The AI debate has shifted
Today, bank executives hardly need to justify whether and why artificial intelligence is relevant to their institution. In September 2025, the European Banking Authority reported that 92 per cent of EU banks were already using AI; in 2026, the European Central Bank noted that more than 85 per cent of the banks it supervises use AI. The strategic gap therefore lies less in adoption than in the depth of utilisation and the ability to realise benefits in the long term.1, 2
This is precisely where the central dilemma lies: banks now have a growing list of AI pilot projects, chatbots and assistance solutions. At the same time, process turnaround times, capacity bottlenecks and the cost-to-income ratio often remain virtually unchanged. The cause is rarely a lack of ideas. What is missing is an operating model that brings together business value objectives, data, technology, governance, change management and accountability for results. The next phase of AI transformation will therefore be determined by economic value transfer. Time savings must be translated into freed-up capacity, better decisions, higher quality or additional business. A demo does not yet constitute value contribution.
The way out of this dilemma begins with a shift in perspective: AI is managed as an enterprise-wide capability. Individual applications remain important, but are embedded within value streams, reusable components and a shared product and operating model. This also changes the investment logic. Management does not fund a collection of isolated projects, but rather a value creation architecture.
The use case trap: Why activity alone does not generate added value
The implementation of an AI roadmap often fails due to a paradoxical pattern: the greater the focus on AI, the more local initiatives spring up. Business units set out their own priorities, IT evaluates new tools in parallel, compliance assesses individual cases, and senior management expects quick results. Whilst this creates momentum, it does not lead to a shared direction.
One key problem is local optimisation. Applications improve individual tasks without changing the end-to-end process. The time saved is then wasted in handover processes, waiting times or media breaks. Added to this is a flawed comparison framework. A corporate GPT, automated document verification and an agent-based process are often assessed using the same ROI template, even though strategic enablement and directly measurable efficiency require different criteria.
Another common pattern is the technical AI pilot without any process-related decisions. The application works, but roles, control points, data access and follow-up steps remain unchanged. The pilot runs alongside the existing process and thus adds to the complexity. Equally problematic is adoption without accountability for results: usage figures rise, whilst no one is formally responsible for translating the time saved into increased capacity, improved service quality or sales success.
The white paper “Escaping the Use-Case Trap” describes this development across the Forming, Storming, Norming and Performing phases. In the early phase, the focus is initially on gaining experience and establishing a robust mandate. In the storming phase, ideas are channelled and prioritised. Only with an integrated operating model are the prerequisites for scaling and measurable benefits created.
Early AI pilots should therefore be deliberately understood as „experience cases“. They initially pursue learning objectives: What level of data quality is achievable? How are control procedures changing? Where is acceptance emerging? Which platform components are reusable? Success in this phase is measured in terms of insights and interoperability. As soon as the bank moves into the scaling phase, these become business cases with clear value and exit criteria.
From use case to value stream: prioritising business model logic over technology
Tangible added value is created when the bank takes its strategic value pools as its starting point. The key question is therefore: which performance or outcome metric should show a visible change over the next twelve to 24 months? Only then is a decision made as to which AI capability is required.
In the Growth value pool, for example, the focus is on conversion rates, share of wallet and time-to-offer. Here, AI can support advisory preparation, offer drafting, next-best-action approaches or subsidy matching. The benefits are reflected in additional revenue and higher conversion rates.
In the Capacity value pool, processing time, volume per full-time equivalent and the degree of automation are the key metrics. Typical areas of application include the verification of KYC documents, preliminary credit checks, contract analyses and complaint handling. The economic benefit arises from hours freed up, avoided new hires or the ability to handle rising volumes without a proportional increase in staff numbers.
The Quality and Risk value pool addresses error rates, rework, findings and loss ratios. AI-based completeness checks, plausibility checks, quality gates and continuous monitoring can reduce error and risk costs.
The Resilience and Knowledge value pool focuses on service levels, peak load capacity, access to knowledge and onboarding time. Corporate GPTs, knowledge assistants, regulatory research and faster information retrieval enhance stability, speed and skills development.
The AI Value Sprint: From bottleneck to robust value proof in six weeks
An impact-driven approach begins with a clearly defined value stream and a measurable bottleneck. For example, in corporate banking, a ‘credit assistant’ is not developed in general terms. Instead, the process from document receipt to a proposal ready for decision-making is examined. Process data, interview results and spot checks reveal where search efforts, missing documents, repeated checks or manual data transfers occur.
The AI Value Sprint begins with a clear value objective, a results-based KPI and a clearly defined process scope.
This is followed by a Process Twin, which captures the actual current workflow, volumes, waiting times and control points.
The appropriate AI capability is then determined, such as extraction, research, evaluation, generation or orchestration. The minimum viable AI service is tested using real-world cases, defined quality thresholds and direct user feedback.
After six weeks, it must be possible to make a sound management decision: scale up, modify or terminate. From the outset, this decision must include consideration of how any potential increase in capacity is to be utilised economically. This creates speed without compromising on governance and cost-effectiveness.
