The Data Steward Agent: How AI Changes the Work of Master Data Governance

In Brief

AI changes data stewardship most effectively when it automates evidence processing before it automates accountable judgment. Many stewardship activities—validation, duplicate investigation, evidence collection, routing and routine correction—contain repeatable work that machines can assist with or perform under explicit policy.

But not every master-data decision should be delegated to an agent. Identity merges, policy exceptions, legal or financial consequences and ambiguous business judgments require a control model that considers evidence quality, business impact, reversibility, policy clarity and sensitivity—not only an AI confidence score.

The Data Steward Agent should therefore operate inside the governance model rather than become the governance model. Its purpose is to make stewardship more scalable, evidence-driven and responsive while preserving explicit authority, human escalation, auditability and accountability.

Data stewardship has always combined two very different kinds of work.

The first is operational: checking required fields, investigating suspected duplicates, correcting quality issues, comparing records, routing requests and gathering evidence.

The second is judgmental: deciding whether two entities should be merged, interpreting ambiguous business context, approving exceptions and determining when an enterprise policy should change.

Traditional MDM often asks the same human Steward to perform both.

The opportunity for Agentic Data Management is not to remove stewardship. It is to separate machine-suitable work from accountable human judgment much more deliberately.

The Steward problem is not simply too much manual work

It is easy to describe Agentic MDM as a productivity problem.

There are too many records, too many exceptions and too few Stewards, so AI should process more cases.

That is only part of the problem.

A Steward may spend time gathering information from several systems before making a decision that takes only a few minutes. Another case may be easy to investigate but difficult to decide because two business units interpret the entity differently.

These are different automation opportunities.

The Data Steward Agent should automate evidence processing before it automates accountable judgment.

Break stewardship into five kinds of work

A useful starting point is to stop treating stewardship as one activity.

I would separate it into five work classes.

Work class Typical activity Automation posture
Validation Required fields, formats, reference values, deterministic policy checks and basic quality controls. Strong candidate for rules and bounded automation.
Evidence Gathering Collect source records, identifiers, relationships, history, external evidence and prior decisions. High-value area for agent assistance.
Recommendation Propose match, correction, classification, survivorship or routing decisions. Suitable where evidence and rationale can be reviewed.
Approval Accept, reject or execute a material change according to delegated authority. Automate only where explicit policy authorizes it.
Policy Judgment Resolve conflicting business interpretations, approve important exceptions or change enterprise rules. Primarily human-led and accountable.

The agent performs work inside the governance model. It does not become the governance model.

What a Data Steward Agent can actually do

Once stewardship is decomposed, the role of an agent becomes easier to define.

Detect

The agent can continuously evaluate master-data changes for missing values, invalid reference codes, unusual patterns, potential duplicate entities and other conditions that deserve attention.

Detection does not require the agent to decide the final correction.

Gather evidence

A suspected duplicate may require comparison of identifiers, addresses, source-system records, corporate relationships, history and previous Steward decisions.

An agent can assemble that evidence before the Steward opens the case.

Recommend

The agent can propose that records be linked, unlinked, corrected, classified, routed or escalated.

Recommendation is especially useful where the agent can show the evidence and policy behind its conclusion.

Execute within policy

Some actions can be safely automated when the rule is deterministic, the evidence is authoritative, the business impact is limited and the action is recoverable.

Examples may include controlled normalization, deterministic reference-code correction or other explicitly authorized transformations.

Escalate

A mature agent should also know when not to decide.

Conflicting evidence, high business impact, sensitive identity changes or unclear policy should move to the authority capable of resolving the case.

Knowing when to escalate is a core agent capability, not an automation failure.

Do not govern the agent with confidence alone

One of the weakest automation designs is a single numerical threshold.

For example:

“If the AI is more than 95% confident, automatically execute the change.”

Confidence may be useful evidence, but it does not describe the consequence of being wrong.

A 99% confidence recommendation to normalize a postal format and a 99% confidence recommendation to merge two legal entities do not carry the same risk.

Confidence can support a decision. Policy determines whether the agent has authority to execute it.

Use five dimensions to determine autonomy

I would assess an automated master-data decision through five dimensions.

