Thoughts on Microsoft Purview News from Microsoft Ignite 2026

Introduction
Another Microsoft Ignite, another avalanche of announcements, previews, new capabilities and enough AI-related terminology to keep us all busy for the foreseeable future.
And, as expected, Microsoft Purview has had some attention this year .
We've seen announcements covering data classification, auto-labelling, Data Loss Prevention (DLP), Copilot protection and security controls for local AI agents. Some are genuinely exciting, others are long-overdue improvements, and quite a few have already been on the roadmap or available in some form of preview.
Personally, I don't think the fact that some of these announcementts aren't entirely new is necessarily a bad thing. Having worked with Microsoft Purview for years, helping organisations implement and improve their data security posture, I've learned that the most valuable improvements aren't always the ones that make the biggest headlines. Sometimes, it's the little things that make our lives as consultants and administrators significantly easier. And sometimes, it's finally addressing limitations we've been working around for years.
Let's dive in!
Table of contents
TL;DR
Microsoft Ignite 2026 highlighted a continued shift in Microsoft Purview towards more accurate and scalable data classification, automated information protection, and stronger security controls for AI-driven workflows. Key announcements include improved built-in sensitive information types (SITs), recurring classification of historical data, a fivefold increase in auto-labelling capacity, and the ability to apply default sensitivity labels to existing SharePoint content. DLP capabilities are also expanding to protect sensitive information accessed by Microsoft 365 Copilot and supported local AI agents, with greater visibility into agent activity and associated risks. On the governance side, Microsoft is consolidating its data governance capabilities into Microsoft Fabric and OneLake. Overall, the direction is towards more automated, content-aware and scalable data security, with protection extending beyond traditional user interactions to AI agents and the information they access.
If you're looking for the full rundown, you can find the Microsoft Ignite 2026 Book of News here. Microsoft's main blog post covering the Purview announcements is available here, alongside a separate update on Microsoft's data governance strategy.
Microsoft Purview news from Microsoft Ignite 2026
There's quite a bit to unpack this year, so here's a condensed overview of the main changes.
Data classification and information protection
Auto-labelling improvements (generally available) – Daily processing capacity has increased from 100,000 to 500,000 files per tenant, simulations can match up to 20 million items, and adaptive scopes allow SharePoint sites to be targeted dynamically.
Default sensitivity labels for existing SharePoint files (public preview) – Library-level default labels can now be applied to existing data at rest, without requiring users to open or edit the files.
Unified classification management (public preview) – A consolidated experience for managing SITs (sensitive information types), exact data match, trainable classifiers and document fingerprinting, with improved searching and filtering.
More accurate built-in SITs (starting January 2027) – Microsoft is improving detection accuracy, starting with more than 100 commonly used SITs, including credit card and US Social Security numbers. The aim is better detection out of the box, with less manual fine-tuning.
Recurring classification refresh (public preview: January 2027) – Purview will be able to reclassify historical content when built-in or custom sensitive information types (SITs) are updated, without requiring users to open or modify the files.
DLP, Copilot and AI agents
Agent risk assessment (public preview) – Purview signals about agent activity and interactions with sensitive information will contribute to the Unified Agent Risk score in Agent 365.
DLP for sensitive content used by Copilot (public preview: January 2027) – Microsoft 365 Copilot can be prevented from processing files containing sensitive information, even when those files aren't labelled. This also extends to files attached to prompts or referenced through links.
DLP for local AI agents (public preview: January 2027) – Purview is extending protection to supported local agents, including GitHub Copilot and Claude tools, with visibility into agent activity and controls for sensitive information accessed through prompts, files and tools. Policies can also be scoped to specific agents.
Data governance: major change of direction
Data governance is moving towards Microsoft Fabric – Microsoft is retiring the standalone Purview Data Governance experience and consolidating its data governance strategy around Fabric and the OneLake catalog. Support for the existing Purview Data Governance capabilities is scheduled to end on 12 November 2027. Importantly, this doesn't affect the continued investment in Purview's data security and compliance capabilities.
Quite a few changes, right?
Now, here's what I actually think about them.
Welka's thoughts
Classification
For me, accurate classification is everything. It's one of the most important parts of any successful Purview deployment because it serves as the foundation for so many other capabilities. Classification helps us discover where sensitive information resides, drives Data Loss Prevention (DLP) and Insider Risk Management policies, supports retention policies, and enables the automatic application of sensitivity and retention labels. It's also what allows organisations to scale information protection without relying entirely on users to classify documents manually. Therefore, if your classifiers aren't accurate, everything built on top of them becomes less reliable.
I've seen how much time can go into creating custom SITs and trainable classifiers, testing them against real business data, reviewing false positives and continuously fine-tuning detection logic. That's why I'm particularly excited about Microsoft's improvements to built-in SIT accuracy.
Getting classification right should be a priority for every organisation implementing Purview. It's not just about finding more sensitive information, but finding the right information consistently, so the controls relying on it can actually do their job.
