The Great Refactor – rebuilding banking systems for the age of AI agents

Financial institutions person deployed instrumentality learning and automation for years. More recently, generative AI has accelerated that shift, with banks deploying copilots, coding assistants and customer-facing tools crossed the organisation.
Despite this, astir banking systems are fundamentally designed astir quality users: retail customers, traders, narration managers oregon portfolio managers, and the engineering teams liable for the software. The basal presumption remains that a idiosyncratic volition log in, navigate a screen, construe the data, travel a process and determine what happens next.
Agentic AI is starting to situation this. Unlike generative tools that chiefly nutrient answers, agentic AI systems tin plan, telephone tools, execute tasks and enactment towards a defined goal.

The caller plan question

This displacement is starting to scope the manufacture and is poised to accelerate rapidly. According to the Cambridge Centre for Alternative Finance’s 2026 Global AI successful Financial Services report, 52% of firms are already piloting oregon deploying agentic AI, portion 81% of respondents expect autonomous agents to beryllium meaningfully achieved by 2030.

Goldman Sachs is moving with Anthropic to make AI agents to velocity up commercialized and transaction accounting, lawsuit owed diligence and onboarding. These are halfway areas wherever banks beryllium connected data, controls, judgement and execution moving together.

The adjacent procreation of fiscal systems volition beryllium driven not conscionable by radical clicking done static interfaces, but besides by AI agents acting connected behalf of customers, employees, traders, hazard and operations teams. In that world, earthy connection becomes much than a chatbot prompt. It becomes a programming syntax: a mode of instructing systems, triggering workflows and coordinating enactment crossed the enterprise.

For banks, this changes the plan question. The contented is nary longer entree to the AI models, but whether their information and IT systems tin beryllium discovered, understood, operated and audited. Systems that cannot beryllium utilized by AI agents hazard becoming functionally unavailable to the adjacent procreation of integer finance. This is the lawsuit for a “Great Refactor”: the contention to crook bequest bundle into platforms that AI agents tin really use.

The hidden constraints

Large fiscal institutions often tally connected bequest exertion estates built connected decades-old code, manual processes and quality interpretation. Critical concern logic whitethorn beryllium buried successful undocumented rules, manual workarounds, fragmented information stores and systems. AWS has recovered that successful galore cases, implicit 40% of concern rules are embedded solely successful codification with nary supporting documentation.