Seven innovative levers for measurable AI benefits
The following framework combines short-term impact with long-term scalability. The approaches are interlinked and can be weighted differently depending on the institution’s level of maturity, business model and risk profile.
The first lever is a dual AI portfolio that manages horizontal enablement and vertical value contributions separately. Horizontal solutions, such as a corporate GPT, create reach, learning curves and reusable platform components. Vertical services optimise specific process steps and are managed using business case metrics. Both tracks require their own budgets, criteria and performance indicators. In this way, a strategic platform does not artificially compete with clearly quantifiable document automation.
The second lever is a ‘single point of contact for employees’. Employees should not have to navigate between numerous bots and specialised applications. A unified front-end interfaces in the background with authorised data sources, language models and specialist services. The user formulates their objective, whilst the platform orchestrates research, document comparison, summarisation or workflow steps. This reduces training requirements, boosts adoption and prevents the emergence of new siloed solutions.
The third lever lies in Agentic Operations. Agentic systems can schedule tasks, launch tools, review results and hand cases over to humans. For banks, this creates a new capacity class between traditional automation and human case handling. In future, the Executive Board will manage digital capacity alongside full-time equivalents and outsourcing. To this end, it must be determined what volumes the system is permitted to process, which decisions are excluded and which quality gates apply.
The fourth lever is an AI Value Ledger, which tracks the benefits from the AI pilot right through to the profit and loss account. A central value record documents, for each AI product, the baseline, the expected benefits, the actual usage, the impact on quality, the operating costs and the capacity decision taken. This makes it clear whether time savings have actually translated into shorter lead times, higher case volumes, avoided staff recruitment or improved customer interaction. The ledger links product management, financial controlling and portfolio management.
The fifth lever is the Process Twin, an alternative to the idealised process diagram. Traditional process models often fail to adequately capture exceptions, queries and individual working methods. A Process Twin combines process mining, case sampling and observations from day-to-day work. This reveals where knowledge searches, media breaks and control loops actually occur. AI is thus aligned with real-world friction losses rather than abstract target processes.
The sixth lever combines a model router with a modular platform. A bank-proprietary orchestration layer separates business services, models and infrastructure. Depending on security requirements, quality, latency and costs, the appropriate model or infrastructure provider can be selected. This architecture strengthens technological sovereignty, enables active price control and reduces dependence on individual providers. At the same time, governance checks and interfaces become reusable.
The seventh lever is ‘learning in the flow of work’. AI competence is developed primarily through application. Power users, context-sensitive assistance, short feedback loops and joint case reviews integrate learning into day-to-day work. Managers steer usage via specific team objectives, whilst technical expertise and critical judgement are deliberately strengthened. As a result, change evolves from a communication campaign into an operational management task.
The economic logic behind these levers can be expressed in a simple formula: the realised value is calculated as the time saved, multiplied by the utilisation rate, the appropriate case volume and the economically usable proportion of capacity. Added to this are quality and revenue effects; operating, transformation and risk costs must be deducted. This logic prevents minutes saved in theory from being prematurely recorded as a contribution to earnings.
The Operating Model: Integrating Value Responsibility, Governance and Technology
A scalable operating model requires three interlinked levels.
At the level of strategic steering, the Executive Board and the AI Review Board define the value pools, risk appetite, investment framework, as well as portfolio and stop decisions.
At the level of AI product ownership, business units, product owners and technical experts jointly manage the end-to-end process, benefit KPIs, adoption, technical quality and continuous development.
The third level consists of the AI platform and KIOps. IT, data management, information security and compliance are responsible for models, data access, service connectors, monitoring, approvals, operations and vendor changes.
The key lies in the interplay: the Executive Board defines the ambition and risk corridor, specialist product teams deliver the value contribution, and a central platform provides reusable technical and regulatory capabilities.
Governance by risk tracks: increasing speed where it is justifiable
Governance acts as a catalyst when recurring decisions are standardised. A risk-based triage assigns new initiatives to a processing track at an early stage.
- In the green track, internal support solutions involving public or non-critical data, subject to mandatory human review via standardised approvals, are implemented quickly.
- The yellow track comprises applications involving internal data, with relevant process implications or automated quality checks; these require in-depth data protection, model and control analyses.
- The red track covers applications with customer impact, high criticality or use in decision-making processes. These require individual assessment, strict human oversight and an explicit mandate from the board.
The regulatory framework remains part of this governance structure. The EU AI Act has been coming into force in stages since August 2024; transparency obligations for certain AI systems have been in effect since 2 August 2026. Banks therefore need to take an integrated view of the AI Act, data protection, DORA, outsourcing management, information security and existing model risk processes. A separate set of AI regulations that is not linked to existing control systems creates additional red tape.3
The roles that make the difference in scaling
The Business Sponsor embeds the relevant value pool within the business strategy and removes organisational barriers; their success is measured by the contribution to results achieved.
The AI Product Owner manages the process, backlog, user experience, KPIs and technical quality, and is therefore responsible for adoption and process impact.
The AI Review Board prioritises the portfolio, determines the risk profile and terminates underperforming initiatives. Its success is reflected in focus and the speed of decision-making.
