By Maximilian Groth
Most AI transformation programs still treat access to data as a technical detail rather than a leadership decision. Max Groth, CEO and Co-Founder of Decentriq, argues that as AI systems gain broader access to organizational and partner data, leaders need to confront a harder question: who decided what AI is allowed to know?
Every AI transformation plan I’ve seen starts with the same question: what can this technology do for us? Faster reporting, sharper targeting, better forecasting, more responsive customer service. Rarely does the plan start with a harder question: what is this technology allowed to see, and who made that call?
That second question is turning out to be the more important one. As AI systems move from isolated pilots into core operations, they start touching data that used to sit behind separate walls: customer records, partner data, financial details, competitive intelligence. Someone has to decide where those walls stay up and where they come down. Too often, nobody actually does. The decision gets made by default, one integration and one dashboard at a time, until an AI system has visibility no single person ever explicitly approved.
Why Is Data Access Being Left to Engineers Instead of Leaders?
In most organizations, the question of what data an AI system can access is treated as an implementation detail, resolved in a sprint planning meeting instead of being raised in front of leadership at all. Someone needs the model to be useful, so they connect it to the systems that make it useful. Each connection is reasonable in isolation. None of them, individually, looks like a governance decision.
But that is exactly what they are. Every data source an AI system touches expands the boundary of what the organization is exposing (to that model, to whoever operates it, and often to any partner whose data now sits alongside it). When this boundary is set by technical convenience rather than deliberate choice, leadership ends up managing the consequences of decisions it never actually made.
This is not a new failure mode. Organizations have struggled with shadow IT and uncontrolled data sprawl for decades. What’s different with AI is the speed and the stakes. A traditional system that oversteps its access accumulates risk slowly, often invisibly. An AI system that oversteps its access can act on that visibility immediately: surfacing it in an output or a decision that a customer or partner sees.
What Happens When Nobody Can Say Who Approved What?
Consider a fairly ordinary scenario: a media agency deploys an AI planning tool to build unified audience strategies across its client roster. The tool works by drawing on campaign and audience data from every client account to identify patterns that improve targeting recommendations agency-wide. It does exactly what it was asked to do, and in doing so, it builds a view of one client’s audience and performance data that ends up informing recommendations for a direct competitor sitting in the same agency’s roster.
Nobody lied. Nobody broke a contract, necessarily. But if a client asks how its data is informing recommendations for a competitor on the same roster, or a regulator asks who approved this level of cross-account visibility, the honest answer in many organizations is: nobody, specifically. The access accumulated. It wasn’t authorized in any way that could be pointed to.
This is where “control” stops being an abstract governance term and becomes a leadership liability. If a company cannot trace which data informed a given AI output, it cannot answer for that output when it matters: to a customer, a partner, a regulator, or a board. The absence of a clear answer is itself the exposure.
Why Does This Get Harder as AI Gets More Capable?
It would be convenient if this problem stayed fixed once a company set some data access policies. It doesn’t, though. Because as models improve, they get better at inferring things they were never directly given. A system that only has access to campaign performance and demographic data can, with enough capability, start reconstructing patterns that look a lot like the sensitive information it was supposed to be kept away from. The boundary a leader thinks is fixed can expand on its own, simply because the model got better at connecting what it already had access to.
This is exactly why a static governance document falls short. The question has to be revisited continuously, and it has to be built into infrastructure rather than left to guidelines that assume good behavior downstream.
This is also where technical architecture and leadership judgment have to meet. Confidential computing and similar approaches exist precisely because “trust us with your data” is not a sustainable model for multi-party AI use cases (whether that’s several media partners pooling audience data, or a bank and a healthcare provider collaborating on a fraud model). The technology can enforce that a given party sees only what they’re supposed to see, and prove it. But choosing to build on that kind of infrastructure is a leadership decision that shouldn’t be left to procurement.
What Should Leaders Actually Do About It?
Rethinking this starts with a set of questions every leader deploying AI at scale should be able to answer, on demand, without convening a task force:
- Who has the authority to expand what an AI system can access, and is that decision logged anywhere?
- Can we demonstrate with evidence which data sources informed a specific AI output or recommendation?
- Are our data partnerships structured so that no single party, including us, ends up with default visibility into everything?
- Is “least privilege” a principle we actually apply to data, or just a phrase we use in security training?
- When something goes wrong, do we know who is accountable for the access that made it possible?
None of these questions require a data science background to answer. They require a leader willing to name who owns each data access decision, so it stops accumulating by default.
AI Transformation Will Be Won or Lost on Trust
What separates the organizations that get this right is leadership that can say, with confidence, exactly what their AI systems know and who decided that was acceptable. That clarity is what will let them move faster with partners and regulators, because rebuilding trust after a failure costs far more than the governance work it would have taken to earn it in the first place.
AI transformation has always been a leadership problem wearing a technology costume. Who controls what AI can know is the leadership problem that AI has finally made impossible to ignore.
Maximilian Groth
