Your DSPM dashboard is green. Your Copilot is still useless.

Open your Data Security Posture Management dashboard. If it looks like most regulated tenants I see, the oversharing flags are clear, the risky-prompt count is low, nothing is exfiltrating. By its own scoreboard, your AI estate is healthy.

But your Copilot still can't find anything useful.

Because the failure that's actually crippling it is the one DSPM was never built to show you.

What DSPM is

Microsoft Purview DSPM is the layer where what should the AI be able to reach? turns into what did it reach? — the observability surface over every AI app and agent in your tenant. One note before you go looking: as of mid-2026 the standalone "DSPM for AI" blade is folding into a single unified DSPM, currently in preview, with the older products relabelled "(classic)".

The three things it shows you

1. AI observability is an inventory of every AI app and agent active in your tenant — including Microsoft Agent 365 agents — flagged by how many are high-risk and how many have touched sensitive data.
2. Activity explorer's AI activities tab shows the prompts and responses themselves: whether an interaction involved sensitive information, and whether a data loss prevention rule fired during it.
3. Data risk assessments scan your hundred busiest SharePoint sites weekly and break each one down across Overview, Identify, Protect, and Monitor tabs — including how many items are shared with anyone, shared org-wide, or shared externally.

Open all three and you can see what your AI reached, and how exposed it got reaching it.

Now notice what all three have in common

Every one of them points the same direction.

The assessments hunt for data reachable too widely. Observability flags exfiltration, risky prompts, unusual access. The Protect tab's remediation actions — restrict access by label, restrict whole sites with SharePoint Restricted Content Discovery, auto-label unlabelled files, auto-delete stale content — every one of them removes reach. The instrument is calibrated end to end for a single failure: the AI touching more than it should.

The failure it can't see

Over-restriction runs the other way, and it trips no sensor.

When an agent can't reach the document it needed, nothing happens. No incident, no alert, no row in a risk report. The agent just returns less, the user shrugs, and "the AI wasn't helpful" never gets escalated as anything. A leak is loud — someone sees what they shouldn't and files a ticket. A lockout is silent. DSPM is built to hear the loud failure. The quiet one passes straight through it.

So a green dashboard is not the same as a healthy estate. Green means nothing is over-reaching. In a cautious, locked-down tenant — exactly the regulated ones where Copilot most often disappoints — over-reach was never the problem. The problem is that the AI can barely reach anything, and the dashboard has no opinion about that.

How to read DSPM for the silence

The data you need is sitting right there. You just have to read it against the grain.

In AI observability, look for agents that hold identities but show almost no meaningful activity — provisioned, then silent. In Activity explorer, look at the interactions that returned nothing useful, not only the ones that returned something sensitive. Across your data risk assessments, compare the access you deliberately provisioned against what ever actually got used; that gap is your over-restriction, measured. None of this is a built-in report. It's you reading the same telemetry for absence instead of excess.

What to do when you find it

Over-restriction gets fixed upstream, in label and rights design — by separating what the information is from who is allowed to reach it, so routine content sits at a default the AI can use instead of being locked away by reflex. DSPM's job here is diagnosis, not cure. It shows you the silence is there. The rights model is where you make it stop.

Go look

Open DSPM in your own tenant this week — and do the harder read. Anyone can confirm the oversharing flags are clear; that's the half the dashboard does for you. Open AI observability and find the agents that exist but barely act. Find the sites your AI never reaches into. The green dashboard is telling you about one failure.

Go looking for the one it can't.