All services Every engagement has a fixed, published price. Book a strategy call
Every property you want from an AI system is inherited from the data beneath it. Why the governing has to start with the data, the five foundations AI needs, and a ten minute readiness test to run before any AI project is approved.
Run the autopsy on almost any AI failure, the discriminatory screening tool, the chatbot inventing policy, the forecast that missed by a mile, and you find the same thing underneath: a data problem that existed long before the model did. Yet most organisations are now racing to govern their models while the data those models learn from has no owner, no quality standard and no traceable history.
That order is backwards, and it is expensive. You cannot govern what a model does if nobody governs what it eats. This article explains why, and what to fix first.
The short answer. Data governance is the foundation of safe AI because every property you want from an AI system, accuracy, fairness, explainability, compliance, is inherited from the data beneath it. A model built on unowned, unmeasured, untraceable data cannot be made safe afterwards. Named ownership, quality standards, lineage and access control have to exist first. Govern the data and AI governance becomes affordable. Skip it and every AI control becomes an expensive apology.
An AI system has no properties of its own. It has the properties of its data, amplified and automated. Accuracy inherits quality: a model learning from records that are wrong, stale or duplicated does not average the problem away, it industrialises it. Fairness inherits history: train a screening model on a decade of biased decisions and it learns the bias as if it were the job description. Explainability inherits lineage: when a regulator or customer asks why the system decided something, the answer lives in where the data came from, and if nobody can trace it, nobody can explain it. Compliance inherits lawful basis: personal data used for training without a lawful basis for that use does not become lawful because a model absorbed it. It becomes an unlawful model. And resilience inherits ownership: when an unowned dataset breaks silently upstream, the model drifts silently downstream, and the first person to notice is usually a customer.
The sequence is predictable enough to schedule. The pilot works beautifully, because pilots run on hand curated data. The production system then meets the organisation’s real data, and quality problems start surfacing as model behaviour: odd scores, confident nonsense, decisions that cannot be explained. Because the symptoms appear in the model, teams patch the model, and the underlying data faults stay put, compounding quietly. Data problems do not disappear inside AI. They hide there, and every month of hiding raises the bill.
The regulatory version of this story is now written into law. The EU AI Act’s high risk requirements include explicit data governance duties: training, validation and testing data must be relevant, representative and as free of errors as possible, with documentation to prove it. An organisation without lineage and quality records cannot evidence any of that, however good its AI governance looks on paper. The register and the approval gate stand on the data underneath them.
| Foundation | What it means | What AI does without it |
|---|---|---|
| Named ownership | A person rather than a team answers for each critical dataset. | Orphan data feeds models, and nobody notices when it breaks. |
| Measured quality | Fit for purpose standards, measured, with thresholds. | Rubbish scales. The model automates the error rate. |
| Lineage | Where data came from and everything that touched it on the way. | Outputs that cannot be explained to a regulator, customer or court. |
| Access and lawful use | Who and what may use each dataset, including a lawful basis for AI use specifically. | Models trained on data nobody approved for training. |
| Lifecycle control | Data kept current, retained deliberately, deletable on request. | Models that remember what the organisation was obliged to forget. |
None of this is exotic. It is ordinary data governance, done with names attached, and it is the difference between AI you can defend and AI you can only hope about.
Before any AI project is approved, five questions about the dataset it will learn from.
Two or more answers of no, and the AI project is a data project wearing ambition. That is not a reason to cancel it. It is the reason to sequence it properly.
Data governance earned its reputation for slowness through boil the ocean programmes. The AI era version is targeted: govern the data your AI needs, first.
See the estate. A Data Governance QuickScan maps ownership, quality, lineage and access in ten days, from £14,500 + VAT, and tells you which datasets are AI ready and which are liabilities. Build the operating model. Data Governance Foundations installs named owners who have accepted the role, stewardship in real diaries and forums that decide, from £45,000 + VAT. Add leadership without the hire. A Virtual CDO keeps data on the board agenda from £7,500 a month, or the whole function runs as Data Governance as a Service from £9,500 a month + VAT. Inside live programmes, Project DG embeds data governance in ERP, migration and AI delivery, where data is profiled before migration is priced, from £22,000 a month.
With the foundations in place, the AI governance blueprint gets dramatically cheaper to run: the register sits on owned data, classifications can be evidenced, and the EU AI Act’s data duties become paperwork rather than archaeology. Strong data governance does not compete with AI ambition. It is the thing that makes the ambition safe to keep.
Because an AI system inherits every property of its data: accuracy, bias, explainability and legal standing. Data governance, meaning named ownership, measured quality, lineage and access control, is the only way to control those properties before a model amplifies them. Governing the model without governing its data is painting over damp.
Data governance controls an asset: the quality, ownership and permitted use of data. AI governance controls systems that act on that data: their approval, behaviour and change over time. They interlock, and the honest sequencing is data first, because AI governance built on ungoverned data cannot evidence its own claims.
Before scaling it, yes. Pilots tolerate weak data because they run on curated samples. Production systems meet the real estate. The practical approach is per use case: govern the datasets your first AI systems will learn from, then widen, rather than delaying AI behind an organisation wide programme.
For high risk systems, the Act requires data governance explicitly: training, validation and testing data must be relevant, representative and as free of errors as possible, with documented practices to show it. Meeting that duty without data ownership, quality measurement and lineage records is practically impossible.
Seeing the estate takes ten days with a QuickScan. Making the first AI use case ready is typically weeks, because readiness is per dataset rather than per organisation. The operating model that keeps data ready is a matter of months, and it pays back on every AI initiative that follows.
Emmanuel Olatunji
Chief Enterprise Architecture Partner
Emmanuel founded Intelance in 2017 and has led architecture, AI governance, cyber and deal engagements across the UK, Europe, the Middle East and Africa. Every Intelance engagement runs under his direction, and every guide he publishes comes from that work.
See his record on LinkedIn About IntelanceThirty minutes with a senior adviser. A clear view of where to start, what it costs and what your team can run themselves.
Talk to a senior adviser
Data Governance
Data Governance
Data Governance