Is AI governance keeping pace with innovation, or are we kidding ourselves?

I asked this question on LinkedIn a few weeks ago and got six likes, two reposts, and not one comment. That’s not nothing, but it’s also not the “let’s hear real perspectives” conversation I was hoping to start. So I’m going to answer my own question first, in public, and leave the invitation open at the end — if you work in this space and think I’ve got it wrong, or only half right, tell me. I’ll update this post with your view, credited, rather than let it sit as one opinion pretending to be a survey.

In short: most organisations aren’t failing at AI governance because the frameworks are wrong. They’re failing because AI governance quietly assumes a mature identity and data governance foundation that most organisations — even good ones — don’t actually have yet. AI didn’t create that gap. It just started charging rent on it.

The frameworks aren’t the problem

NIST AI RMF, ISO/IEC 42001, the EU AI Act’s risk tiers — none of these are badly written. Read any of them and the substance is sound: know what you’re running, assign ownership, assess risk before and after deployment, keep evidence, give people a way to flag when something’s gone wrong. I built a readiness check around exactly that shape of question, and nobody who’s taken it has told me the questions feel unreasonable.

What I keep seeing in practice is organisations treating AI governance as a new, separate discipline — a policy to write, a tool to buy, a checklist to run — bolted on top of whatever already existed. And what already existed, in a lot of mid-market tenants I work in, is this: permissions that were never cleaned up, a SharePoint and OneDrive sharing model nobody’s audited in years, service accounts with more access than any human would be given, and no real inventory of what’s connected to what.

Copilot and agentic tools don’t create that problem. They index it, surface it, and act on it faster than a person would have. An AI agent that can read everything a user’s account can read isn’t a new risk category — it’s the old over-permissioning problem, now with a much faster reader attached.

What I actually see, walking into a tenant

Four things come up often enough that I’d call them the real gap, not the exception:

Nobody can say where the sensitive data actually is. Not in a “we have a DLP policy” sense — in a “if I ask you to name the five SharePoint sites an AI agent absolutely should not be able to touch, can you name them in under a minute” sense. Most can’t, and Purview sensitivity labels that exist only in a policy document don’t change that. A 2026 ShareGate survey of Microsoft 365 environments put a number on exactly this mismatch: 93% of organisations said they were confident their governance could support AI use responsibly, while 29% admitted AI tools had already surfaced sensitive data they shouldn’t have reached.

Agents and integrations don’t have an identity model. A human has a name, a manager, an access review, an offboarding process. A service principal or an agent connector frequently has none of that — it was set up once, given broad permissions because narrowing them was fiddly, and nobody’s looked at it since. Conditional Access policies get built around users and devices; the same rigour rarely gets applied to what’s effectively a non-human identity with its own reach into the tenant. Netwrix’s 2026 identity security research found a blunt version of the consequence: a 43% breach rate at organisations where AI expanded access, against 11% where it didn’t.

There’s no inventory. Not of models, not of use cases, not of which team is using which tool with which data. You can’t govern what you can’t name, and most organisations I’ve asked genuinely don’t have a list — not a bad list, no list.

Nobody’s told people when they’re actually supposed to check the output. Everyone agrees “human oversight” matters in the abstract. Far fewer organisations have said, concretely, which categories of AI-assisted output need a second pair of eyes before they go external, and who that second pair of eyes belongs to.

None of these four are AI problems. They’re identity and data governance problems that AI adoption is now making visible and urgent, because the systems touching that mess are faster and more far-reaching than the humans who used to be the bottleneck.

Why this matters more for mid-market than enterprise

Large enterprises usually have some of this foundation already — an identity team, a data governance function, people whose job is exactly the unglamorous plumbing work above. It’s mid-market organisations, the ones running Microsoft 365 Business Premium with Entra ID P1 and no dedicated security headcount, where I see the gap most sharply. They’re adopting Copilot and agent tooling at the same pace as everyone else, often genuinely enthusiastically, without the underlying identity and data hygiene that the frameworks quietly assume is already solved.

That’s not a reason to slow down adoption. It’s a reason to sequence it properly: fix the access model and know your data before you widen what can read it, not after.

The invitation, still open

I’d still like to hear from people actually doing this work — CISOs, AI governance leads, security specialists, anyone navigating this inside a real organisation rather than from a vendor deck. Tell me where I’ve got this wrong, or what I’m missing from a sector or scale I don’t see often. Comment, message me, or just reply to the original post — I’ll fold genuine perspectives into an update here, credited, because a roundtable beats a monologue every time I can actually make one happen.

And if you want a concrete starting point rather than another framework summary, the AI Governance Readiness Check on this site walks through the same four areas above in about five minutes, entirely in your browser, with nothing submitted anywhere.