Microsoft Just Governed 82 AI Agents Per Employee. Nobody Asked What Any of Them Know.
Microsoft's Agent 365 solves enterprise AI governance: identity, access, observability. It doesn't solve knowledge accuracy. That gap is still wide open.
For every person in your company, 82 AI agents are running on your network — most with elevated access privileges. That's not a projection. It's Microsoft's own security data. This week the company launched Agent 365: a centralized control plane that gives IT teams visibility and governance over those agents. What they're doing, who they belong to, whether they're permitted to act. It's a real product that solves a real problem. It answers nearly every question about enterprise AI governance except the one that matters most: what do those 82 agents actually know?
The governance sprint is over. Governance won.
Vasu Jakkal, Microsoft's Corporate Vice President for Security, described Agent 365 as "air traffic control for AI agents." The analogy is good. ATC knows every plane in the air — its flight path, its pilot, its clearance level. That's what Agent 365 does. Cross-silo visibility, security categorization, access control, all wrapped into the new Microsoft 365 E7 bundle at $99/user/month.
The enterprise AI governance conversation matured fast. A year ago nobody had a product for this. Now the world's largest enterprise software company has shipped one, and they're bundling it with Copilot and E5-level security in a single SKU.
But here's the thing about ATC: it knows every plane is in the air. It doesn't check whether the cockpit charts are out of date.
Agent 365 governs the agents. It doesn't govern what the agents know. And in most enterprise environments, what the agents know is wrong.
The enterprise AI stack in 2026 has four layers. Three are covered.
Agent 365 didn't launch alone. In a 72-hour window this week, four other enterprise AI governance products came out roughly simultaneously: AvePoint's AgentPulse Command Center (shadow AI discovery), OneTrust's expanded AI governance platform (real-time guardrails), Netskope's One AI Security (BYOAI prevention), and now Microsoft's Agent 365.
The market just placed its bets. The enterprise AI governance layer is mature, competitive, and backed by major vendors. That sprint has been extraordinary to watch.
Look at the full stack:
- Identity layer — Who are the agents, who created them, what accounts do they run under → Agent 365, Imprivata
- Access layer — What can they do, what data can they reach, what actions are permitted → Agent 365, OneTrust, Netskope
- Model layer — Which AI model are they running on, which vendor, which version → GPT-4, Claude, Gemini
- Knowledge layer — What documents do they read, are those documents current, accurate, consistent → ???
The first three layers have products. Competitive products, in fact — multiple vendors fighting for the same budget. The fourth layer has nothing at Microsoft's scale. It barely has anything at any scale.
Your ME7-licensed organization gets governed agents with clean identities and correct permissions. Those governed agents are still reading the SharePoint pages from 2022. The pricing spreadsheet someone last touched before the reorg. The HR policy doc that contradicts the one uploaded six months later. The Copilot integration didn't fix any of that. Agent 365 doesn't either.
The bundle ensures compliance. It doesn't ensure accuracy.
Microsoft's position is stranger than it looks
Here's the uncomfortable part: who's sitting right on top of this gap.
Microsoft owns the entire document lifecycle. Word and SharePoint create the documents. OneDrive stores them. Teams hosts the conversations where knowledge lives informally. Copilot reads all of it. Agent 365 governs the agents that act on everything above. No other vendor has that view of enterprise document health.
If any company could ship an accuracy layer — something that actively detects stale content, flags contradictions across document sets, identifies where knowledge has drifted from reality — it's Microsoft. They have the signals. They see when documents were last updated. They know which Copilot answers triggered negative feedback. They understand which sections of SharePoint nobody visits anymore.
They haven't built it. That's a product gap, not a technical limitation. Whether the priority lands internally or a specialist platform fills it first, the market question is the same: who ships the layer that makes governed agents actually reliable?
82 agents at scale is a lot of wrong answers
The enterprise AI governance conversation happened faster than anyone expected, and the products that came out this week are genuinely useful. Agent 365 solves a real problem that most enterprises have been struggling with.
But identity and accuracy aren't the same problem, and the industry hasn't stopped long enough to notice. We now have robust infrastructure for knowing which agents exist, what they can access, and whether they have permission to act. We have essentially nothing for knowing whether the information those agents are acting on is correct.
Eighty-two agents per employee is an enormous multiplier. Every piece of outdated policy, every stale pricing sheet, every contradictory procedure document — it gets amplified by that factor across every interaction those agents have. Governance makes the problem visible. It doesn't make the problem smaller.
The next conversation in enterprise AI is already overdue.