The rapid adoption of LLM agents and generative AI has exposed fundamental cracks in enterprise architecture practices that were built for deterministic systems. In banking environments, the core bottleneck has never been writing new code but integrating it into complex legacy landscapes governed by regulatory and security constraints.
Cognitive Dissonance of Productivity and Architectural Debt
Industry claims of massive AI-driven speedups clash with evidence from the randomized METR 2025 test. Experienced developers working on mature codebases spent 19% more time when using AI tools, even though they subjectively believed they had accelerated by 20%. GitClear data shows copied code now exceeds refactored code, while code churn rates jumped from 3.3% to 7.1% after widespread AI adoption. DORA metrics confirm rising throughput alongside degrading stability, turning cheap code generation into a systemic architectural risk.
Architects must therefore shift from writing documents to implementing guardrails-as-code, fitness functions, and deterministic sensors that prevent hallucinated or structurally flawed code from reaching production.
Regulatory Reality: Two Different Globes
While the EU AI Act and U.S. SR 26-2 guidance have relaxed timelines for high-risk AI systems, Russian banks operate under stricter mandates. 187-FZ on critical information infrastructure and 850-P on operational resilience require detailed mapping of business processes to infrastructure objects, with architecture teams now owning this mapping. The June 2026 3-MR guidance from the Bank of Russia mandates human-in-the-loop confirmation for AI used in critical processes such as payments when security risks are high.
Architects must also accommodate 243-FZ distinctions between sovereign and national models by building abstraction layers and routers that allow model swaps without disrupting operations.
Anatomy of AI Integrations: Where Templates Break
At the solution-architecture level, reviews now begin by classifying models as either components or actors. Components confined to sandboxes face standard OWASP GenAI LLM Top 10 risks, while agents with memory, API access, and tool-calling capabilities require the stricter OWASP Agentic Top 10 controls, including execution sandboxes and strict tool-call limits.
Agent identity management has become critical. Static API keys create confused-deputy attack surfaces; instead, each agent receives an independent identity and short-lived per-invocation delegation tokens that carry task-specific context and user permissions. Audit logs must clearly distinguish human-initiated actions from agent-executed ones.
Vendor-inherited AI poses another stealth risk. Default-off policies now treat activation of any embedded Copilot as a major architectural change requiring full review, with a dedicated registry maintained to track inherited AI exposure.
Tooling, Git, and the End of Architecture Review Boards
Russian institutions rely on homegrown or registry-approved systems rather than Western EA tools. Architectural documentation lives in Git as Markdown, PlantUML, and Structurizr files, which LLM agents parse more effectively. Internal abstraction layers supporting the Model Context Protocol have restored documentation value by making it machine-readable.
Traditional Architecture Review Boards have been disbanded in favor of a federated advice process. Domain solution architects make decisions with guild input, while regulatory guardrails remain mandatory. Updated Architecture Decision Records now capture deterministic versus probabilistic boundaries, data-flow topology, human-in-the-loop triggers, exit paths, and refusal/abstention metrics.
Non-functional requirements have evolved to include token-cost limits per transaction, full-cycle latency measurement, and explicit modeling of model uncertainty behavior.
