Moving beyond the AI Pilot Trap in Banking Operations

Generative AI is everywhere in banking – but it's rarely used effectively.

There's a big gap between pilot projects and real business operations. Three real-world examples show how to make the transition to industrial use.

Generative AI has sparked high expectations within the banking sector. Nearly every institution is experimenting with models and applications. However, only a few have managed to transition from proof of concept to productive deployment. Many initiatives remain stuck in the pilot stage or represent incremental developments of previously rigid chatbots or rarely used knowledge databases. 

Key reasons include poor data quality, insufficient risk controls, unclear business cases, and missing governance structures. This is confirmed by a recent study by PwC, in which 56% of more than 4,400 surveyed CEOs state that their AI investments have so far led neither to significant revenue growth nor to cost reductions. According to McKinsey, as many as 75% of companies achieve no or only marginal (below 5%) ROI from their AI initiatives.

At the same time, the technological baseline has changed. Traditional automation approaches have been largely exhausted in many areas. Where documents are unstructured, special cases occur, or contextual understanding is required, rule-based systems quickly reach their limits. 

Agentic automation enables a transition from purely learning systems that model human thinking and neural networks toward systems capable of replicating human decisions in a reproducible manner. As a result, platforms are emerging that not only process information but also consider context, prepare decisions, and orchestrate and execute complex business processes along defined control points. In doing so, they overcome the structural limitations of purely intelligent automation.

From AI Vision to Scalable Banking Operations

The central challenge, however, lies less in the technology itself than in its operational implementation. Individual financial institutions are often unable to operate agent-based automation at the same depth, with comparable cost structures, or with the same level of operational professionalism as specialized BPO providers. These providers industrialize such technologies across numerous processes, integrate them into modular platforms, and embed them within robust governance structures. Three practical examples from the BPO context illustrate how generative AI in the field of agentic automation is already being deployed productively and profitably today.

Case 1: AI Upgrade for Banking Automation Platforms

One approach is to start where established automation platforms are already in place but whose capacity for further innovation has reached its technical limits. 

In the first example from the insurance and healthcare sector, generative AI was deliberately integrated as an additional intelligent component into an existing on-premise platform.

Cases that were previously technically limited and too costly to automate – such as unstructured invoices, unclear insurance references, or individual special conditions – become manageable through trained language models. These systems learn to interpret documents contextually, extract relevant passages, and perform initial assessments.

To keep the process both lean and highly cost-efficient, only those data fields that the local system cannot clearly identify are transmitted to the cloud for AI processing. This targeted hybrid approach with sequential processing saves resources and significantly improves analytical precision. 

Impact

• Automation rates increased by up to 75 percent within three months, depending on document type. 
• Manual adjustments required up to 50 percent fewer full-time equivalents (FTEs). 
• Processing times and payout durations were reduced, while the quality of output for customers improved significantly. 
• At the same time, the platform core remained unchanged. Existing processes, security architectures, and regulatory frameworks could continue to be utilized.

Case 2: Rethinking Complex Business Processes Process End-to-End with Automation andGenerative AI

The full potential of generative AI becomes evident where existing systems are not merely enhanced but entire process chains are redesigned. 

The second example from the leasing and lending business illustrates this approach. 

The document-intensive verification and processing workflow was restructured as an end-to-end journey – from omnichannel data capture and digitalization to professional verification steps, customer communication, and the structured handover to risk management for the final credit decision. 

To enable this, SPS developed its own platform – SPS GPT – In collaboration with an IT partner. The platform is based on the latest large language model (LLM) technology and serves as a new tool for efficient data processing.

Generative AI does not function here as an isolated tool but as part of an orchestrated automation model. It interprets complete documentation, identifies inconsistencies at an early stage, checks dozens of business rules, and prepares or even conducts customer communication. Experts intervene only where human judgment is required keeping a human-in-the-loop approach.

It’s easy to see, from the firm’s perspective, why it’s worth the investment in a brand-new workplace that facilitates this sort of model. Removing the admin and traditional time sinks enables fee-earners to generate more value.

Impact

A decisive lever here lies in process design: validations were moved upstream, media disruptions were reduced, and responsibilities and escalation paths were clearly defined. 

• Processing times decreased by 70 percent, and the need for manual rework was reduced by more than 80 percent.
• Customer dialogue also improved – particularly because follow-up questions were formulated more clearly, contextually, and individually. 
• The core banking system remained untouched, which is a key factor in a highly regulated environment. 

This example demonstrates that the extensive use of agentic automation requires, in some cases, a fundamental rethinking of processes and the involvement of numerous internal functions. At the same time, it highlights the significant potential across entire process chains.

CASE 3. Agentic Automation: Sequential AI Agents in Action in Banking Processes

A preliminary stage of autonomous AI agents is agent-based automation. Specialized AI modules take on clearly defined tasks sequentially and in parallel along a process – such as analysis, communication preparation, or decision support. An example from motor vehicle claims processing illustrates this principle.

A first AI module consolidates extracted data from all claim-related documents – such as claim notifications, policies, police reports, supporting evidence, and repair cost estimates – verifies coverage, plausibility of the incident description, and cost consistency, and produces a structured professional claim assessment. 

A second module uses this assessment as context to generate status updates, follow-up questions, or decision rationales, adapting tone and content to the specific case. 

A third module creates an audit-proof decision dossier in which assessment, communication, and recommendations are logically linked and transfers it to a claims handler or downstream system.

Control points and human-in-the-loop mechanisms are firmly embedded. Depending on the task, agents act sequentially or in parallel, and quality assurance as well as governance are integral parts of the design. As a result, decision quality increased, processes became more standardized, and resources were deployed more efficiently.

Technology is only the smaller part of the equation

All three examples illustrate the potential of Agentic Automation. At the same time, they clearly show that technology alone does not determine success. Data, processes, and people are equally important. These four dimensions shape the approach SPS pursues in its implementations.

Technology: Technology is not viewed as a standalone application but as an integrated component of existing architecture. Decisions regarding platforms, cloud models, or language models are guided by cost-effectiveness, scalability, technical sovereignty, and compatibility with existing delivery models. Processes and roles are taken into account from the very beginning.

Data: From this perspective, the data foundation takes center stage even before technology itself. Experience from early projects has shown how quickly initiatives stall when data quality, governance, and responsibilities are not clearly defined. Clean data reduces hallucinations, prevents pilot projects from being discontinued, and accelerates scaling.

Processes: At the process level, generative AI is not deployed selectively but as part of a learning production system. Measurable targets, scalable use cases, and continuous improvement cycles ensure that short-term effects translate into sustainable productivity gains.

People: Organizational embedding ultimately determines success: new roles, skill profiles, human-in-the-loop mechanisms, governance structures, and early employee involvement. Transparency instead of “black box AI”, along with clearly demonstrated value for teams, supports sustainable organizational adoption.

Make or Buy: A Question about the operating model

For banks, the issue is no longer isolated AI experiments but the establishment of scalable operating models with clear impact on time, cost and quality. 

The make-or-buy question becomes a strategic decision point. Institutions must decide which capabilities to develop internally and where specialized partners can add the most value. 

Partners can reduce technological and operational complexity – for example, through proven modular platforms, compliance structures, and industrialized data processing. At the same time, responsibility for governance, quality, and regulatory requirements remains with the institution. 

The decisive factor is therefore not the technology decision alone, but its integration into the institution's own operating model.