Why data governance so often remains a dead letter
Many organisations have already launched data governance. What remains is a charter, a committee that meets less and less often and a catalogue filled in once, then left as it was. The concrete questions, meanwhile, go unanswered. Who approves the definition of an ‘active customer’? Who authorises access to this table? Who corrects this anomaly?
The causes can be identified. The approach targeted the entire data estate at once, with no priority. Roles were assigned on paper, with no allocated time or decision-making power. The catalogue was entrusted to the technical team, which knows the tables but not their business meaning. Finally, no rule was translated into a control: the policy specifies who may access the data, the platform lets everyone in.
The test of useful governance is simple. When a question arises about a piece of data, it finds an owner and an answer within a reasonable time.
Who is accountable for what: the roles to establish
Governance rests on a small number of roles, held by named people. Their titles vary from one organisation to another; their distribution much less so.
| Role | Held by | Accountable for |
|---|---|---|
| Data owner | A business manager in the domain | The definition, rules of use, expected level of quality, access authorisations |
| Data steward | A business or functional expert close to operations | Documentation, investigation of anomalies, maintaining the glossary |
| Data custodian | The team that operates the systems | Storage, security and technical application of the rules |
| Data protection officer | The person designated under the GDPR | Advice and oversight on the processing of personal data |
Two clarifications avoid many misunderstandings. The owner does not own the data in the legal sense: they are accountable for it to the organisation. And stewardship is not a full-time job in most domains, but an explicit part of an expert’s role, recognised by their line management.
An arbitration body completes the arrangement. It settles disagreements between domains (two competing definitions of the same indicator, for example) and approves common policies. It does not need to be new: an existing committee can carry this mandate.
Glossary, catalogue, lineage: three instruments, three uses
The business glossary fixes meaning. It contains the definitions of terms and indicators, approved by the owners. It is what puts an end to debates about what an order, a contract or a customer is.
The data catalogue says what exists. It lists datasets, their description, their owner, their classification and their known level of quality. A catalogue maintained by hand ages quickly: technical metadata must be collected automatically, with the stewards adding meaning and context.
Lineage says where a piece of data comes from and where it goes. It is used to explain a figure, to measure the impact of a change in a source application and to prove, during an audit, the journey of a piece of information. It is reliable when extracted from transformation code, much less so when reconstructed through interviews.
Policies that apply, the GDPR included
A data policy only has an effect if it translates into a control. Four families are enough to begin with: classification (what level of sensitivity), access (who, for what use, with what approval), retention (for how long, then what deletion) and sharing with third parties.
For personal data, the GDPR, Regulation (EU) 2016/679, gives these policies a mandatory basis. Its Article 5 sets out in particular the principles of purpose limitation, data minimisation, accuracy and storage limitation. Its Article 30 requires a record of processing activities to be kept. The catalogue and lineage do not replace this record, but they provide the material that makes it possible to keep it up to date and to respond to requests to exercise rights. We work with your data protection officer, without taking the place of either them or your legal advisers.
AI adds a requirement. Article 10 of the AI Act, Regulation (EU) 2024/1689, subjects the training, validation and test datasets of high-risk AI systems to data governance practices. Our entry on the AI Act presents its overall logic, and the subject continues in our AI governance offering.
Our approach: one domain, real cases, then extension
Our consultants work within your teams; decisions remain yours.
Choosing a domain and its critical data. Together we select a domain where the stakes are visible: data that feeds a leadership indicator, a regulatory report or an AI project. The scope is deliberately narrow.
Establishing the roles. Owners and stewards are named, their mandate is written down, their time is negotiated with their line management. Without this step, everything else remains theoretical.
Working through real cases. A disputed definition, a sensitive access request, a recurring anomaly: rules emerge from resolving these cases, not from a drafting exercise. Each decision made is recorded and reusable.
Tooling, then extending. The glossary and catalogue are put in place once the roles are working. The following domains adopt the model, adapted to their context. In highly decentralised organisations, this extension takes the form of federated governance, as described by the data mesh: a small number of common rules, applied by each domain.
Five failings that drain governance of its effect
- Starting with the tool. A catalogue with no owners to feed it becomes an out-of-date inventory.
- Naming without mandating. An owner who has neither the power to decide nor the time to do so is a name in a table.
- Housing governance in IT alone. Technical teams apply the rules; they cannot arbitrate the meaning of business data.
- Writing policies without controls. An access rule that is not configured on the platform binds no one.
- Confusing governing with correcting. Governance decides the expected level of quality and who deals with deviations. Measurement and controls fall under data quality, which is carried out in parallel.
Measuring that governance is working
The indicators chosen at framing concern the effect of governance, not the volume of documents produced.
- The share of critical data with a named owner and an approved definition.
- The time taken to process an access request, from request to decision.
- The share of exposed datasets that carry a description, a classification and an owner in the catalogue.
- The number of definition disagreements brought before the arbitration body, and the number actually settled.
- The use of the catalogue by business users.
The last sign is qualitative. When project teams spontaneously consult the owner of a piece of data before using it, governance has become part of working habits. This change in practice cannot be decreed: it is prepared and supported over time.