A trade may be one transaction to a bank, but it does not remain recognizable in the same way as it moves through the post-trade landscape. Different vendor platforms can assign their own identifiers, field names and data structures to the same underlying transaction.
For a leading global bank, this made reconciliation more complicated than comparing two sets of records. A trade could pass through multiple systems, yet there was no common key that reliably linked its representations from beginning to end.
The challenge was compounded by limited visibility into the vendor platforms themselves. The bank could work with the transaction data and logs they exposed, but not necessarily the internal architecture or data models that produced them. Analysts therefore had to infer how fields related by examining schemas, data types, values and trade context.
That interpretive work could require significant effort before reconciliation even began. Every new data source or requirement introduced more schema analysis, entity mapping and configuration work. The bank needed a faster way to establish those relationships without weakening the human oversight, explainability and repeatability expected of a critical financial process.
Persistent focused on separating the work that required interpretation from the work that demanded consistent execution.
Using AI to Make Sense of Disparate Trade Data
The first opportunity was the work that happened before records could be matched: determining how the same trade was represented across disparate systems.
Persistent’s approach uses schema information and representative data to identify potential relationships between fields. Trade-specific context helps the model recognize when differently named or structured data elements may refer to the same underlying trade attribute.
The recommendations give trade analysts a stronger starting point. The AI assigns a confidence score to each proposed mapping, helping analysts prioritize which recommendations are more likely to be valid and which require closer examination or further refinement.
The AI does not decide which relationships will govern reconciliation. Analysts review the proposed mappings against their knowledge of the transaction and the requirement, approving what is valid, rejecting what is not and investigating ambiguity.
This shifts their role from discovering every relationship field by field to applying expert judgment where it matters most. Rather than replacing analyst expertise, AI helps direct it toward validation, uncertainty and approval.
The confidence score guides where analysts focus their attention; it does not determine whether a mapping moves into operation. That decision remains with the subject matter expert, creating a clear handoff from AI-assisted interpretation to controlled execution.
Designing for Intelligence Without Giving Up Control
Once the mappings are approved, the problem changes. Interpretation gives way to execution, where the priority is to apply the same logic consistently and traceably.
Once approved, the mappings become part of a controlled configuration. Reconciliation jobs then apply the agreed mappings, rules and tolerances to process records and publish results for business review.
AI is not part of this routine execution path. Once a mapping has been validated, deterministic processing applies the controlled configuration each time the reconciliation runs. Authorized users can review the results and make adjustments where required, without making every transaction dependent on a model-generated decision.
Building Human Oversight into the Process
Human oversight is built into the progression from interpretation to execution. Subject matter experts approve the configuration before it is used, while version control and audit logging provide a record of what changed, what was approved and how a reconciliation run was performed.
Security and governance are part of the architecture rather than surrounding controls. Guardrails protect sensitive information used by the AI-assisted components, and the solution has been reviewed through the bank’s architecture and AI governance processes. Together, these controls create a traceable path from a proposed relationship to the logic used in execution.
Creating a Faster Path to Reconciliation
Reducing the work required to interpret new schemas and establish mappings can shorten the path from a reconciliation requirement to an operational process. Analysts can spend less time searching for relationships manually and more time validating recommendations, resolving ambiguity and applying judgment.
Because AI is not involved in every reconciliation run, the bank can benefit from AI-assisted mapping without making routine processing dependent on continued model usage. Deterministic execution preserves the consistency and traceability required in a controlled financial process.
As validated mappings and reconciliation patterns accumulate, teams can build on work that has already been approved rather than beginning every new requirement with a blank page. The approach creates the potential for:
- Less manual effort in schema analysis and entity mapping
- A faster route for onboarding new reconciliation requirements
- Greater business ownership through analyst-approved configurations
- Lower recurring AI dependency during routine processing
- Stronger traceability from mapping approval through execution
- A reusable starting point for future reconciliation use cases
Scaling Without Adding Complexity
The broader value lies in giving the bank a repeatable way to absorb new data sources and requirements without increasing manual effort at the same rate as complexity.
New schemas will still need to be understood and new relationships will still require expert validation. The difference is that the bank now has a framework in which AI can help interpret variability, analysts remain accountable for what is approved and deterministic processing preserves consistency in operation. Over time, validated mappings and controlled configurations can provide a stronger starting point for what comes next.
The result is not AI everywhere. It is AI applied precisely where interpretation is hardest, with the discipline to step aside once the answer must be repeatable.
