It is very common for organizations to talk about AI governance in terms of AI compliance, but the two are very different things. An organization that has gone through an AI transformation and now considers its AI use “compliant” does not necessarily have good AI governance. AI compliance is a point in time. AI governance is continuous.
When we think of AI compliance, our focus is on demonstrating that the decisions made, or actions taken, at that point in time were true and were supported – a DPIA was completed, a vendor contract was signed, a policy was published. Each of these is evidence that a decision was justified then but not evidence of what the system is doing today.
AI governance asks a different question: is the system we reviewed six months ago still operating the way it was assessed to operate? Is the AI agent we approved and provisioned a month ago still performing in the way we approved it and with the same permissions?
An organization can pass every compliance checkpoint and still be indefensible the moment a regulator asks how the system operates today. This is the gap where legal defensibility sits.
Three Levels, One Gap
Most organizations sit at one of three levels of AI governance, and the distance between them is bigger than it looks from the inside.
- Good AI governance means the basics are in place. Policies and guidelines exist. Roles and responsibilities are defined. Risk assessments have been performed. Teams have received awareness training and AI use is monitored periodically. This reduces risk and establishes the foundations of governance. But where those activities are largely periodic, it can still provide little more than a snapshot of the organization’s governance posture at a particular moment.
- Great AI governance means that structure has been operationalized. There is a clear AI inventory and use case management, not just a policy that says one should exist. Controls are risk-based and aligned to the obligations that apply. Training is ongoing, not a one-time module. Monitoring is continuous rather than periodic. Documentation and evidence are maintained as the system changes, not just at launch. This builds trust and drives consistent outcomes but it still isn’t the same as defensibility.
- Defensible AI governance means you can prove it. Decisions are documented with clear rationale, not just recorded as having happened. Human oversight is defined, declared and evidenced, not assumed. There is end-to-end traceability from input to outcome. The evidence is audit-ready, not something that needs to be reconstructed when someone asks for it. The system is resilient, ethical and aligned with laws and standards as they evolve, not just as they stood at go-live.
Good governance reduces risk. Great governance improves performance. Defensible governance gives you the evidence to stand behind your decisions when they are challenged.
The Underlying Question
Ask any organization whether their AI is governed, and they’ll say yes. Ask that same organization to produce the evidence for one specific AI decision what system and version was in use, what data or inputs were relevant, who approved the use case, what controls applied, what monitoring was in place, who owned the outcome, and what evidence supported the decision. Their confidence starts to diminish, and that gap is almost always the gap between Great and Defensible because you can show that the policy exists, the inventory exists, but what you lack is proof that any of it held up for this one decision, on this one day.
What Does Defensive AI Governance Actually Require?
Getting from Great to Defensible isn’t about adding another policy. It’s about being able to evidence a review, a decision, a control and an action across the full pre- and post-deployment lifecycle. In practice, five things need to be true at once:
- A live inventory. You cannot defend what you cannot describe. A specific, current record of what each system does, what data feeds it, and what decisions it touches is needed.
- A documented legal and risk basis. A system-level mapping of the legal, regulatory and risk obligations that apply, such as GDPR Article 22 on automated decision-making, the Equality Act 2010’s indirect discrimination provisions, the EU AI Act’s risk-tiering. But a legal and risk basis that only checks the compliance box isn’t enough. Bias testing tells you whether the outcomes are skewed. Fairness asks a harder question, which is what was the process the system used to reach a decision – one you’d be comfortable explaining to the people it affects. A system can pass a bias test and still make decisions on proxies nobody would defend out loud.
- A sanctioned path to use it. A framework that only says no gets bypassed. An approval process that is long, cumbersome and onerous doesn’t stop AI adoption – it merely pushes it into shadow AI, where there is no visibility and no evidence at all. This is also where controls live – access restrictions, output monitoring, human-in-the-loop checkpoints where the decision actually matters. A control that exists on paper but has never been tested is a claim without actual evidence. Defensibility requires evidence that the control operates in practice. That means documentary, operational, technical and, where appropriate, independent assurance that it performs as intended.
- A named owner. Every AI system and approved use case needs clear accountability. Someone must be able to explain why the system was approved, the conditions attached to its use, who owns the relevant risks and controls, and who is responsible when those conditions change. Accountability cannot disappear into a committee.
- Ongoing monitoring. Models change. Use cases expand. Permissions change. Controls fail. Organizational needs and legal obligations evolve. Continuous monitoring should test whether the assumptions supporting the original approval still holds and whether the system continues to operate within its approved parameters. Scheduled reviews provide an additional safeguard, but material change should trigger reassessment when it happens – not simply at the next review date.
Gather Evidence as it Happens
If an organization is making defensibility its goal, they must remember that it is not a launch-day sign-off, filed away and produced only at the next audit. It is built through a continuous evidence trail across the governance lifecycle. Decisions, reviews, controls and actions should be evidenced when they happen, not reconstructed later. That evidence needs to show who made the decision, what informed it, what happened next and whether the assumptions and controls supporting it continued to hold. When circumstances change, the organization needs to be able to show how it responded.
That evidence only holds up if it was created at the time, clearly shows who was responsible, tells the full story and can be found when it’s needed. A record reconstructed after something goes wrong is very different from evidence created as the governance actually happened.
These five elements work together. An inventory creates visibility, clear legal and risk alignment establishes the basis for use, controls put that into practice, ownership creates accountability and ongoing monitoring ensures those decisions remain appropriate as things change. Defensibility then comes from being able to connect those elements and produce the evidence that shows how AI was governed in practice.
Start Here for Legally Defensible AI
Organizations do not need to implement everything at once. A practical starting point is to establish three things:
- A current AI inventory: identify the AI systems and use cases that are actually in use across the organization, including those that have not been formally approved.
- A named owner for each system or use case: establish who is responsible for its use, risks and ongoing oversight.
- A documented legal and risk basis for each use case: identify the legal, regulatory and organizational requirements that apply and document how they have been addressed.
These measures provide the foundation for implementing appropriate controls, monitoring AI use and reviewing decisions as systems and circumstances change.
