Across financial institutions, artificial intelligence is moving from experimentation into operational use. The point of differentiation is increasingly shaped by how these capabilities are integrated into workflows, data environments, and decision-making processes. As adoption becomes more widespread, organizations that delay this integration are likely to experience differences in speed, cost structure, and analytical depth.
Much of the attention around AI continues to focus on visible applications such as chat interfaces, content generation, and productivity tools. These signals are relevant, although they do not fully capture the structural changes taking place. In this article, we explore how the expansion of AI capabilities is translating into new operating models in financial institutions, with a particular focus on balance sheet management and Asset and Liability Management (ALM).
Audio Article
This transition is not driven by a single breakthrough, but by the accumulation of incremental improvements that compound over time. As capability increases, the set of tasks that can be performed without direct human intervention continues to expand. This creates a gradual but persistent shift in how work is structured and executed across functions.
A more accurate perspective comes from observing how the range and complexity of tasks AI systems can perform autonomously continue to expand.
Recent measurements of AI capabilities show a clear progression. Tasks that previously required seconds or minutes of human effort now extend to work requiring hours of expert time. This progression follows an exponential pattern, with the time horizon of autonomous task execution increasing rapidly over relatively short periods.
This pattern is not always immediately visible in day-to-day observation. AI capability tends to advance through successive stages, with periods of incremental improvement followed by more significant increases in performance. Short-term observations can therefore underestimate the broader trajectory.
Empirical evidence supports this interpretation. Surveys of AI researchers have shown consistent underestimation of future capability. Milestones expected decades ahead have begun to materialize much earlier. The effect of compounding improvements contributes to this gap between expectation and reality.
In practical terms, the evolution of AI capabilities can be understood through several dimensions:
The duration of tasks that AI systems can complete without human intervention
The complexity of reasoning required to complete those tasks
The ability to maintain context across multiple steps and data sources
The integration of external tools and systems into the execution process
These dimensions progress together, expanding the range of tasks that can be performed autonomously. The result is a gradual extension of AI from isolated assistance toward more complete task execution.
For financial institutions, the relevance of these developments depends on how AI is incorporated into operational workflows. The distinction lies in whether AI is used as an isolated tool or integrated into systems, data, and processes that define the organization’s environment. In practice, adoption tends to follow a progression.
Employees use AI systems to support drafting, summarization, and research tasks. Productivity gains are often immediate, although they remain at the individual level. The AI operates without access to institutional context, and each interaction requires manual input of relevant information.
AI systems are extended with access to institutional knowledge and operational tools. Informational agents provide responses grounded in internal frameworks, documentation, and historical decisions. Operational agents interact with systems to retrieve data, perform calculations, and generate outputs. At this stage, AI becomes part of the analytical infrastructure, combining context with action.
AI systems execute workflows based on triggers such as data updates, schedules, or predefined conditions. Deterministic workflows handle structured processes, while coordinated agent systems manage more dynamic tasks. Outputs are produced continuously, with reduced reliance on manual initiation and greater integration across systems.
This progression reflects an expansion in both capability and integration. Each phase builds on the previous one, extending the range of tasks that can be performed and the degree to which AI is embedded within the organization’s operating model.
Software development provides an early example of this progression. The domain benefits from structured inputs, clear validation through testing, and sustained investment in tooling. These characteristics have enabled earlier adoption of AI capabilities. Development practices have evolved from individual assistance to more structured approaches in which specifications guide implementation by AI systems.
In this environment, the role of the developer has evolved alongside the tools. Greater emphasis is placed on defining system behavior, architecture, and requirements with precision. Implementation becomes increasingly supported by AI systems capable of translating these specifications into working code.
The underlying capability extends beyond software engineering. The ability to translate a well-defined specification into a working implementation is becoming more accessible across functions, and other domains are beginning to follow a similar trajectory.
This is starting to change how organizations approach problem-solving. Traditional models rely on centralized development resources, where requests are prioritized and implemented over time. With these AI capabilities becoming more widely available, teams can define requirements with greater precision and obtain working solutions more directly. The primary constraint becomes the clarity and completeness of the specification.
In balance sheet management and Asset and Liability Management (ALM), these changes have direct implications. These functions depend on continuous analysis, regulatory interpretation, and the integration of multiple data sources. As AI systems gain access to institutional data and regulatory frameworks, they support the analytical processes required for reporting, risk management, and decision-making.
The types of activities affected include:
Analysis of balance sheet dynamics across currencies, products, and business lines
Interpretation of regulatory metrics such as IRRBB, LCR, and NSFR
Identification of drivers behind changes in financial indicators
Monitoring of data quality and consistency across reporting processes
Preparation of reporting materials and analytical summaries
These activities require both data access and domain understanding. As AI systems combine these elements, they contribute to the production of analysis in a more continuous and scalable manner.
Human judgment remains central within this model, supported by AI in reducing the effort required to produce analysis and increasing the capacity to interpret and act on it. The allocation of work changes, with operational tasks increasingly handled by systems and interpretative tasks remaining with professionals.
Platforms such as Mirai AI illustrate how this architecture is being implemented in practice. By combining domain-specific expertise, access to institutional data, and agent-based infrastructure, these systems extend beyond general-purpose AI tools and align with the requirements of balance sheet management.
This architecture brings together several components:
Domain knowledge that reflects regulatory frameworks and institutional practice
Direct access to balance sheet data and internal systems
Agents capable of performing analytical and operational tasks
Workflow structures that enable continuous execution rather than on-demand interaction
In this model, agents operate with both context and action. Domain knowledge provides the interpretative layer, while system connectivity enables interaction with data, models, and reporting processes. Mirai AI Agent reflects this approach by combining regulatory and domain expertise with live access to balance sheet data, allowing analysis to be generated in a way that is specific to the institution’s position.
The integration layer is equally important. The Mirai MCP Server enables these capabilities to be accessed as part of a broader agent ecosystem, allowing institutions to connect their own systems, workflows, and agents without requiring bespoke integrations for each use case. This supports the transition toward coordinated agent environments, where multiple systems interact through a shared infrastructure.
The development layer further extends this model. Mirai AI Modeling provides an environment where analytical models can be built, calibrated, and iterated using direct access to structured data and AI capabilities. This reduces the distance between conceptual modeling and implementation, allowing teams to define requirements and obtain working outputs more directly.
The adoption of these capabilities is already visible across the industry. Public disclosures and industry research indicate:
AI is contributing to a significant share of code generation in large organizations
Reported levels above 25% in some institutions, with higher concentrations in specific use cases
Substantially higher percentages among newer technology-driven firms
A sharp increase in interest in multi-agent systems, reflecting a shift toward coordinated architectures
These signals align with the broader transition from individual productivity to integrated capability and automation.
The broader pattern reflects a transition from isolated tool usage to embedded infrastructure. As these systems become more integrated into workflows, data environments, and analytical processes, differences in adoption approaches are likely to influence operational performance.
AI is becoming part of the operational architecture of balance sheet management, and its integration into workflows, systems, and analytical processes is already underway.
Download the whitepaper and discover how AI agents are transforming balance sheet management in practice.