Written for decision-makers and practitioners in large financial institutions who need a realistic, evidence-based picture of what AI can — and cannot — do for the EA discipline.
Enterprise Architecture has long been one of the most strategically important yet chronically under-resourced disciplines in large organisations. EA teams at major banks carry an extraordinary mandate — aligning technology to business strategy, managing risk across vast and complex application landscapes, governing change, and ensuring regulatory compliance — yet they often operate with tools and processes that have changed little in the last decade.
The emergence of capable AI systems in 2024–2026 creates a meaningful opportunity for the EA discipline. Not a revolution, but a genuine shift in what is tractable: AI can reduce the manual overhead of landscape maintenance, assist in drafting assessments, and make architecture knowledge more accessible across the organisation.
The vision at the centre of this paper is living architecture: a continuously maintained, queryable model of the enterprise that reflects reality more closely than today's static documentation. This is a compelling direction — but it is important to state clearly what it is not: it is not a system that reasons autonomously over your architecture, produces reliable decisions without human review, or eliminates the need for skilled enterprise architects.
This paper is written for practitioners and decision-makers who need a realistic picture of what AI can and cannot do for EA, what the prerequisites are, what can go wrong, and how to pursue the opportunity in a way that is defensible to regulators, auditors, and experienced architects.
The mandate, day-to-day realities and structural capacity problems of EA in large banks — and why AI has the potential to close the gap.
Precise analysis of where LLMs bring genuine value (drafting, synthesis, retrieval) versus where they fail unreliably (hallucination, incomplete data, deep causal reasoning).
Why data quality is the non-negotiable prerequisite — and what must be done before AI can be effective: canonical registers, entity resolution, data ownership.
Incorrect impact analysis, hallucinated references, false confidence, audit rejection, and errors at scale — with concrete mitigations for each.
Accountability, explainability, reproducibility, and prompt versioning — what a regulated European bank must put in place before going to production.
A three-layer architecture (retrieval, reasoning, validation) with a build vs buy recommendation and a full reference table across EA repository, ITSM, document layer, retrieval, LLM and audit.
From AI as a drafting assistant (weeks) through targeted integration (months) to the living architecture platform (6–18 months), with key principles and funding considerations.
A week-by-week PoC plan for AI-assisted impact analysis structuring, with explicit success criteria, kill criteria, and skills requirements for both EA and implementation teams.
This paper was authored by Rob Vugts, GenAI Engineer & Senior Technical Architect with 30+ years of experience in software engineering. Rob has not worked as an enterprise architect himself, but has spent much of his career working alongside EA teams in large organisations — observing their struggles first-hand and actively helping to solve them. In his role as a technical architect he supplied EA colleagues with the data they needed most: infrastructure inventories, application landscape data, TCO analyses, end-of-life schedules, and capacity planning. He has also built tools that made it easier for EA teams to collect, query, and maintain that information.
That vantage point — close to EA practice but from the technical delivery side — shapes the perspective of this paper. The problems described are not theoretical; they are the ones Rob watched experienced architects struggle with over many years. The paper is written with that practitioner's reality in mind, not from a vendor or consultant's distance.
Rob works through AI-chitect, advising organisations on AI-assisted development, Azure Cloud architecture, and the integration of AI into complex enterprise workflows. The paper reflects the state of available technology and practice as of April 2026 and has been developed with specific attention to the requirements of a regulated European banking environment, including DORA, EBA ICT risk guidelines, and the EU AI Act.
This paper was developed in collaboration with LLMs (large language models). The core ideas, framing, and conclusions were shaped and curated by Rob, with LLMs serving as thinking partners throughout — helping to articulate and stress-test ideas, improve structure, identify blind spots and alternative viewpoints, verify current technologies, and provide critical reviews to avoid overstating AI’s capabilities or understating its risks.
That last point is deliberate. One of the core arguments of this paper is that AI outputs require human validation, and that overconfident AI-generated content is a specific risk in governance contexts. It would be inconsistent — and dishonest — to produce a paper making that argument without applying the same scrutiny to the paper itself. The critical review process, conducted with LLM assistance, was specifically aimed at catching places where the paper might be too optimistic, too vague, or insufficiently grounded in evidence.
The full 22-page paper is available as a free PDF. Share freely with attribution — do not reproduce or adapt without permission.
© 2026 Rob Vugts / AI-chitect. Share freely with attribution. Do not reproduce or adapt without permission.