2024 – Present · Microsoft Purview

Purview DSPM &
AI Observability

Microsoft Purview · Data Security Posture Management for AI · Microsoft Ignite 2025

Co-Design Partnership Enterprise Security UX AI Agent Governance Data Visualisation Ignite 2025

Shipped Agent Overview page in Microsoft Purview for the Zava Procurement Agent, showing status, risk level, a sensitive activity trend chart, and recommended actions

Project Details

TL;DR

By 2024, Purview already had a lot of the signal it needed to police AI agents — Copilot instances, third-party bots, automated pipelines were touching sensitive data across the enterprise, and the underlying governance data was largely already being captured. What didn't exist yet was a way to turn that signal into real agent accountability.

I worked closely with another designer and our PMs to propose that governance experience, building directly off raw customer findings rather than a synthesized brief. Our first version was ambitious: full activity logs and audit trails, tracing every flagged interaction down to the individual step.

That direction landed well, but wasn't feasible to build at that scope. We pivoted — scoping down to detailed agent logs that flag the interactions actually worth a security admin's attention, paired with an in-depth Agent Details experience covering access, owner, connected tools, and risk.

The hardest part of the project wasn't the original idea, it was cutting real, valuable scope down to something buildable without losing what made it worth pursuing in the first place.

DSPM for AI shipped at Microsoft Ignite 2025, carrying that scoped-down direction forward.

Project Details

The Details

The Problem

By 2024, AI agents — Copilot instances, third-party bots, automated pipelines — were spreading through enterprise systems faster than security teams could track them. Purview's data governance foundation was already strong, and much of the raw signal needed to govern these agents was already being captured. What was missing was a way to turn that signal into accountability: when an agent overshared a file or quietly pulled data through a connected tool, security teams had almost no way to reconstruct what actually happened.

Wide planning board mapping out Build view versus Runtime views, IRM alert structure decisions, and references including a competitor walkthrough and threat graph explorations — Structural mapping, early on — deciding how Build view and Runtime should relate, and reviewing reference material (including a competitor tool walkthrough) before committing to a direction.

Research

Our PMs brought us in early, working directly off raw customer findings rather than a finished brief — pain points were still being compiled as we started converting them into concepts. We paired that with secondary research into how AI agents were actually being adopted across enterprises, and ran it against live customer feedback sessions to check the risks we were designing for matched what teams were actually hitting in production. Close, early collaboration between PMs and design is what let us move from raw findings to a real proposal quickly.

Insight

Purview already had most of the underlying data. The opportunity was building real governance tooling around it — full activity logs and audit trails, so nothing about an agent's behavior needed to be reconstructed after the fact. We proposed that version first: comprehensive, and built to hold up under real security scrutiny.

Multiple iterations of the agent connection graph and side drawer pattern, showing different states including a populated agent map, a user detail panel, and a reviewer agent detail state — Iterating on the graph-and-drawer pattern — testing how the agent connection map (users, knowledge sources, tools, agents) paired with a drill-down drawer across several states and data densities.

Decisions

That full activity-log-and-audit-trail proposal landed well — but wasn't feasible to build at that scope in the time we had. Rather than shelve the ambition, we scoped it down: instead of an exhaustive log of everything an agent did, we designed detailed agent logs that specifically flag the interactions worth a security admin's attention, paired with an in-depth Agent Details experience covering access scope, owner, connected tools, and risk.

That scoping decision was the hardest part of the project. It meant deciding, feature by feature, what could be cut without cutting the thing that made the idea worth building — and then redesigning around that smaller footprint so it still felt like the best possible version of the experience, not a compromise.

We kept alerts scoped to a specific agent rather than grouped generically by alert type, so an admin investigating one agent could see everything relevant to it in one place — consistent with the smaller, agent-centered shape the project ended up taking.

Outcome

What shipped is the scoped-down version: every agent gets a dedicated page — status, access scope, owner, connected knowledge sources and tools, risk level, and a sensitive-activity trend over time — plus a detailed log that flags which specific interactions need a security admin's attention, short of the full exhaustive audit trail we'd originally proposed. It's a smaller footprint than our first pitch, but our usability studies proved that seconds are enough for security admins to clearly answer their question of “does this agent need attention right now” without cross-referencing three other tools.

Impact

DSPM for AI was announced at Microsoft Ignite 2025, positioning Purview to govern AI agents alongside traditional data assets. It was a successful launch, garnering a strong positive reception. The scoped-down direction — detailed, flagged agent logs and a full Agent Details experience — carried through into what shipped, even though the more ambitious original proposal didn't make it into this round.

1 / 6

AI-Powered Privacy Manager →

Purview DSPM &AI Observability