The steward queue was always the bottleneck. In 2026, the agents that clear it became the platform.
For thirty years, master data management in a bank ran on a queue. A match falls below the auto-merge threshold, a new counterparty arrives without a clean identifier, a sanctions screening throws an ambiguous hit, a customer record forks across the core banking system and the CRM — and each one lands in a work list for a human data steward to adjudicate. The steward opens the candidate records side by side, picks the surviving values, applies the fix in one system, and moves to the next item. The model worked. It simply never scaled to the volume of a modern financial institution, and everyone in the function knew it.
In the first half of 2026, the vendors that own the MDM category stopped treating that queue as a fact of life and started treating it as the thing to automate. Reltio shipped autonomous stewardship agents. Informatica announced the industry's first Agentic Multidomain MDM. SAP acquired Reltio outright for an undisclosed sum widely reported in the billions. The common thread is a single architectural claim: the data steward's judgement, applied at machine speed and under full governance, is now a software capability rather than a headcount line. For financial services — the most data-intensive, most heavily regulated, most fraud-exposed industry in the economy — that claim changes the operating model of the entire data function.
The economics that made manual stewardship untenable
The case for automating data stewardship in financial services was never about elegance. It was about arithmetic that had stopped working.
Start with the compliance perimeter. Financial institutions spend an average of 72.9 million dollars a year on KYC and AML operations, and the average institution employs 642 full-time staff doing nothing but that work. A large share of that effort is, at its core, master data work: resolving whether two records are the same legal entity, reconciling ultimate beneficial ownership across jurisdictions, keeping a customer's risk attributes current, and maintaining the "golden source" that downstream sanctions and transaction-monitoring systems trust. When the entity resolution underneath is weak, the false-positive rate on screening climbs, the investigator headcount climbs with it, and the cost line compounds.
Then add the regulatory pressure on data quality itself. More than a decade after the Basel Committee published BCBS 239, only two of the thirty-one global systemically important banks are fully compliant with its principles for risk data aggregation and risk reporting. Deloitte found that sixty-eight percent of banks expect BCBS 239 to improve business steering, but only twenty-one percent say they have actually achieved it. The European Central Bank's final RDARR guidelines now interlock with DORA — if the risk-data infrastructure fails during a stress event, an institution can breach both regimes at once — and supervisors began assessing banks against the tightened guidelines during SREP examinations from 2025 onward. The regulator's expectation is no longer a one-time clean-up. It is end-to-end lineage, automated validation, and continuous monitoring. None of that is achievable on a manual stewardship model.
The market has priced this in. The agentic AI in financial services market is estimated at 7.78 billion dollars in 2026, up from 5.51 billion in 2025, and is projected to reach 43.52 billion by 2031 at a compound annual growth rate above forty-one percent. The broader AI-in-data-management market, where the banking, financial services and insurance sector already holds the largest share, is valued at 38.67 billion dollars and forecast to reach 314.27 billion by 2035. Fraud detection and anti-money laundering account for the single largest slice of that spend. The money is moving toward exactly the workloads that sit on top of master data.
What "autonomous data stewardship" actually means
The phrase is doing real work, so it is worth being precise about what changed. Embedded AI in MDM is not new — match scoring, fuzzy comparison, and survivorship suggestions have used machine learning for years. What is new in 2026 is the shift from AI as an embedded helper inside the platform to AI as the operating model of the platform. The agent does not suggest a merge for a human to approve. Within defined guardrails, the agent resolves the record, applies the change across the connected estate, writes the lineage, and escalates only the cases that fall outside its mandate.
Informatica made the clearest statement of the pattern at Informatica World 2026, announcing what it called the industry's first Agentic Multidomain MDM: a continuously running system in which autonomous agents cleanse, steward, enrich and match master data in real time, explicitly positioned as a replacement for the slow, manual, human-dependent MDM model. Its new Data Steward Agent automates the labour-intensive work of resolving quality issues and matching records, and the company frames the design goal as ensuring every piece of master data "carries its own instruction manual" so that downstream autonomous agents can make high-integrity decisions with full lineage and transparency. CLAIRE now operates as a fully headless, multi-agent intelligence layer with native Model Context Protocol support, exposing every data-management capability as a governed service any agent can invoke. The headless data management and data quality agents reached general availability in spring 2026; Agentic Multidomain MDM is scheduled for the fourth quarter.
Reltio reached the same destination from the customer-operations side. AgentFlow delivers autonomous agents that handle routine stewardship, governance and custom workflows on trusted enterprise data, opening the door to genuinely hands-off data operations. AgentFlow Unstructured turns documents, transcripts, PDFs and scanned files into governed context — directly relevant to a KYC function drowning in onboarding paperwork — and its reference-data agents update RDM from a simple file upload, with no mappings, no ETL, and no custom UI work. The strategic signal sat on top of the product one: in March 2026 SAP announced its acquisition of Reltio, completed it in May, and stated the rationale plainly — make SAP and non-SAP data AI-ready by governing and exposing the golden record as a trusted, context-rich data product for both analytics and AI agents.
Where the agents land first in financial services
Autonomous stewardship does not arrive everywhere at once. In financial services it lands first in four places, each chosen because the cost of the manual version is highest and the rules for a correct answer are clearest.
