Out of the box, Purview relies on regex and keyword matching — producing a 5:1 noise ratio on payment cards and 563:1 on passports. This only works on file types Purview can read: legacy formats like .doc, .xls, and .ppt, plain .csv files, and .msg email archives are not natively supported and are silently skipped. Trainable classifiers use machine learning but shift the entire burden to the customer: you must source 50–500 positive and 150–1,500 negative samples per pattern, run weeks of training cycles, and review results manually. Once a classifier is published it cannot be retrained — if it underperforms, you delete it and start over from scratch.

Purview is purpose-built for M365 and Azure. File servers, unstructured cloud storage, and data sources outside the Microsoft ecosystem receive limited or no coverage. Auto-labeling in SharePoint and OneDrive is largely restricted to modern Office formats — .docx, .pptx, .xlsx — with little to no support for .csv, .msg, .tmp, or list item attachments. Unstructured data outside that narrow set is exactly where classification risk is highest and Purview’s reach runs out.

Data Loss Prevention policies in Purview are only as strong as the classification underneath them. The benchmark demonstrates that out-of-box Purview misses the majority of sensitive files while simultaneously flooding security teams with false alerts. Files in archive containers like .zip and .rar are not inspected at all — their contents are invisible to DLP policy entirely. The result is both coverage gaps and alert fatigue at the same time.

Any time a classification pattern changes or your organization wants to look for something new, Purview requires a full re-scan of your entire data estate. There is no incremental or targeted update. In large environments this means weeks of scanning time and repeated cost — every single time.

Purview can enforce DLP on outbound email and apply labels in transit — but it does not classify or label Exchange mailbox data at rest. .msg and .eml files exported or archived outside Exchange are also unsupported for auto-labeling. Years of email containing sensitive information sitting in mailboxes or archive folders remains unclassified, untagged, and outside the scope of your data protection policies.

Base Purview licenses are just the starting point. OCR and trainable classifiers both require license uplift. Worse, OCR is metered — you pay per document. Combined with the requirement to re-scan every time a pattern changes, organizations end up paying OCR costs repeatedly for the same documents that haven’t changed. The cost of getting to accuracy is open-ended.