Good AI starts with data, not algorithms or tools. Every AI outcome, especially in the public sector where decisions affect people’s lives, depends entirely on whether the underlying data is reliable, well-governed, and fit for purpose.

Public sector organisations, like NHS partners, are keen to use AI but face understandable uncertainty regarding safety and responsibility. Much of the valuable information is trapped in older systems not designed for machine learning. The central challenge: How can AI be applied safely, practically, and sustainably while respecting privacy and public expectations?

The core challenge: AI value rests on data integrity

The common focus on the AI model misses the point; most of the work and most of the potential risks lie in the data itself.

For data to be truly AI-ready, it must be:

  • Accurate and fit for purpose: Directly supporting the desired outcome.
  • Trustworthy and Representative: Avoiding unintentional bias.
  • Governed: Supported by clear metadata and compliant with legal/privacy requirements.

Many complications and biases arise quietly, during training or as part of a knowledge base. Addressing these data issues early is far more effective and less costly than fixing problems later.

A reliable start: Outcome first, data second

Teams often rush to select a tool or model. A more dependable approach is to begin with the desired outcome-the decision or insight AI is intended to support.

This shifts the focus to: “What data do we need to achieve this outcome safely and effectively, and how will human oversight be maintained?” This frames governance and documentation as essential support structures for safe AI, not as administrative roadblocks.

Aligning data with ethical and regulatory imperatives

A data-first approach is the most practical way to meet broader ethical and regulatory expectations:

  • Ethical frameworks: Reliable data is the foundation for meeting core ethical principles like fairness and transparency, central to frameworks like Hippo’s Responsible AI Framework.
  • Regulatory adherence: Robust data governance provides the evidence required for compliance with frameworks such as the UK’s Pro-Innovation Approach to AI Regulation. This approach ensures adherence to principles like Safety, Fairness, and Accountability.

Practical steps for AI-ready data

Preparing data means establishing reliable processes that improve dependability over time:

  1. Treat metadata as essential: Embed metadata (origin, ownership, update history) early, through automated pipelines, to help models accurately interpret data.
  2. Focus on quality, not volume: Smaller, well-labelled datasets curated with subject matter experts often outperform huge, unstructured collections (especially vital in healthcare).
  3. Use synthetic data thoughtfully: Useful for early exploration and concept testing of sensitive information, but should not be relied upon for final performance validation.Work smart with unstructured data: Techniques like Retrieval-Augmented Generation (RAG) makes previously unusable data usable, but their effectiveness relies on good governance and clear metadata.

Governance and oversight: The enablers

Governance is an enabler, not a barrier. When embedded early, it provides clarity, reduces risk, and ensures human oversight.

Effective data governance includes:

  • Knowing data sources and provenance.
  • Understanding rights and restrictions (privacy, copyright).
  • Maintaining transparent audit trails.
  • Protecting sensitive data (anonymisation, access controls).
  • Integrating security measures: Clear access controls, auditable change history, and monitoring for adversarial inputs.

Ensuring human accountability:

In high-stakes public sector decisions, accountability is paramount:

  • Human-in-the-Loop: For decisions like patient triage, AI must be a support tool, with clear human review and override processes.
  • Contestability: Citizens must be able to challenge decisions. This requires accessible data and model audit trails to ensure transparency.
  • Clear Responsibility: Accountability must be mapped to a human or team; responsibility cannot default to the model.

The payoff: Predictable progress and public trust

A data-first approach leads to predictable, sustainable progress and measurable benefits:

Benefit Impact
Increased Trust Public and teams understand how systems work and why decisions are made.
Safer experimentation Clear data guardrails allow exploration without legal/ethical breaches.
Consistent behaviour Representative data supports stable and equitable model performance.
Scaling capability Good data practices support repeatability and scaling successful projects.

Closing thoughts: Focus on the data 

By prioritising accuracy, governance, and representativeness, and by embedding human oversight and accountability, teams build safer, more consistent systems. This approach aligns AI development directly with fundamental public-sector principles.

In the public sector, responsible AI isn’t just a technical goal – it’s imperative for data governance and public trust

Discover more