What We Learned Shipping AI Agents Into an FCA-Regulated Compliance Workflow

2026-07-08 · Source: The AI Journal · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, AI in Regulated Industries · Depth: Advanced, medium

Summary

Redberry's team at PP Mobius, a platform for FCA-regulated Tavistock Protect, successfully deployed AI agents into a formal compliance review workflow. Facing a bottleneck of approximately 600 monthly client fact finds, they developed an architecture using independent micro-agents for specific checks, rather than a single monolithic AI. The system categorizes checks into deterministic "System checks" (code-based), "AI checks" (individual micro-agents with confidence scores and reasoning traces), and "Manual checks" (human judgment). Built on the Laravel AI SDK, agents run asynchronously, streaming results to reviewers. Regulatory demands for auditability shaped key features: every check result carries evidence, reviews are versioned and immutable, and all data exports as audit-defensible PDFs, ultimately improving engineering practices.

Key takeaway

For AI Architects or MLOps Engineers building agentic systems for operational workflows, especially in regulated environments, you should adopt a decomposed, evidence-first architecture. Focus on creating independent micro-agents for specific tasks, ensuring each result includes full reasoning and confidence levels. Prioritize immutable versioning for all reviews and integrate audit-defensible export capabilities from the outset. This scaffolding—traceability, versioning, and human oversight—is crucial for transforming agent demos into reliable infrastructure.

Key insights

AI agent success in regulated industries hinges on robust surrounding architecture, not just model performance.

Principles

• One check, one agent, one traceable result.
• Deterministic logic stays in code, not AI.
• Store reasoning and evidence, not just the verdict.

Method

Decompose compliance into System, AI, and Manual checks. Dispatch individual micro-agents for AI checks, each providing a trace and confidence score. Aggregate results asynchronously, ensuring immutable versioning and audit-defensible exports.

In practice

• Implement micro-agents for independent, auditable assessments.
• Route low-confidence AI results to human reviewers.
• Design structured PDF exports for regulatory audit defensibility.

Topics

• AI Agents
• FCA Regulation
• Compliance Workflow
• Microservices Architecture
• Audit Trails
• Financial Services
• LLM Applications

Best for: AI Architect, AI Engineer, MLOps Engineer

Related on AIssential

• AUDITFLOW: Executable Symbolic Environments for Structured Financial Reporting Verification — AuditFlow combines adaptive search with deterministic symbolic verification for robust financial audit accuracy.
• AI in Design Verification: From Experimentation to Measurable Capability — AI in design verification must demonstrate measurable capability improvements within integrated workflows, not just local task productivity.
• Agent Security Meets Regulatory Reality -- A Practitioner Systematization of Autonomous-Agent Threats and Controls in Regulated Financial Systems — Regulated financial systems amplify LLM agent security threats, demanding production-scale auditability and authorization controls.
• Agentic AI Observability: The Foundation of Trusted Enterprise AI — Agentic AI observability provides crucial visibility into multi-agent system reasoning, enabling trust and control in enterprise deployments.
• The Agentic Shift: Why Engineers Are Becoming Workflow Architects — Agentic AI transforms engineering by autonomously planning and executing tasks, requiring human oversight for safety.
• AI Agent Sandboxing for SaaS: How Builders Let Agents Work Without Letting Them Roam — AI agent sandboxing enables safe, controlled automation by enforcing strict boundaries around agent actions and resources.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The AI Journal.