The KIOps Engineer ensures operations, monitoring, model changes, cost control and technical resilience, and is therefore responsible for stability and scalability.
AI Power Users test real-world scenarios, provide continuous feedback and support colleagues in their day-to-day work. They drive adoption and the learning curve.
The 100-Day Agenda for the Executive Board
Integrating AI into the business model does not require a preparatory phase lasting several years. Within 100 days, management can lay a robust foundation, generate two to three proofs of value and prepare the ground for scaling decisions. This requires a clear definition of the scope.
In the first 30 days, three strategic value pools are defined and baselines for volume, time, quality and costs are established for the prioritised processes. At the same time, the AI Review Board and its decision-making powers are formalised, and risk trails and initial platform guidelines are defined.
Between day 31 and day 60, the bank carries out two AI Value Sprints in prioritised value streams. Existing corporate GPT or platform components are reused, real users and power users are involved, and benefits, quality, risks and operating costs are measured transparently.
From Day 61 to Day 100, the winning initiatives are scaled up and underperforming initiatives are systematically discontinued. The AI Value Ledger is integrated into controlling and portfolio management; capacity and process decisions are made; and a roadmap for the platform, roles and the next value streams is approved.
Five decisions that only the Executive Board can make
Firstly, the Executive Board must define the value focus and decide which two to three strategic value pools are to be prioritised and which initiatives are to be deliberately put on hold.
Secondly, it defines the risk corridor: which data, processes and levels of decision-making are permissible, and in which areas does human responsibility remain essential?
Thirdly, it decides on the capacity strategy and thus whether the time freed up is to be used for growth, quality improvement, reducing workload or achieving structural cost savings.
Fourthly, a decision on autonomy is required: which components must remain controllable and interchangeable, and where is it more cost-effective to procure standard services?
Fifthly, the Executive Board must adapt the management systems and determine which targets are to be incorporated into divisional and team management, so that joint responsibility is taken for adoption and the impact on results.
The Executive Board should deliberately avoid a technology-driven large-scale programme without prioritised value streams; a constantly growing list of AI pilot projects without decisions on when to halt them; a governance process that treats every case as a high-risk system; productivity promises without capacity decisions; and change communication without practical application in day-to-day work.
Making added value measurable: From usage rates to impact on results
The most significant change in management relates to the measurement logic. Many organisations measure logins, active users or responses generated. These metrics indicate acceptance, but do not yet reflect economic impact. A robust KPI system links the five levels:
- usage,
- process,
- quality,
- business and
- risk.
At the usage level, active users, suitable cases and repeat usage are considered. The key management question is whether the application is actually being used in the relevant work context.
At the process level, the focus is on turnaround time, touch time and the degree of automation. The key question here is whether the end-to-end process is changing or whether merely a single step in the workflow is being accelerated.
The quality level captures error rates, rework, completeness and consistency, and assesses whether performance is becoming more reliable and traceable.
At the business level, volumes per full-time equivalent, conversion rates, service levels and avoided hires are measured. The decisive factor is whether a visible contribution to results or customer value is generated.
The risk level examines model deviations, overrides, and data protection and security incidents. It answers the question of whether the benefits remain within the defined risk corridor.
The capacity bridge is particularly important. If an AI assistant reduces the time taken for a task by 30 minutes, this initially represents only technical potential. It is only when this is multiplied by suitable use cases, actual usage and economically poolable capacity that the realisable value becomes apparent. This transparency protects against unrealistic expectations whilst also highlighting which organisational decisions are required.
An AI Value Ledger should therefore be updated monthly and reviewed quarterly within the portfolio. Products with a growing impact are allocated a scaling budget. Products with low usage or no impact on processes are revised or discontinued. This consistency is a key distinction between a culture of experimentation and professional AI management.
Conclusion: AI becomes a lever for the business model when the bank industrialises its capabilities
Artificial intelligence has arrived in the banking sector. The competitive advantage now stems from the ability to generate repeatable value from the technology. To achieve this, banks must move beyond the logic of individual use cases and organise AI across their value streams, decision-making and operational processes.
The key management task comprises four elements: a clear focus on value, a dual portfolio of horizontal enablement and vertical services, a modular platform with risk-based governance, and a measurement framework that translates time savings into an impact on results.
Supplemented by learning within the workflow, this creates an operating model that combines innovation and control.
The next step is not simply to gather more ideas. It is an executive board decision on where AI should have a measurable impact within the business model and which structures will enable this impact in the long term.
Seven questions for the next executive board meeting
The next executive board meeting should focus on seven questions:
- Which three value pools should be visibly influenced by AI over the next twelve months?
- Which end-to-end value stream is suitable for the first robust AI Value Sprint?
- Which platform components can all business units reuse?
- How do we distinguish between investments in learning and AI products aimed at immediate returns?
- Who is responsible for realised benefits and the associated capacity decisions?
- Which risk pathways accelerate straightforward cases and safeguard critical applications?
- And which initiatives should we halt in order to gain focus and implementation momentum?
Recommended reading
