By: Georgette Virgo
AI can spot the likely duplicates. It cannot always be trusted to decide what becomes one.
That is the fault line in modern master data management. A model may recognize that two customer profiles, supplier records, or product listings probably refer to the same thing. But a mistaken merge can overwrite history, distort reporting, or send an AI system confidently in the wrong direction.
LakeFusion MDM, available as a Databricks App in Databricks Marketplace, combines AI‑assisted entity matching with stewardship workflows for decisions that require human judgment. Configurable confidence thresholds and review queues help route ambiguous matches to stewards, who can approve, reject, or adjust proposed decisions.
The important distinction is that this is not a promise that people review every match. The design appears intended to automate high‑confidence decisions while routing uncertainty and exceptions to people. Exactly where that line is drawn, and whether organizations can make it mandatory for particular data domains, is a question of configuration and governance rather than AI alone.
When Matching Needs Business Judgment
LakeFusion’s matching engine combines deterministic rules, vector similarity, and large language models to identify records that may represent the same customer, supplier, product, provider, location, or asset. Semantic, phonetic, exact, and partial matching methods help it find connections across fragmented enterprise data that conventional rules alone can miss.
But similarity is not the same as identity. Two legal subsidiaries may share a billing address and still need to remain separate. A franchise can share its parent company’s name while representing a distinct business entity. In cases like these, the matching engine can surface the connection, but it cannot determine the right outcome without the business rules, source context, and institutional knowledge behind the data.
That is where stewardship matters. LakeFusion routes ambiguous or high‑impact matches into workflows where business stewards can review, adjust, approve, or reject proposed decisions. Source‑level lineage and auditable overrides add context to the process, helping teams understand both the evidence behind a match and the action taken.
The goal is not to put a person in front of every match. It is to automate the routine decisions while reserving human judgment for the records where a wrong merge would carry real operational, regulatory, or customer consequences.
Two Layers of Accountability
Trust in master data requires more than knowing where a value came from. It also requires understanding how a decision was made.
Unity Catalog provides the technical layer: lineage that traces data across tables and columns, including supported upstream and downstream dependencies. LakeFusion adds the operational layer, giving data stewards a governed workflow to review and manage the matching decisions that shape golden records.
Together, these capabilities create a more complete accountability model. Lineage explains the data’s path through the platform. Stewardship explains the judgment applied when records are matched, merged, or kept separate. One shows how a value moved; the other helps show why the value became authoritative.
That distinction matters as mastered data feeds analytics, business applications, and AI. Organizations need visibility into both the technical flow of information and the decisions that determine which version of a customer, supplier, product, or provider record the business ultimately trusts.
The Stakes Rise When AI Can Act
As master data moves beyond dashboards and into applications and AI agents, the cost of a wrong match rises with it. An inaccurate customer count can be corrected in a report. A flawed supplier, product, or account record can carry further consequences when downstream systems use it to trigger actions.
AI can narrow millions of records to the most likely matches, apply rules consistently, and handle routine cases at scale. But it cannot independently know when two similar entities must remain distinct because of a contractual relationship, a regulatory boundary, or the way a business actually operates.
That is why the steward still signs off. Human review is not a rejection of AI‑assisted MDM, nor does it require a person to inspect every record. It is the control point for ambiguous, high‑impact decisions, where an organization must decide what becomes authoritative, document why, and stand behind the result.
For LakeFusion and Databricks, the architectural bet is that this work can happen closer to the data, within the governance environment customers already operate. But the central lesson is broader: as AI gains the ability to act on enterprise data, mastering that data remains a question of accountable judgment. AI can recommend the match. The steward decides whether the business should trust it.