Historical classification
The improvements to historical classification are equally exciting. One of the biggest challenges in Purview assessments is getting an accurate picture of sensitive information across an organisation's entire environment, including years of SharePoint and OneDrive content that nobody has touched in a long time.With recurring classification refresh, improvements to built-in and custom SITs can be reflected in historical classification results without requiring users to modify existing documents.
For me, the real value is being able to build a more accurate picture of where sensitive information resides and how much of it exists across the estate, rather than relying primarily on recently processed content. I'd love to see this evolve into better reporting and comparisons over time, particularly across built-in and custom SITs and trainable classifiers. Being able to distinguish newly discovered sensitive information from improved detection accuracy would make a real difference during data security assessments.
Ultimately, classification shouldn't be a one-off implementation exercise. It needs to evolve alongside the organisation, its data and the risks we're trying to manage.
Auto-labelling for sensitivity labels
The auto-labelling improvements are a welcome addition, particularly for organisations dealing with millions of documents across SharePoint and OneDrive. A fivefold increase in daily processing capacity (from 100,000 to 500,000 files), larger simulations (20 million items), and adaptive scopes should make scaling information protection considerably easier.
But faster auto-labelling is only beneficial if the classification behind it is accurate. Otherwise, we're simply scaling our mistakes. The last thing we want is thousands of documents incorrectly labelled, potentially with encryption that disrupts legitimate access.
That's why I think the increased simulation capacity is just as valuable as the processing improvements. Proper testing and gradual deployment remain essential.
Another improvement I'm particularly pleased about is the ability to apply default sensitivity labels to existing files in SharePoint libraries.
Previously, library defaults applied to new or modified documents, leaving us to figure out how to label everything already sitting there. This often meant scripts, Power Automate flows or some other creative jiggery-pokery.
Being able to label existing content without users having to open or modify files is a long-overdue improvement that should make baseline protection much easier to implement.
There are still preview limitations to consider, but it's one less workaround to worry about, and I'll happily take that.
Data Loss Prevention and local AI agents
Another announcement that caught my attention is the expansion of Purview DLP to local AI agents.
Let's face it, these days it's hard to find someone in technology who hasn't used GitHub Copilot or another AI coding assistant to write scripts, troubleshoot issues or simply vibe code something together.
Whilst I can appreciate the productivity benefits, I've always been concerned about the limited visibility and control over what data these tools can access and process. We're no longer talking about someone simply copying and pasting information into a chatbot. AI agents can interact with local files, source code, project directories and other tools, potentially accessing credentials, customer information or confidential business data. Just because a developer has legitimate access to sensitive information doesn't mean an AI agent should automatically be able to process or share it.
That's why I like Microsoft's direction here. Rather than simply allowing or blocking AI tools, Purview is moving towards controlling what sensitive information supported agents can interact with. Of course, I'll reserve judgement until I've had a chance to test it properly. I'm particularly interested in how well DLP handles interactions between agents, local files and external tools, and how much visibility administrators will have into what's being accessed or blocked.
There's plenty of potential here, but the real test will be whether we can enforce meaningful data security controls without unnecessarily disrupting legitimate development workflows.
Copilot DLP
The improvements to Microsoft 365 Copilot DLP are another welcome addition. The reality is that most organisations have plenty of unlabelled or incorrectly labelled documents, and sensitive information doesn't stop being sensitive just because someone forgot to label it.
Being able to prevent Copilot from processing files based on their actual content, rather than relying solely on sensitivity labels, adds an important layer of protection. Of course, this brings us straight back to the importance of accurate classification.
It's also worth remembering that DLP isn't a substitute for good information governance. Preventing Copilot from processing sensitive information doesn't fix the underlying permissions that allowed access to it in the first place. Both need attention.
What I'll be keeping an eye on over the coming months
Announcements are one thing, but seeing how these features perform in real environments, with real business data, is another.
I'm particularly interested in testing the accuracy of the enhanced SITs, how effectively recurring classification refreshes improve visibility into historical data, and whether the increased auto-labelling capacity makes large-scale deployments easier to manage.
For local AI agents, I'll be looking at how well DLP controls work in realistic scenarios, the visibility they provide into agent activity, and whether the new risk signals are genuinely actionable.
And, as always, there's the less exciting but equally important stuff: licensing, supported workloads, reporting and operational overhead. These are often the things that determine whether a feature is genuinely practical to implement at scale.
Conclusion
Purview has come a long way in recent years. It's certainly not perfect, but the improvements announced at Ignite show that Microsoft is addressing some of the practical challenges we've been dealing with for years, whilst also adapting to the growing adoption of AI.
What I particularly like is the continued investment in the fundamentals. As exciting as Copilot and AI agents are, effective data security still starts with understanding what information we have, where it lives and how sensitive it is.
For me, that's the biggest takeaway from Ignite 2026. Better classification, more scalable protection and stronger controls for AI-driven workflows are all steps in the right direction.



