AI DLP for screenshots: verify the OCR step

To evaluate DLP for screenshots, verify that optical character recognition runs on the exact delivery path before interpreting a classifier result. Then test whether the extracted text supports the expected match and whether the policy controls the eventual upload. Activation, extraction, classification and delivery are separate evidence stages.

Microsoft Purview's OCR documentation provides a useful worked scope. This guide reviews those documented conditions and proposes synthetic acceptance tests. No tenant was tested, and coverage of every AI client, pasted image or Copilot attachment remains unestablished.

Check activation and scope first

Purview OCR is optional and must be enabled at tenant level. Microsoft describes an Azure pay-as-you-go subscription, Syntex billing configuration, and an administrator selecting OCR locations and included or excluded groups. These prerequisites deserve explicit confirmation before a screenshot enters an evaluation.

The documentation says settings generally take effect about an hour after activation and only images uploaded after enablement are scanned. An old file producing no finding therefore gives little evidence about a newly enabled configuration. Upload a fresh, deliberately fictional test image through the chosen route.

Record the location, group membership, policy version and activation time. Keep any billing decision with the authorised account owner. A documentation review or a security pilot brief supplies no authority to enable a chargeable feature.

The broader workflow still needs an explicit classification and policy checkpoint:

Use the limit for the correct location

Microsoft lists different supported file types and limits by location. The OCR page states a 20 MB maximum for Exchange and Teams, and 50 MB for SharePoint, OneDrive and endpoints. Its image-resolution range is 50 × 50 to 16,000 × 16,000 pixels. Extraction covers the first two million characters.

Exchange also has a limit of 20 embedded images scanned per supported file. Endpoint OCR excludes paths already excluded from Endpoint DLP. The documented default endpoint bandwidth limit is 1,024 MB per device per day, after which image scanning stops.

These conditions apply to the locations Microsoft lists. Check each browser upload path against its own documentation. The separate Endpoint DLP documentation lists JPEG, PNG, TIFF and BMP extensions for OCR monitoring. Its image list and the OCR page differ on endpoint image-only PDF coverage. Confirm that case before treating it as supported.

Follow the actual screenshot route

Uploading an image file, pasting image data, attaching it inside an AI application and placing it in a document are different actions. Write one evaluation row per action, including the browser or client and the destination account.

Edge for Business inline DLP documents its own activation and file-size conditions. That page does not establish OCR coverage for screenshot uploads. The Purview OCR location list alone also leaves pasted images and particular Copilot prompt-image paths unresolved.

Ask the supplier to identify the enforcement component for each row and provide supporting documentation or a controlled demonstration. Preserve an explicit unknown where the evidence stops. Use the AI DLP comparison guide to keep that route-level question visible during selection.

Make extraction errors observable

Microsoft's Azure OCR limitations guidance identifies resolution, contrast, rotation, lighting and text density as accuracy factors. It recommends evaluating the implemented solution. That page describes Azure OCR; it supplies no Purview-specific accuracy guarantee or confirmation that Purview exposes the same confidence values.

For a controlled test, render a short fictional reference record into an image and keep a plain-text version. Use a classifier whose expected result is already established for that test record. Include a harmless near-match that should pass without a sensitive-data finding.

Change one image property at a time: text size, contrast, rotation or compression. Record the output expected from each case before running it. Use the languages and document styles your approved workflow actually encounters. Hold handwriting-support claims until the relevant product documentation and test result agree.

Verify the match and the destination

Retain separate fields for in-scope status, scan completion, classifier result, policy action and observed delivery. A missing alert alone cannot distinguish an excluded image, a failed extraction and a scanned image that contained no match. Record unavailable scan evidence as unavailable.

Use a controlled destination and synthetic content to check what actually arrives. Compare the matched image with the harmless near-match, and test boundary sizes only within an authorised environment. A warning, a logged finding and a prevented upload have different observable outcomes.

Give the employee a supported way to complete legitimate work when the image cannot be evaluated. Document that path and its limits. Record unknown scans explicitly as unevaluated.

Finally, keep image classification separate from the sensitivity-label checks on generated output. Both may matter to the same task, but each needs its own evidence. Bring the route matrix and paired test images to a Promptective demo to discuss a bounded evaluation.