Data stewardship and AI governance: how they fit together

Data governance sets the rules: the policies, standards and decision rights for how data is managed. Data stewardship is how those rules are carried out day to day. Stewards are the named people accountable for specific datasets. They keep definitions clear, quality acceptable, access appropriate and problems resolved. Governance decides what good looks like; stewardship makes sure it happens, dataset by dataset.

AI governance is oversight of the AI systems that use data: which uses are approved, how risky each one is, how it is tested and monitored, who reviews its output and who is accountable when it goes wrong. It adds what stewardship cannot cover: the behaviour of a system, not just its inputs. A perfectly stewarded dataset can still feed a model that behaves badly.

Stewardship feeds governance. AI governance depends on stewards to say what data a system may use, where it came from, how reliable it is and who can approve a new use. In return, AI governance tells stewards which datasets matter most, because they now feed models that make decisions. In the MASTER-AI™ sequence, steward knowledge supports define (inventory and ownership) and observe (data lineage and drift).

Object. Stewardship looks after data assets; AI governance oversees systems and the decisions they make. Risk. Stewardship guards quality, privacy and misuse of data; AI governance adds bias, safety, autonomy and accountability for outcomes. Pace. Data changes slowly and predictably; a model's behaviour can shift when it is updated or retrained. Evidence. Stewards keep catalogues, glossaries and lineage; AI governance keeps a model inventory, risk classifications, test results and monitoring.

Four places: approving data for training or use in a model, tracing lineage from source to output, assessing privacy risk, and handling incidents when a bad output is traced back to bad data. A data protection impact assessment is both a stewardship and an AI governance task, and where personal data is involved UK GDPR applies to both. Duplicated effort in these areas is the usual sign the two are not coordinated.

Not for long. Controls on a model cannot fix a problem in the data behind it, and without stewards nobody can say with confidence what data a system is using or whether it should be. You can start AI governance first, but you will quickly need the stewardship answers: who owns this dataset, and may we use it for this?

Stewards can and should contribute, but they should not carry AI governance alone. Stewardship is about specific datasets. AI governance needs someone accountable for every AI use, with authority across the business and a line to the board. Smaller organisations often combine the roles in a small group, as long as accountability for AI is named and not assumed.

Start with an inventory of where AI is used and what data each use touches. That one list shows which datasets need a steward first and which AI uses need approval first. Name an owner for each, agree who can approve a new use, and set a review date. Then build the rest in order of risk, not all at once.