By Brij Mohan, Vice President – Principal Software Dev at LPL Financial, specializing in AI-driven data governance and agentic architectures
For years, master data management (MDM) has been treated as a necessary but reactive function inside financial institutions. Issues were found after the fact, fixed manually and managed through rule sets that grew more fragile with every new data source. That model no longer holds up.
Data today sits at the center of everything: regulatory compliance, real-time decisions and client experience. Yet many organizations are still running the same batch-oriented, rule-driven MDM approaches they built a decade ago. The world in which their data operates has changed dramatically. Their data infrastructure largely hasn’t.
I’ve spent years working on large-scale financial data platforms, and the pattern is consistent: Organizations are rarely blind to their data problems. The harder challenge is resolving them fast enough without creating downstream impact elsewhere. That’s where traditional MDM tends to fall apart.
Why The Old Approach Is Breaking Down
A typical financial institution pulls data from dozens of systems, CRMs, trading platforms, custodians, risk engines and legacy databases, each generating millions of records daily in different formats with overlapping identifiers. The compounding effect is the real problem. Duplicate records skew advisor workflows and corrupt reporting pipelines. Incomplete attributes propagate across systems. In one implementation I was directly involved in, a small percentage of duplicate client records created operational friction that rippled across multiple downstream platforms. Small problem, large blast radius.
Regulatory frameworks like BCBS 239 have also raised the bar on data accuracy, lineage and timeliness. Yet according to a 2024 industry analysis published by PwC and referenced in an Informatica industry report, only two of 31 global systemically important banks are fully compliant with all its principles. At some point, continuously patching a reactive system stops being a viable strategy.
A Different Operating Model
What I’ve seen work is a shift from periodic correction to what I think of as autonomous data stewardship (ADS): embedding data reliability directly into the pipeline rather than bolting it on afterward.
The idea is straightforward. Instead of waiting for downstream systems to surface a problem, organizations detect issues closer to ingestion. Streaming architectures identify anomalies as data arrives. Predictive models flag likely quality degradation before it affects business processes. Data quality stops being a cleanup exercise and becomes an operational capability.
Why Rules Alone Aren’t Enough (And Why Pure AI Isn’t The Answer Either)
Traditional MDM systems rely heavily on deterministic rules. The problem is that real-world data rarely behaves deterministically. Small variations in a client name, mailing address or entity relationship are enough to create missed matches or false duplicates at scale.
The tempting answer is to replace rigid rules with large language models (LLMs). LLMs are remarkably good at understanding context, handling variation and interpreting messy data. But financial systems operate differently from consumer AI applications. A trade confirmation, reconciliation workflow or regulatory filing cannot be “probably correct.” It either is or it isn’t.
That’s why determinism matters more in financial environments. Outcomes must be explainable, reproducible and auditable. A reconciliation workflow can tolerate delayed processing. It cannot tolerate uncertainty in the final outcome.
Deploying purely probabilistic models inside high-stakes financial workflows introduces a different category of risk.
A more effective strategy is often a hybrid architecture that uses each approach for what it does best. LLMs handle fuzzy, context-dependent interpretation, understanding semantic similarity, entity relationships and unstructured variations. Deterministic systems and confidence thresholds then make the final classification or execution decision. The model informs the workflow, but governed systems control the outcome.
This is the part of the ADS model that often gets overlooked. It’s not about choosing AI over rules or rules over AI. It’s about building layered systems where probabilistic intelligence supports deterministic decision-making without sacrificing precision or explainability.
What AI Agents Actually Change
The most significant shift is how stewardship decisions are made. In an ADS model, specialized AI agents can handle distinct parts of the workflow, detecting anomalies, enriching records, resolving conflicts and validating outcomes. They operate within structured boundaries designed to produce consistent decisions at a scale no human team could realistically match.
The goal isn’t to remove humans from the process. It’s to focus human judgment on the decisions that genuinely require it, rather than spending time on repetitive triage.
That reallocation of effort is a meaningful operational change. Industry data supports this: A 2024 MDM market analysis found that AI-powered implementations reduced manual stewardship workload by 31% and improved entity resolution accuracy by 21%. This reallocation can also help lead to faster data integration during mergers and acquisitions, improved consistency across systems and lower operational overhead overall.
But the deeper shift is cultural. When data can be trusted in near real time, teams stop working around it. Decisions can be made faster. Reporting becomes cleaner. Data can function as a strategic asset rather than a recurring operational concern.
A Practical Starting Point
This transition does not require organizations to replace existing infrastructure overnight, but it does require an honest assessment first. Before introducing automation, leaders need visibility into what data exists, who owns it and where the highest-risk quality gaps are. Without that visibility, automation efforts often address symptoms instead of root causes.
The most practical starting point is introducing real-time data quality monitoring into existing pipelines. From there, automated resolution can be layered in incrementally, starting with high-volume, low-risk issue types where rules are already well understood.
One of the more underestimated challenges in this transition is that data stewardship has traditionally lived within specialized teams. ADS distributes that responsibility more broadly across engineering, operations and compliance, which requires active change management. Trust in these systems comes from explainability, auditability and gradual adoption, not from flipping a switch.
Closing Thought
Autonomous data stewardship is not simply another AI trend. It’s a response to growing regulatory pressure, increasing data complexity and the practical limits of manual operations.
The organizations that get ahead of this won’t just have cleaner data. They’ll have a structural advantage in how quickly and confidently they can operate.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
