Banks are investing heavily in AI in the hope that it can improve productivity and reshape how work gets done. Yet the results have been slower to materialise. Only 32% of senior executives can point to tangible gains in business productivity, while 41% do not expect AI to move the needle for years. In financial services specifically, 52% of firms have adopted agentic AI, but just 14% consider it central to their organisational strategy and competitive advantage.
Part of the challenge lies in the complexity embedded in banks’ daily operations. Employees often rely on experience to navigate disconnected systems, handle exceptions and make decisions that formal processes do not fully account for. This accumulated ‘people debt’ leaves valuable operational knowledge held by individuals rather than captured in the organisation’s systems and processes. If AI is introduced without first uncovering that context, banks risk carrying existing inefficiencies and workarounds into a new generation of technology.
Why undocumented knowledge limits AI agents
Financial institutions run on formal policies, but they also rely on thousands of informal decisions made by experienced employees. These workers know how to move work between systems that were built at different times, hold different information and often do not communicate with each other. They know when an exception needs escalating, when a client request requires human review and when a workaround is masking a deeper process issue. Much of this knowledge is not captured in systems or standard operating procedures.
This reliance on undocumented judgement becomes a problem when AI agents are expected to take on work that has traditionally relied on human judgement. To an AI agent following the documented process, a pause or extra step may appear unnecessary. In practice, that intervention may prevent an incorrect decision from affecting a client, creating a regulatory issue or exposing the bank to financial loss. Financial services firms therefore need to understand where human judgement protects the business and where employees are compensating for outdated or poorly designed processes.
For example, when a customer makes an unusual payment, an experienced employee may draw on the customer’s history, the circumstances surrounding the transaction and previous cases to decide whether it is likely to be routine or needs further investigation. An AI agent may understand what step usually comes next, but it will not necessarily understand the context behind that decision unless the relevant knowledge has been captured. Without this understanding, AI risks reproducing existing weaknesses and workarounds, rather than removing them.
Building context before automating work
AI can help firms close the people debt gap, but only if they use it to understand their work before they automate it. That starts with a simple question applied to every manual judgement in a process: is this step protecting the business, or is it patching a broken process? The first kind of knowledge needs to be captured and built into how an agent works; the second needs to be fixed, not automated.
Answering the question means finding the decision points in the first place, and the challenge is that most of it is invisible and resides outside standard operating procedures. Emerging “work graph” platforms tackle this directly: rather than relying on documented processes, they passively map how work actually flows across communication channels, ticketing systems and enterprise records, then extract the unwritten knowledge employees apply along the way. That turns judgement which has lived in employees’ heads into a living context model the wider organisation and AI agents can build on.
For example, take customer onboarding. Experienced employees know how to check documents, validate information and resolve unusual cases. Some of these judgements may never have been written down because they were developed through years of handling similar cases. Mapping the cases they escalated, the ones they cleared, and the reasoning between the two makes that judgement visible. The bank must still define when the AI agent can proceed, when it should ask for review and when a human needs to make the final decision.
The same exercise separates necessary safeguards from compensating workarounds. An extra check that catches genuine risk should be built into how the agent works; one that exists only because two systems don’t talk to each other should be designed out before anything is automated. By building a clearer picture of how work actually gets done, firms give AI the context to handle tasks reliably, while keeping people involved where experience and judgement remain necessary.
Setting the boundaries for AI agents
Incorrect or fabricated outputs can have serious consequences, especially in a regulated environment. Because agents can access information and take action across systems, implementing reliable data and effective guardrails before they begin making decisions is essential. This means institutions need to know where the data comes from, who is responsible for it and what information AI is permitted to access.
To achieve this, banks should route AI agents through a central point of control that records how they are used, rather than allowing them to operate freely across the organisation. This provides a consistent view across the enterprise and makes it easier to investigate problems when they arise.
The employees who understand how work actually gets done also have an important role to play in setting these controls. Their expertise can help define the boundaries of an agent’s role, identify when human judgement is still required, and ensure AI is used to rethink and improve how problems are solved, rather than simply being bolted onto existing processes.
Creating the foundations for AI-enabled banking
Banks will not get the most from AI simply by deploying agents across existing processes. They first need to understand the knowledge, decisions and workarounds those processes depend on, and decide which should be preserved or which should be removed.
Building the right foundation is the key to success. Governance should define where agents can act, when human involvement is required and who remains accountable for the outcome. The supporting infrastructure must also give agents access to the context they need to operate reliably. Just as importantly, employees need to be part of shaping how AI is introduced, bringing the experience required to redesign processes rather than simply replicate them.
Banks that take this approach can use AI to simplify how they operate, strengthen institutional knowledge and build greater resilience for the future.
Meena Athinathan
Meena Athinathan is a Vice President and Strategic Business Unit Leader of Banking at Cognizant. An AI builder and technology services provider, Meena drives strategic vision, fosters high-level CXO engagements, and builds strong relationships with analysts, advisors, and industry partners. She drives the adoption of AI-led operating models and solutions for banks, enhancing operational efficiency, customer experience, and data-driven decision-making to deliver transformative business outcomes.
