities

The first identity is the one seen by humans.

The second identity is the one generated by machines from accessible data.

These two identities can be aligned. But they do not have to be.

A company may position itself with extreme clarity in human-facing channels while its digital footprint remains fragmented across the broader web ecosystem. Conversely, a firm might possess technically structured data without having built a compelling human brand.

AI Perception Management does not seek to replace traditional reputation management. It complements it.

Public relations remains important. SEO remains important. AEO is becoming more important. Data governance remains important. Corporate communications remains important.

Siloed barriers are dissolving, and a new interface is emerging between these disciplines. It addresses the fundamental question of how machine-generated perception is constructed from corporate data.

The New Question for the Boardroom

In a few years, a question may become as routine in bank boardrooms as discussions surrounding brand awareness, cybersecurity, or customer retention:

What do the leading AI systems see when they analyze our enterprise?

Erste Group, UBS, and Deutsche Bank are each providing different preliminary answers—not through a shared industry standard, but through their operational trajectories.

Erste has demonstrated how proprietary financial expertise can be paired with generative AI. UBS is managing a monumental post-merger technical integration while deploying a group-wide AI footprint. Deutsche Bank is embedding generative AI directly into operational workflows like document processing and customer service.

None of these developments prove that AI currently controls the public reputation of these banks.

However, they demonstrate that artificial intelligence is penetrating deeper into the informational and operational fabric of major financial institutions.

And for that reason, the very definition of corporate reputation is evolving.

Moving forward, the critical question may no longer be solely: What do our customers think about us?

It will increasingly be: What information about us is being processed by machines—and what narrative about our enterprise is being synthesized as a result?

This is not yet an established market standard. But it is a question that will inevitably grow in significance as generative AI becomes ubiquitous.

EYECONIFAY defines the strategic framework to address this question as AI Perception Management.

Not with the illusion of being able to control artificial intelligence. But with a far more grounded objective: To render an enterprise's digital identity as current, unambiguous, consistent, and verifiable as possible - before machines tell a story about it.

Sources

• [1] National Institute of Standards and Technology (NIST), "Retrieval-Augmented Generation (RAG)", Glossary. NIST describes RAG as the integration of a generative AI system with a separate information retrieval system or knowledge base.
• [2] NIST / TREC, "Retrieval-Augmented Generation Track". Research focused on retrieval mechanisms, evidence validation, source attribution, and generated response quality.
• [3] Erste Bank, "Erste Bank launches Austria's first financial AI", October 5, 2023. Press release documenting the financial AI prototype, deployment of OpenAI/ChatGPT, proprietary knowledge base, and testing phase metrics.
• [4] Erste Group, Annual Report 2025 – Strategy. Strategy documentation emphasizing modern technologies and AI as core to expanding financial advisory accessibility.
• [5] UBS, "UBS successfully completes the client migration in Switzerland – another key milestone in the integration journey", March 18, 2026. Press release detailing the global client migration from Credit Suisse to UBS infrastructure, encompassing 1.2 million clients and 80,000 system tests.
• [6] UBS, Fourth Quarter 2025 / Annual Reporting. Reporting documenting over 380 live AI use cases, 780 applications in development, and nine corporate transformation initiatives.
• [7] Deutsche Bank, "Technology Transformation". Corporate disclosure detailing dbTextract, dbLumina, and AI deployment metrics (97% accuracy, 40% processing time reduction).

Methodological Note

This article explicitly differentiates between verifiable facts, logical deductions derived from those facts, and proprietary EYECONIFAY concepts. "AI Perception Management," "Digital Cannibalism," and "Invisible Churn" are utilized as defined within the EYECONIFAY intellectual framework. Assertions regarding Erste Group, UBS, and Deutsche Bank are grounded in the cited public records. Where reliable public data is unavailable to support quantitative benchmarks, no values are asserted.