Dimension Question
Evidence Quality Is the recommendation supported by authoritative and sufficiently complete evidence?
Business Impact What happens to business processes, customers, financial reporting or downstream systems if the decision is wrong?
Reversibility Can the change be safely reversed after execution?
Policy Clarity Does an explicit approved rule define the correct action?
Sensitivity Does the decision involve legal identity, financial controls, privacy, regulatory obligations or another sensitive business consequence?

Four decision bands are more useful than one threshold

These dimensions can then determine the appropriate operating mode.

Decision band Treatment Typical conditions
Auto-execute Agent executes and records evidence. Clear policy, authoritative evidence, limited impact and strong recovery path.
Recommend Agent proposes the action for review. Strong evidence but meaningful ambiguity or moderate impact.
Approval Required Agent performs analysis; human authority approves before execution. Significant business consequence, identity impact, policy exception or sensitive domain.
Human-led Agent provides evidence only. Insufficient evidence, unclear policy, strategic judgment or difficult-to-reverse impact.

Autonomy should therefore be treated as a controlled privilege rather than a permanent maturity level.

The same action can belong to different bands in different domains

There is no universal automation policy for “duplicate resolution” or “data correction.”

An incorrect duplicate merge involving a low-impact reference-data record may be easy to recover from.

An incorrect merge involving two strategic customers or two legal suppliers may affect transactions, contracts, compliance or risk reporting.

The technology function may be identical while the business consequence is completely different.

Automation policy belongs to the business-data domain, not to the algorithm alone.

Human-in-the-loop should mean authority, not ceremonial review

“Human-in-the-loop” can become a weak control if the human merely clicks Approve after an AI system has already made the practical decision.

Human oversight is useful when the person receives enough evidence to challenge the recommendation and has actual authority to reject or change it.

NIST's Generative AI Profile similarly notes that different AI contexts may require different levels of human oversight, review, tracking and documentation. :chatgpt-content-reference{index="2"}

Human-in-the-loop should mean meaningful human decision authority—not a human signature attached to an opaque machine conclusion.

IBM MDM shows what supervised automation can look like

Current IBM MDM capabilities provide a useful example of human and machine work interacting without assuming full autonomous governance.

IBM's February 2026 release added side-by-side record comparison, similarity analysis, entity history and configurable potential-match workflows for Data Stewards.

Its June 2026 release went further by allowing manual Steward decisions—such as link, unlink and potential-match decisions—to contribute to algorithm-tuning recommendations. IBM also added automatic relationship discovery and improved quality-remediation workflows.

That is an important design pattern.

The machine supports the Steward, and Steward decisions become evidence that can improve the machine.

The market is moving toward explicit Data Steward Agents

Informatica from Salesforce made this direction explicit at Informatica World in May 2026.

It announced Agentic Multidomain MDM and a Data Steward Agent intended to automate labor-intensive quality-resolution and record-matching work.

The availability statement matters, however.

Informatica listed Agentic Multidomain MDM and the Data Steward Agent as targeted for Q4 2026 and explicitly stated that customers should make purchasing decisions based on released and available functionality rather than unreleased roadmap features.

Product announcements show the market direction. Enterprise architecture should distinguish that direction from capabilities that are demonstrably available in production.

SAP reinforces the continuing role of accountable stewardship

SAP Master Data Governance continues to define explicit Data Steward roles for customer, supplier, material and other master-data domains.

Current SAP documentation describes the Steward as a role responsible for master-data quality and provides specific authorization for processing, maintenance and replication activities.

SAP's central-governance model also uses workflow in which business users raise change requests and a master-data specialist or Steward reviews the requested change.

This matters because Agentic MDM should extend an existing authority model rather than erase the distinction between requester, Steward, owner and technical system.

The future Steward becomes a supervisor of evidence and policy

If agents perform more repetitive investigation, the Steward role should change.

Traditional emphasis Emerging emphasis
Manual exception search Reviewing prioritized, evidence-rich exceptions.
Individual record correction Identifying and removing recurring root causes.
Manual duplicate comparison Adjudicating ambiguous identity evidence and supervising match policy.
Following workflow steps Designing authority boundaries and exception policy.
Applying static rules Evaluating whether rules still reflect the business.

The future Steward should spend less time processing known cases and more time governing uncertain ones.

The agent needs a feedback loop, not only a task queue

A Steward Agent should not simply process cases and forget the outcome.

The operating model should use completed decisions to improve future detection, prioritization and recommendations.

