Agentic Banking Architecture

A Practitioner's Guide

Banking technology is approaching a critical inflection point. For sixty years, banks have built vertical applications - each containing its own business logic, data, and integrations. Lending systems, payments platforms, treasury applications, CRM — hundreds of silos connected by an ever-growing web of ETL, APIs, and reconciliation.

Agentic AI changes the architecture, not just the automation. When goal-oriented agents can reason, execute multi-step processes, and call tools, business logic no longer needs to live inside applications. It migrates into agents. Applications thin down to data access layers. Integration complexity — the dominant cost in banking technology — collapses.

This guide is a practitioner's reference for how that transition works. It covers the target architecture, the governance model, the migration path, and the organisational implications. It's written from the perspective of someone building and running AI platforms inside a regulated bank, not from a vendor or consulting vantage point.

It's also a living document. Sections are published as they're ready and updated as I learn more about what works.

• Section 1 The Thesis — Why Agentic Changes Everything The shift from applications-that-contain-logic to agents-that-execute-logic-on-shared-services. Why this is structural, not incremental. Sixty years of banking technology in context.
• Section 2 Reference Architecture A complete logical architecture for agentic banking: five horizontal layers, development agents, observability agents, and the common services that hold it together.
• Section 3 Governance Architecture How you govern agents in a regulated environment. Policy codification pipelines, the observation bus, three lines of defence, and the agent identity question.
• Section 4 Application Architecture — How Silos Dissolve What happens to existing applications when business logic migrates into agents. From vertical silos with duplicated data to horizontal processes on shared ledgers.
• Section 5 Migration Path — Four Phases How to get there from here. Wrap the legacy, migrate the logic, thin the applications, dissolve into ledgers. A phased approach with incremental value delivery.
• Section 6 The Deterministic-to-Probabilistic Spectrum A framework for classifying banking processes from fully deterministic to genuinely adaptive, and why it matters for governance and migration sequencing.
• Section 7 Data-Centric Process Architecture A framework for modelling banking processes around data states and transformations, facilitating transformation of processes to agentic execution.
• Section 8 The Delivery Organisation What happens to the org chart when AI handles coordination. Builders report to business domains, not technology. The CTO owns the platform, not delivery.
• Section 9 Accountability Distributed, not ambiguous. How banking already distributes accountability for automated systems — and why the same pattern applies to autonomous agents.
• Section 10 The People Transformation What changes inside the delivery cycle when agents handle coordination. Strategy, build, governance, knowledge — activity by activity. Roles eliminated, evolved, and emerging.
• Section 11 Risks and Opportunities Four structurally new challenges in agentic transformation: getting the deterministic-probabilistic boundary right, codifying tacit knowledge, aligning organisational transformation with technology, and managing the human-agent boundary as a moving variable.
• Section 12 Governance Readiness — Extend vs Build New Banks are more ready on accountability than they think, and less ready on enforcement. Three frameworks that extend to agents, and five gaps that have to be built.
• Section 13 Agents in the Boardroom Seven uses for board agents, from interrogating the board pack to turning decisions into policy — and why this is a governance question, not a productivity one.
• Section 14 The Customer Experience Four models for how customers interact with their bank once agents arrive — and why agent-initiated transactions are missing an accountability anchor, not a scoping mechanism.

This guide grew from a series of LinkedIn posts on agentic AI architecture for banking. The original posts and their comment threads — which significantly shaped the thinking here — are linked from each section.