By John Joseph, Senior Director, North America Revenue, Xceptor
The financial services industry is moving beyond AI that simply generates information towards deploying agents that make decisions in live operational environments. While this can deliver significant gains in efficiency, streamline regulatory compliance, and reduce the risk of human error, it has also increased scrutiny of data governance and control.
During a panel discussion at Data Management Summit New York, one theme emerged consistently: many firms are still struggling to define what governance means for unstructured data.
In the session, we asked the audience: what is the hardest part about applying data governance consistently to AI-ready data today? The top response was governing unstructured content to the same standard as structured data (35%), followed by assigning clear accountability for AI inputs (35%).
The results reflect what we’re seeing across the capital markets industry. As AI makes more operational decisions and data becomes increasingly unstructured, firms are recognising that reliable outcomes depend on effective data governance.
Identifying critical data
One of the most common misconceptions is that deciding what data is critical requires different criteria for structured and unstructured data.
Consider an interest rate swap. The transaction may contain dozens of critical data elements (CDEs) that drive risk calculations, reporting requirements, and operational decisions. Whether those elements exist in a database table, spreadsheet, or PDF does not change their importance. CDEs are defined by regulatory, legal, financial, and operational requirements, not by the format in which information arrives.
The real challenge is that structured data arrives in defined formats with established schemas, making validation relatively straightforward. Unstructured information does not follow the same patterns. A document may contain varying layouts, images, tables, handwritten notes, inconsistent formatting, and different terminology, all of which must be interpreted and transformed into a structured format.
This is where the risk of errors increases. An AI model may misinterpret a numerical value, confuse similar characters, or fail to recognise important contextual information embedded within a document. Something as simple as confusing a dollar sign with the number five can materially affect downstream decisions.
The criteria used to identify CDEs have not changed, but unstructured data requires a different approach for extraction, interpretation, and validation.
Governance frameworks for unstructured data
Managing unstructured data begins with extracting information from large volumes of documents in varying formats. Unlike traditional rules-based automation, which relies on predefined structures and field locations, AI agents can extract and interpret data regardless of the format or layout.
Once information has been identified, it needs to be transformed into a structured format that can be consumed by downstream systems and AI applications. This includes classifying data, assigning metadata, and applying business rules that determine how it will be used.
The final step is verification. Automated controls can compare extracted data against reference sources, business rules, and other trusted datasets to identify discrepancies and exceptions. Many institutions still rely heavily on human intervention to investigate and resolve breaks, which can be time-consuming, inconsistent, and difficult to scale. AI will increasingly be used to identify and prioritise exceptions, surface root causes, and help teams focus on the highest-value tasks.
However, whether an AI agent suggests an action or takes one autonomously depends on the level of risk and confidence involved. Where certainty is high and risk is low, such as resolving an exception that has been seen many times before, an agent may be able to resolve it automatically. Where confidence is lower and risk is higher, the agent may provide a recommendation for human review and approval. Combining agentic AI, workflow automation, and human oversight with clear accountability is what makes this viable.
Equally important is how information is accessed. Many firms do not want to spend time extracting and replicating large volumes of data into central systems. They increasingly want data to remain where it resides, using API-based connectivity to retrieve only the information required at the point of need. This reduces duplication, improves control, and ensures processes and AI agents are working with the most current data available.
Maintaining control in AI-driven operations
As AI becomes embedded within operational workflows, governance extends beyond individual data elements.
The future operating model is one where people, workflow automation, and agents work seamlessly across a connected ecosystem. As a result, operations leaders need visibility across the full chain. This includes a clear view of what data is being used across systems, workflows, and agents, how decisions are made, and full audit trails.
In this environment, governing unstructured data still requires a similar framework to structured data: identifying critical information, establishing ownership, validating quality, and maintaining accountability. It’s the approach for extracting and validating data that differs, and this is where Agentic AI shines.
For example, Xceptor’s Confirmations Extraction Agent extracts and transforms large volumes of unstructured data without needing to manually build new templates.
By combining AI with business rules and multiple validation steps, accuracy rates for AI outputs reach the high 90s. Human-in-the-loop controls and confidence scores mean firms decide what goes straight through and what gets reviewed, with a full audit trail at every step.
Through this approach, underpinned by decades of post-trade expertise, we help firms use AI to manage unstructured data in real-world operational environments, without increasing risk or compromising control.
Find out more: https://www.xceptor.com/solutions/post-trade-operations
Subscribe to our newsletter