Detect → Gather Evidence → Recommend → Decide → Act → Observe → Learn

The learning stage may improve matching logic, exception routing, business rules or the criteria that determine whether similar future cases require human review.

IBM's use of manual Steward decisions in matching-tuning recommendations demonstrates one concrete version of this feedback principle.

A mature Data Steward Agent should learn from governed outcomes—not merely from its own previous predictions.

Avoid circular machine learning

Feedback creates another risk.

If an agent makes a recommendation, the system automatically executes it and that execution is then treated as evidence that the recommendation was correct, the model can reinforce its own assumptions.

Learning data should therefore distinguish between machine-generated outcomes and independently validated outcomes.

An action performed by the agent is not automatically evidence that the agent was right.

Auditability must connect the whole decision chain

Traditional MDM audit logs often show who changed which field and when.

Agentic stewardship requires more context.

Entity → Evidence → Recommendation → Policy → Authority → Action → Outcome

For a material automated decision, the enterprise should be able to reconstruct which entity was involved, what evidence was used, what the agent recommended, which policy authorized the action, who or what held execution authority and what happened afterward.

That evidence is necessary for investigation, contestability, recovery and continuous improvement.

Every autonomous action needs a recovery design

Enterprises often ask whether an agent is accurate enough to automate a decision.

They should also ask what happens when the agent is wrong.

Can the change be undone? Have downstream systems already consumed it? Can the original state be reconstructed? Will reversing the master change reverse the business transactions created after it?

These questions determine how much autonomy is responsible.

Automation should expand only where the enterprise knows how to recover when automation is wrong.

More automation should be earned through evidence

Agent autonomy should not increase simply because a pilot has been running for a fixed number of weeks.

A stronger approach uses exit criteria.

Is evidence consistently available?

Are incorrect recommendations understood?

Can changes be reversed safely?

Are policy boundaries clear?

Can monitoring detect harmful drift?

Is accountability explicit when an automated decision causes downstream impact?

Autonomy should expand only when the answers become stronger.

Do not automate a broken stewardship process

Agentic technology can accelerate a bad governance design just as easily as a good one.

If ownership is unclear, an agent may route work faster to the wrong authority.

If enterprise definitions are inconsistent, the agent may enforce inconsistency more efficiently.

If Stewards spend most of their time repairing one upstream process, automating the repair may hide the fact that the upstream process should be fixed.

Agentic MDM should remove repetitive work without institutionalizing the reason that repetitive work exists.

Governance must define what the agent is allowed to learn

Machine learning introduces another governance question.

Which human decisions should become future training or tuning signals?

A Steward override may indicate that the model was wrong. It may also represent a one-time business exception that should not change future behavior.

The system should therefore distinguish ordinary corrections from exceptional decisions.

Human feedback should improve the agent only when the enterprise understands what that feedback means.

The Data Owner remains accountable for policy

An agent may propose a new quality rule after observing hundreds of similar exceptions.

That does not mean the agent should redefine enterprise policy by itself.

Data Stewards can evaluate operational evidence. Data Owners or other accountable business authorities should retain responsibility for material policy changes and decision-right boundaries.

That broader post-go-live ownership model is developed in MDM #9, Why Enterprise MDM Governance Fails After Go-Live — and How to Make Ownership Real .

The agent may identify that policy should change. Accountable governance decides whether policy actually changes.

Measure governance outcomes, not only automation volume

An Agentic MDM program can easily optimize the wrong metric.

“Percentage of cases automatically resolved” sounds attractive, but higher automation is not always better.

I would monitor a broader set of outcomes.

Measure What it reveals
Recommendation accuracy Whether agent recommendations align with validated outcomes.
Override pattern Where humans repeatedly disagree with the agent and why.
Decision latency Whether automation actually improves business responsiveness.
Exception recurrence Whether root causes are being removed or cases are merely processed faster.
Recovery quality Whether incorrect automated decisions can be detected and reversed safely.
Steward capacity shift Whether human effort is moving from repetitive processing toward higher-value governance work.

The objective is not to maximize automated decisions. It is to improve the quality, speed and governability of master-data decisions.

Seven questions before deploying a Data Steward Agent

1. Which Steward activities are evidence processing, and which require accountable business judgment?

2. What evidence must the agent gather before making a recommendation?

3. Which actions can be auto-executed under approved policy, and which require human approval?

4. Can the organization explain why an agent recommended or executed an important decision?

5. Can an incorrect automated change be contained and reversed?

6. Which human decisions are suitable learning signals, and which represent one-time exceptions?

7. Who remains accountable when an automated master-data decision affects a business process?

If these answers are unclear, the organization is not ready for greater agent autonomy regardless of model accuracy.

