Skip to main content

Building an AI Governance Framework That Actually Holds Up

Published On

Oct 20, 2026

Category

  • Privacy
  • AI Governance
Building an AI Governance Framework That Actually Holds Up

Most organisations did not decide to adopt AI. It arrived — through a feature toggle in a SaaS tool, a developer experimenting with an API, a marketing team pasting copy into a chatbot. By the time anyone asks “what are we actually doing with AI?”, the honest answer is usually “more than we can list.” Governance is how you turn that sprawl into something you can see, justify, and defend.

This is not about slowing teams down. A good framework is mostly about knowing what you have, deciding how much scrutiny each use deserves, and keeping a record. The rest follows.

Start with an inventory, not a policy

A policy written before you know what you are governing tends to be aspirational and ignored. Begin with an inventory instead. For every AI use — embedded features, standalone tools, in-house models, API calls — capture a few facts:

  • What it does, in one plain sentence.
  • Who owns it internally and which vendor or model sits behind it.
  • What data goes in (especially personal data, and whether any of it is special-category).
  • What comes out, and what decisions or actions ride on that output.
  • Whether a human reviews the output before it affects anyone.

You will be surprised twice: by how many entries appear, and by how many touch personal data. Keep the inventory somewhere it will actually be maintained, and assign an owner. An inventory nobody updates is a snapshot, not a control.

Tier the risk

Not every use deserves the same attention. A model that drafts internal meeting notes is not the model that screens job applicants. Sort each entry into rough tiers based on the consequences of getting it wrong:

  • Low — internal productivity, no personal data, no decisions about people. Light touch.
  • Medium — touches personal data or customer-facing content, but a human stays in the loop. Standard review.
  • High — affects access to a service, employment, credit, safety, or legal standing; or processes sensitive data at scale. Full scrutiny before launch.

Tiering keeps your governance proportionate. It also tells you where to spend your limited DPIA and review time.

Where the EU AI Act fits, at a high level

The EU AI Act layers a product-safety logic on top of data protection. Without reproducing the legal text, the structure is worth knowing because it maps cleanly onto your tiers:

  • Prohibited practices — a narrow set of uses considered unacceptable, such as social scoring or certain manipulative or biometric techniques.
  • High-risk systems — uses in areas like employment, education, essential services, and critical infrastructure carry the heaviest obligations: risk management, data governance, logging, human oversight, and documentation.
  • Limited-risk uses — mainly transparency duties, for example telling people they are interacting with an AI system or that content is AI-generated.
  • Minimal-risk — the bulk of everyday uses, with few specific obligations.

There are separate provisions for general-purpose AI models. The point for governance is simpler than the detail: the higher the tier you assigned, the more likely the Act expects formal records and oversight — and the more those records overlap with what GDPR already asks of you.

DPIAs for AI and automated decisions

Where AI processes personal data in a way likely to be high-risk to individuals, GDPR Article 35 calls for a Data Protection Impact Assessment. AI use cases hit the common triggers often: large-scale processing, profiling, evaluation, and systematic monitoring.

Article 22 is the one people miss. It restricts decisions based solely on automated processing that produce legal or similarly significant effects on a person. If your AI decides — not merely suggests — who gets a loan, a job interview, or an account suspension, you are in Article 22 territory and need a lawful basis, safeguards, and a route for the person to contest the outcome and get human review.

A workable DPIA for an AI use answers, plainly:

  • What personal data is processed, for what purpose, and on what lawful basis.
  • What could go wrong for the people involved — bias, error, opacity, scope creep.
  • How likely and how severe those harms are.
  • What measures reduce them, and what residual risk remains.

Write it before deployment, revisit it when the use changes, and keep it. The document is also your best evidence of accountability if anyone asks.

Human oversight that means something

“Human in the loop” is only a safeguard if the human can actually intervene. A reviewer clicking approve on a queue of two hundred AI decisions an hour is not oversight; it is rubber-stamping. Real oversight needs:

  • A person with the authority and competence to override the system.
  • Enough context and time to judge the output, not just acknowledge it.
  • A way to detect when the model drifts or starts failing in new ways.
  • A clear record of who decided what, and when.

Design the workflow so that doing the careful thing is the easy thing. If overriding the AI is slower and more annoying than accepting it, your oversight will quietly erode.

Vendor and model due diligence

Most AI risk now enters through suppliers. Before you rely on a model or tool, ask the unglamorous questions:

  • Where is data processed and stored, and is it used to train the provider’s models?
  • What are the data retention and deletion terms?
  • Is there a data processing agreement, and do international transfers have a valid mechanism?
  • What does the provider document about training data, known limitations, and bias testing?
  • What happens to your dependency if the provider changes terms or disappears?

Treat a model the way you would treat any subprocessor handling personal data — because that is what it is.

Data lineage

You cannot govern what you cannot trace. For higher-tier uses, keep a basic map of where input data comes from, how it flows through the system, where outputs go, and how long anything is kept. Lineage is what lets you answer a data subject request, investigate a bad output, or honour a deletion — without a week of archaeology.

A starting checklist

  • Maintain a living inventory of AI uses, each with an owner.
  • Assign every use a risk tier with simple, written criteria.
  • Flag anything that may be high-risk under the EU AI Act or Article 22.
  • Run a DPIA before launching high-risk or personal-data-heavy uses.
  • Define real human oversight for consequential decisions.
  • Run vendor and model due diligence; secure DPAs and transfer mechanisms.
  • Document data lineage for higher-tier systems.
  • Set a cadence to review the inventory and reassess tiers.

None of this requires a large team. It requires a habit: see it, tier it, document it, review it. Start with the inventory this week — everything else has somewhere to attach once you can see what you actually have.

This article is general information, not legal advice. For your specific obligations, consult qualified counsel or your data protection officer.