Data governance in 2026: catalog, stewardship, and the AI question

The thesis

Data governance for most of the last decade was treated as a policy-and-committee exercise. Policies were written. Stewards were named. Committees met. The work rarely reached the data. In 2026 the governance teams that are getting traction have shifted to an operating model anchored in the catalog and supported by named data product owners. The committee work has shrunk. The catalog work has grown. The stewardship work has become more practical and less ceremonial. This shift is the most important governance trend in the discipline, and it is not new technology; it is a different way of running the function.

Governance has shifted from policy to operating discipline

Five years ago a typical governance program at a mid-to-large enterprise produced a binder of policies, an org chart of stewards, and a calendar of committee meetings. The policies were rarely consulted after publication. The stewards’ roles were poorly defined. The committee meetings became status updates without decisions. In 2026 the programs that are getting traction look different. The catalog is in active use. Data product owners are named per domain, with documented responsibility for the data their domain produces and consumes. Stewards are subject-matter contributors to the catalog rather than ceremonial role-holders. The committee work has shifted to lightweight quarterly review and approval rather than monthly status. The governance team’s work product is not a policy document; it is a running catalog plus a stewardship and ownership practice that the rest of the organization actually uses.

The catalog is the spine of modern governance

We made this point in our architecture article and we will repeat it here because it is the single most consequential platform decision a governance team makes in 2026. A modern data catalog (Collibra is our most common recommendation, with Atlan, Alation, and the hyperscaler-native catalogs in active consideration) is the system of record for the governance practice. It holds the business glossary, the data lineage, the ownership assignments, the data product metadata, the quality monitoring, and the policy enforcement integrations. Governance programs without a modern catalog fall back on spreadsheets and SharePoint. Governance programs with one have a place where the work lives. The decision is no longer optional. Tier-one platform decision, every time.

Stewardship, done practically

Data stewardship has been one of the most over-titled and under-resourced roles in the data discipline. The fix has not been to give stewards more authority; it has been to redefine the role around concrete deliverables. The stewards we see succeeding in 2026 are contributors to the catalog with specific responsibilities: maintain the business glossary entries for their domain, approve schema changes to the data products their domain owns, sign off on data quality monitoring thresholds, and act as the named contact for domain-specific data questions. The work is bounded. It is part of the steward’s existing role, not a ceremonial addition to it. The committees that previously consumed steward time have shrunk. The catalog where the work lives has grown.

AI governance is converging with data governance

A year ago there was an argument about whether AI governance would be a separate discipline from data governance. That argument is settling. AI workloads consume data. AI outputs are data. The access controls, lineage, quality monitoring, and ownership patterns that govern other data assets apply, with adjustments, to AI workloads. The adjustments are real (model lineage, prompt and output logging, evaluation against governance criteria) but they are extensions of the existing discipline, not a replacement for it. The governance teams that are positioned for AI are the ones that already have a working catalog and a stewardship practice. The ones that do not are scrambling to build both at the same time, often under pressure from the audit committee. We say so in client conversations because the alternative (“we will start AI governance later”) is not a defensible plan when the AI workloads are already in production.

Federation is winning over centralization

The centralized “one governance team that owns everything” model has not aged well. The organizations that have made governance work are running federated models. A small central function owns the catalog, the policies, and the cross-domain decisions. The domains own the data products inside their scope. The stewards inside each domain do the practical work. The model echoes the data mesh organizational pattern but does not require a full mesh architecture to deliver value. The federation is not a slogan. It is an operating model with documented ownership at the domain level and clear escalation paths to the central function for cross-domain issues. The governance teams running this pattern are noticeably faster, more responsive, and more trusted by the business. The ones running pure centralization are usually behind on cataloging the estate.

Concrete takeaways for the governance roadmap

• Treat the catalog as a tier-one platform decision. Pick one, stand it up, run the program inside it.
• Redefine stewardship around concrete deliverables. Catalog contribution, schema approval, quality threshold sign-off, domain contact. Bounded scope.
• Federate ownership. Central function owns the catalog and cross-domain decisions. Domains own their data products.
• Treat AI governance as an extension of data governance. Same catalog, same ownership, same stewardship, with model-specific adjustments.
• Shrink the committee footprint. Replace monthly status meetings with quarterly review and approval. Move the practical work into the catalog.