The Data Steward Agent is an operating-model change

The most important mistake would be to treat the Data Steward Agent as another feature added to the Steward workbench.

Its real significance is organizational.

Detection, evidence gathering and recommendation can increasingly move to machines. Human attention can move toward ambiguous decisions, policy design, exception authority, root-cause improvement and supervision of automated behavior.

That redistribution of work creates the foundation for a broader capability developed in the next article: systems that can not only detect master-data problems but also recover from selected failures under explicit governance.

The goal is not to automate stewardship away. It is to make stewardship more scalable, evidence-driven and focused on accountable decisions.


Sources & Further Reading

IBM — What's New in Master Data Management, IBM Software Hub 5.4
IBM's June 2026 release documentation covering matching-algorithm tuning recommendations, incorporation of manual Steward decisions, relationship discovery and quality-remediation capabilities.
IBM Documentation

IBM — What's New in Master Data Management, IBM Software Hub 5.3.1
IBM's February 2026 documentation covering side-by-side record comparison, similarity analysis, entity history and potential-match workflows for Data Stewards.
IBM Documentation

IBM — Discovering Relationships in Your Master Data
Current IBM documentation describing rule-based automatic relationship discovery and explicit identification of relationships created through discovery rules.
IBM Documentation

Informatica from Salesforce — Trusted Data Foundation for AI Agents
May 2026 announcement introducing Agentic Multidomain MDM and the Data Steward Agent. At announcement, both were identified for Q4 2026 availability, with an explicit caution that unreleased functionality should not determine purchase decisions.
Informatica

SAP Learning — Introducing Central Governance of Master Data
Current SAP learning material describing change-request workflows in which business users initiate master-data changes and master-data specialists or Stewards review them.
SAP Learning

SAP Help Portal — Master Data Governance for Customer: Data Steward
Current SAP MDG documentation defining a Data Steward role with authorization for master-data processing and replication and responsibility for master-data quality.
SAP Help Portal

NIST — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
NIST guidance noting that different generative-AI contexts may require different forms of human oversight, review, documentation and governance depending on risk and context of use.
NIST


Method Note

This article uses Data Steward Agent as a practitioner operating-model concept for AI-assisted master-data stewardship. Product vendors may use the term differently, and the article does not imply that all current MDM platforms provide equivalent autonomous-agent functionality.

The work classes Validation / Evidence Gathering / Recommendation / Approval / Policy Judgment, the risk dimensions Evidence Quality / Business Impact / Reversibility / Policy Clarity / Sensitivity, and the decision bands Auto-execute / Recommend / Approval Required / Human-led are practitioner frameworks developed for this article.

The feedback loop Detect → Gather Evidence → Recommend → Decide → Act → Observe → Learn and audit chain Entity → Evidence → Recommendation → Policy → Authority → Action → Outcome are likewise conceptual operating frameworks rather than vendor-specific reference architectures.

No universal confidence threshold, automation percentage, Steward-effort reduction or fixed timetable for expanding agent autonomy is assumed. Appropriate authority should reflect evidence quality, business consequence, reversibility, policy clarity, regulatory sensitivity and organizational governance.

Vendor examples distinguish released capabilities from announced future functions. IBM examples reference capabilities documented in 2026 releases. Informatica's May 2026 announcement identified Agentic Multidomain MDM and the Data Steward Agent as Q4 2026 capabilities at the time of announcement; enterprises should verify current production availability directly during evaluation.

Reviewed: October 2026

Previous: MDM #2. AI-Ready Master Data: Why Trusted Business Context Matters More Than Clean Data

Next: MDM #4. Self-Healing Master Data: From Error Detection to Controlled Recovery

This article is part of the Enterprise MDM Series exploring how master data is evolving for AI-ready and agentic enterprises.

Comments

Popular posts from this blog

How to Read the 2026 Gartner MDM Magic Quadrant: Five Leaders, Two Acquisitions, One Buyer Framework

AI Agents: Chatbots, RPA and Agentic AI Explained

Why Enterprise MDM Governance Fails After Go-Live — and How to Make Ownership Real