When I speak with organisations about AI deployment, the conversation usually follows a familiar pattern.
It starts with genuine enthusiasm about what the technology can do and how easily users can now access information, but before long a difficult question tends to surface: have we actually thought about what data it serves and to whom?
In most cases, the honest answer is no, not fully. And that’s a problem worth taking seriously, because the gap between how ready an organisation’s data is for AI versus the reality is where a significant amount of risk lives.
In this blog, we explore why AI safety is ultimately a data governance issue, and how gaps in access, accountability and control can create risk long before organisations realise it.
Most organisations have spent years tightening controls around their structured data. Information stored in core systems and applications is typically governed through defined roles, permissions and review processes, and for the most part those controls do what they are supposed to do.
Documents are a different story. Permissions might be configured at a folder level, but the logic of who should and shouldn’t have access to a given piece of information doesn’t automatically translate to an AI environment. A M365 Copilot query isn’t made in the context of the person asking the question, rather it can operate with different access than that individual would have themselves. The boundaries that existed before are no longer the boundaries that apply.
If data isn’t AI-ready, here’s what can happen: an employee runs a Copilot query and receives information they shouldn’t have access to. In many cases, the person will likely flag this as an issue and a remediation process occurs. However, what if the information was surfaced to someone with not-so-honourable intentions? Or the access breach is missed, and the data is unwittingly leaked via another channel? These scenarios are not only plausible, but very real risks.
Remember when we used to talk about shadow IT, where employees would spin up their own spreadsheets or Access databases because the sanctioned tools weren’t meeting their needs? Well, the same pattern is now playing out with AI. People are bypassing M365 Copilot by running their own queries in non-sanctioned consumer LLMs, feeding in information that no one in the organisation has assessed or approved.
I’ve seen organisations where leadership has permitted this with the reassurance that “we’re just not putting any client data in there, so it’ll be fine.” Unfortunately, these organisations often find out the hard way that this can still result in a data breach, a compliance finding, or an AI-generated output that’s embarrassingly or dangerously wrong. These missteps can come with very real consequences in the form of fines and reputational damage.
So, if AI gives someone access to information they shouldn’t have, who’s responsible?
The chain of accountability gets blurry with AI. If a company hasn’t put in the right guardrails to prevent improper data access, then the organisation carries the risk. Adding AI to a data estate without a proper governance framework means your organisation is, in effect, guilty through omission.
This will become more an issue as AI becomes more deeply embedded in how decisions are made.
When I start talking about data governance, compliance and metadata, people’s eyes start to glaze over. So, when I consult with organisations as to what rigorous data governance looks like, I brieak it down to three things:
1. A conscious, structured decision about what your AI environment can access and and why. Not an assumption or default setting, but a deliberate set of boundaries with accountability behind them. That means clearly defined data ownership, access procedures and privacy rules that have been thought through within the context of AI, not just the systems that existed before it.
2. Standard data governance practices that are being followed. Roles and responsibilities assigned. Data owners and stewards who understand what they’re responsible for. A common understanding of how information should be handled, shared and protected.
3. Getting people on side. Governance frameworks only work if the people inside an organisation understand their role in maintaining them. That includes technical teams but goes well beyond them. Everyone who uses AI is a participant in the governance model. Training and clear expectation-setting are a part of the framework, not optional extras.
The AI value proposition also needs to be defined up front. Putting good data governance foundations in place won’t produce an ROI that justifies the investment by itself, but without those foundations, the AI you’re deploying will produce lower-quality outputs, create higher risk and generate costs that dwarf the cost of getting governance right in the first place.
The organisations that approach AI as a business transformation, rather than a technology rollout, tend to get this. They understand that the value they’re trying to create must be identified at the start, not retrofitted when the problems emerge.
Ready to DELETE uncertainty from AI? Discover where your sensitive data is exposed, identify potential risks, and build a roadmap for secure AI adoption. Book a Data Security Envisioning Workshop with a Data#3 Microsoft specialist and take the first step towards confidently securing your AI journey.
Contact a Data#3 Microsoft specialist today to take the first step towards confidently securing your AI journey.
Information provided within this form will be handled in accordance with our privacy statement.