The first is customer onboarding and KYC. Entity resolution, duplicate suppression, and attribute enrichment are textbook stewardship tasks, and they sit directly on the critical path of revenue. The operational evidence is already concrete: one large Dutch financial institution combining AI across its KYC and compliance processes reported a ninety percent reduction in onboarding time and a thirty percent cut in staff workload. When the golden record assembles itself as the documents arrive, onboarding stops waiting on a queue.
The second is AML and financial-crime data. The golden source that feeds sanctions screening and transaction monitoring is a master-data artefact, and the quality of that source is the single biggest lever on false-positive rates. Autonomous stewardship keeps the source current, reconciles beneficial ownership, and maintains the entity graph that distinguishes a genuine match from a coincidence of names — which is why fraud and AML attract the largest share of AI-in-data-management investment in the sector.
The third is risk data aggregation under BCBS 239 and RDARR. Supervisors now expect automated validation, end-to-end lineage and continuous monitoring rather than periodic manual reconciliation. An agent that stewards the master data continuously, and records why every value survived, produces exactly the audit trail the examination is looking for — and closes the gap between the sixty-eight percent of banks that expect business-steering benefit and the twenty-one percent that achieve it.
The fourth is reference data management — currency codes, product hierarchies, counterparty classifications, regulatory taxonomies. It is unglamorous, high-volume, rule-bound work, which makes it almost ideal for an agent that can ingest a file and update the reference set without an ETL project behind it.
The data foundation still decides the outcome
The enthusiasm needs a counterweight, and the counterweight is the same one that has governed every wave of enterprise AI: the agent is only as trustworthy as the data and governance beneath it. The adoption numbers make the point bluntly. Ninety-nine percent of organisations plan to put agents into production, but only eleven percent have actually done so, and the obstacles are not model quality — they are data, governance and security. Forty-eight percent of organisations cite governance concerns and thirty percent flag privacy as the blocking issues.
For a bank, an autonomous steward acting on incomplete lineage or an ungoverned policy is not a productivity tool; it is an unbounded operational and regulatory risk that executes faster than a human could catch it. This is precisely why the credible vendors lead with governance, lineage and transparency rather than autonomy for its own sake, and why "every record carries its own instruction manual" is a design principle rather than a slogan. It is also why SAP was willing to acquire a master-data platform to anchor its AI strategy: the entity-resolution and golden-record layer is the thing the agents reason over, and owning it is owning the foundation. The Model Context Protocol matters here too — by exposing stewardship capabilities as governed, invokable services, MCP lets a bank compose agents across Reltio, Informatica, Salesforce and the cloud platforms without rebuilding the architecture, while keeping the governance boundary intact.
The practical implication is uncomfortable but clarifying: an institution cannot buy its way out of a weak data foundation by adding agents on top. The agents make a strong foundation dramatically more productive and a weak one dramatically more dangerous.
The operating model, not just the tooling, is what changes
The deepest change is organisational. When agents clear the routine queue, the data steward's job does not disappear — it moves up the value chain. The work shifts from adjudicating thousands of low-ambiguity matches to governing the policies the agents enforce, designing the guardrails and escalation thresholds, and reviewing the genuinely hard exceptions the agents are right to refuse. Case-study evidence already shows golden-record automation saving more than five thousand steward-hours a year in a single deployment; that time does not vanish from the budget, it redeploys to higher-judgement work.
That is a different role, and it requires a different operating model: clear policy ownership, explicit decision rights about what an agent may resolve autonomously versus what it must escalate, continuous monitoring of agent decisions, and a feedback loop that tunes the guardrails as the agents learn the institution's data. Banks that treat autonomous stewardship as a tooling upgrade will get a faster version of the queue. Banks that treat it as an operating-model change will get a data function that scales with the business instead of with headcount.
What financial-services leaders should do now
The sequencing follows from everything above. First, fix the foundation before pointing agents at it — entity resolution, golden-record quality, and lineage are prerequisites, not parallel workstreams, and the eleven-percent production figure is a warning about skipping this step. Second, choose the beachhead by cost and clarity: KYC onboarding and AML golden sources offer the highest manual cost and the clearest rules, which makes them the right place to prove the model. Third, design the guardrails before switching on autonomy — define what an agent may resolve, what it must escalate, and how every decision is logged for examination, because a regulator will eventually ask. Fourth, redesign the steward role deliberately, moving the team from queue-clearing to exception governance rather than letting the change happen by attrition. And fifth, treat MCP posture and agent-readiness as procurement criteria for every platform in the data estate, because the institutions building composable agentic stacks are already asking the question.
Autonomous data stewardship is not a forecast for financial services any more. The platforms shipped, the largest enterprise software vendor in the world bought the category leader, and the regulatory and cost pressures that make the manual model untenable are all moving in the same direction. The question for 2026 is no longer whether agents will steward master data in a bank. It is whether the data foundation underneath them is good enough to let them.
Apptad partners with banks, insurers and capital-markets firms to industrialise autonomous data stewardship — the golden-record foundation, the governance and lineage that make agents trustworthy, and the operating model that moves stewards from clearing queues to governing exceptions — across Reltio AgentFlow, Informatica CLAIRE Agents and Agentic Multidomain MDM, STIBO, Salesforce Data Cloud, Databricks, Snowflake and the surrounding ecosystem.
If your KYC, AML and BCBS 239 programmes are running faster than the data foundation underneath them can support, that is the conversation worth having (opens in new tab).
