Background

The bank’s month-end close required reconciling data across more than 40 subsidiaries, a process that consumed 800 analyst hours and still produced material misstatements. Point-solution RPA bots automated individual steps, but they broke whenever source schemas changed and could not reason across the full picture.

In early 2026, the bank commissioned ELIAMA to design a permanent operating system: a private, deterministic AI layer that could remember every ledger relationship, reason about exceptions, and post adjustments directly into the core banking platform.

Applicability

The Autonomous Reconciliation OS now covers all general ledgers, cash positions and intercompany balances across the bank’s European and North American entities. Because its outputs materially influence adjustment decisions and regulatory reporting, accuracy, explainability and auditability were mandatory from day one.

The system was intentionally scoped to the bank’s controlled environment. No data leaves the private VPC, no model is trained on client information, and every recommendation is traceable to source records.

Requirements

The build required three integrated layers. The Memory Layer unifies structured ledger data and unstructured supporting documents in a private RAG and knowledge graph. The Logic Layer dispatches deterministic reconciler agents that collaborate, escalate exceptions, and propose adjustments. The Action Layer posts approved adjustments via the core banking API under strict human-in-the-loop approval.

LLMOps monitoring, prompt versioning, cost tracking and continuous evaluation were embedded from deployment so the system stays measurable and accountable as the bank’s data evolves.

Results

Reconciliation cycle reduced from 14 days to 4 hours.

Zero material misstatements in three consecutive quarters.

70% of analyst time reallocated from repetitive matching to exception investigation.

Key takeaways

Point-solution automation decays. Sustainable AI power comes from a coherent operating system, not a chain of brittle bots.

Sovereign AI systems require memory, logic and action layers under the client’s control — not borrowed interfaces.

Human reviewers remain in command when deterministic guardrails, citations and approval gates are engineered into the architecture from the start.