The new design question
This shift is starting to reach the industry and is poised to accelerate rapidly. According to the Cambridge Centre for Alternative Finance’s 2026 Global AI in Financial Services report, 52% of firms are already piloting or deploying agentic AI, while 81% of respondents expect autonomous agents to be meaningfully achieved by 2030.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
Goldman Sachs is working with Anthropic to develop AI agents to speed up trade and transaction accounting, client due diligence and onboarding. These are core areas where banks depend on data, controls, judgement and execution working together.
The next generation of financial systems will be driven not just by people clicking through static interfaces, but also by AI agents acting on behalf of customers, employees, traders, risk and operations teams. In that world, natural language becomes more than a chatbot prompt. It becomes a programming syntax: a way of instructing systems, triggering workflows and coordinating activity across the enterprise.
For banks, this changes the design question. The issue is no longer access to the AI models, but whether their data and IT systems can be discovered, understood, operated and audited. Systems that cannot be used by AI agents risk becoming functionally unavailable to the next generation of digital finance. This is the case for a “Great Refactor”: the race to turn legacy software into platforms that AI agents can actually use.
The hidden constraints
Large financial institutions often run on legacy technology estates built on decades-old code, manual processes and human interpretation. Critical business logic may be buried in undocumented rules, manual workarounds, fragmented data stores and systems. AWS has found that in many cases, over 40% of business rules are embedded solely in code with no supporting documentation.
This creates a new form of agent-blind technical debt. AI agents cannot rely on institutional memory. They need systems they can access, interpret and operate. This is why the construction of a new financial operating system is necessary – from deterministic, hand-coded logic and human operation to model-driven, natural-language mediated and agentic.
There are signs that major banks recognise this. Citi is using AI to migrate data from legacy systems and automate coding, with CTO David Griffiths saying the agentic approach can be 2 to 20 times faster for certain tasks. Moreover, Morgan Stanley has used its DevGen.AI tool to review 9 million lines of legacy code and translate them into plain-English specifications. This is a machine legibility story: turning old logic into something that can be understood, explained and modernised.
Just as the 6Rs gave institutions a way to think about cloud migration, the age of AI agents requires a new set of design principles. Agents need access to systems through secure pathways. They need those systems to be annotated so that data, rules, and constraints are machine-readable. Capabilities need to be actionised, so agents can call tools and APIs rather than simply read information. Workflows need to augment human teams by assembling steps around intent. Business logic needs to be abstracted into policies and constraints that agents can reason over. And where appropriate, parts of the lifecycle can be autonomised with clear governance, testing, and oversight.
What agent-ready systems unlock
Done properly, this is transformative. Users no longer struggle through complex menus or fragmented workflows; instead, they describe the outcome they want, and the system can assemble the right steps around that intent. For customers, that could mean more personalised financial guidance, faster onboarding and more responsive support. For front-office and operations teams, it could mean faster reconciliation, better exception handling and fewer handoffs between systems.
It also changes how technology is delivered. Product teams can test workflows in natural language before committing a full engineering sprint. Developers can use architectural metadata to generate code that is better aligned with system constraints
The governance benefit is just as important. In traditional systems, an audit requires scrutinising syntax, logs, code, and process evidence. In a properly designed agentic system, audit can become more direct, query the system in plain English and operate around the clock.
The next competitive divide
The next banking transformation is not just about adopting AI. It is about preparing for a market in which more financial activity is initiated, interpreted and coordinated by software.
This is where large banks face both a disadvantage and an opportunity. Fintechs and newer firms may have cleaner stacks, fewer legacy constraints and more flexibility to design around agentic workflows from the start. Incumbent banks have scale, data, trust and deep client relationships, but also more complex estates to refactor. Cambridge’s research found fintechs ahead of traditional financial institutions in agentic AI adoption, at 57% versus 45%.
The most competitive institutions will not simply be those with the best apps or the largest AI teams. They will be those whose systems can be safely discovered, understood, operated and audited by agents. The next users of banking infrastructure may not be human. Banks that prepare for that shift will be better placed to move faster, serve clients more intelligently and preserve their relevance in the next generation of digital finance.
Kruttik Aggarwal, Director of Engineering at Lab49
