“Can I Share This?” Building UC Data Classifier (Gatekeeper), a RAG Agent for Everyday Data Questions at UC 

UC Tech Project: Can I Share This? with image of Beena Albert

By Beena Albert, Business Intelligence Tech Lead

Generative AI tools make it easy to summarize, clean, and analyze datasets in seconds. Because it’s so easy, it’s tempting to upload a spreadsheet without stopping to check if it holds sensitive information. However, once that file leaves UC boundaries, there’s no undo button.  

And it isn’t just about AI. In our day-to-day work, the same questions pop up all the time Can I put this dataset on SharePoint? Is it okay to email this spreadsheet over Outlook? 

To help staff, faculty, and students make those calls, UC has clear data classification policies. We also have great data and security experts who can weigh in when needed, but if you’re newer to UC, like me, or just trying to move fast on a deadline, translating complex policy documents into a practical decision for the dataset sitting in front of you can feel overwhelming. 

This is the problem I am trying to solve through Gatekeeper: a retrieval-augmented generation (RAG) agent in Microsoft Copilot Studio that turns UC policy into quick, practical guidance on how to handle and share your data safely. 

(About the name: Gatekeeper isn’t about policing people. Think of it more like a guide who knows the rules and helps you get your data onto an approved platform safely, cutting initial classification triage from hours of policy hunting down to under a minute.) 

The Problem: Bridging Policy and Everyday Decisions 

Under UC’s Classification of Information and IT Resources standard, institutional data falls into four protection levels: 

  1. P4 — High (Statutory): Data protected by strict law or regulation (e.g., SSNs, medical/HIPAA records, financial account numbers). Requires explicit approval and high-security storage. 
  2. P3 — Moderate (Proprietary): Sensitive institutional or personal data not covered by P4 statutes (e.g., performance reviews, or “toxic combinations” that re identify individuals). 
  3. P2 — Low (Internal): Non-public operational information meant for internal UC use only (e.g., internal memos, draft workflows, unpublished schedules). 
  4. P1 — Minimal (Public): Information intended for public release (e.g., published course catalogs, press releases, public web pages). 

Architecture 

Built with Agent Builder in Copilot Studio for Microsoft 365, Gatekeeper uses a RAG pipeline so it pulls authoritative UC policy text first before doing any reasoning. When you submit metadata, the agent runs it through OWASP defenses, retrieves relevant UC policies, classifies fields individually and in combination, passes high-risk or conflicting items through an evidence auditor, and generates the final assessment.  

To ensure privacy and reliability, Gatekeeper is built on five core guardrails:

  1. It works strictly with metadata, inspecting only field names and descriptions and never actual cell values, so it never ingests sensitive records.
  2. Policy always overrides general model knowledge, requiring every definitive rating to cite a specific UC policy document down to the section or page.
  3. It scores each column individually before evaluating how fields interact, flagging both immediate toxic combinations and risks that could emerge if extra columns are added.
  4. Fourth, when information is uncertain, Gatekeeper identifies that uncertainty rather than making a confident guess. It tags every field as Verified, Inference, Provisional, or Requires Review. If context remains missing after one quick follow-up question, it states its assumptions, defaults to the safer, higher protection level, and marks unknown row counts as Scale Risk: INDETERMINATE.
  5. Since user text, uploaded schemas, and web snippets can hide prompt injections, Gatekeeper treats all external content strictly as data rather than instructions, enforcing a clear hierarchy of trust that puts security controls and system instructions first, followed by UC policy, supporting references, and lastly, user requests. 

In Practice, Validation, and What’s Next 

When you share a list of column headers like Employee ID, name, department, job title, salary, home address, and performance rating (never the actual rows of data), Gatekeeper asks once for any missing context, runs an evidence audit on high-risk or conflicting items, and returns a breakdown covering Field/Column Name, Protection Level, Status, Plain-Language Rationale, UC Source, Legal Context, Confidence, Volume/Sensitivity, Scale Risk, Toxic Combinations, Required Handling, Escalation Office, and Next Step. Anything needing human sign-off routes straight to the right team: UISL, Privacy, Information Security, Registrar, Legal Counsel, or Export Control. 

I fully understand that I’m trying to solve a complex problem. Next, I am preparing the tool for penetration testing, working on memory refreshes, and exploring whether we can pair it with another RAG-based agent dedicated to tracking and pulling the latest UC policies.  

What I’ve learned most from this journey is how important it is to share ideas with colleagues; the feedback and encouragement I’ve received across UC have been instrumental in shaping this work. 

Have a tricky data sharing scenario you’d like to test? Reach out to Beena Albert. 

Author

Beena Albert

Beena Albert
UC Office of the President
Business Intelligence Technical Lead
Information Security Committee (HCM DS – ISC)

UC Tech Project Series  

View more UC Tech stories here.
Tell us about the tech project you or your team has completed, is working on, or is preparing to work on. Your story, if accepted for publication, will be shared out on the UC Tech News site, in the UC Tech News newsletter (room/timing permitting), and on the UC Tech News LinkedIn page. Please use the UC Tech News blog intake form to submit your UC Tech Project story: https://app.smartsheet.com/b/form/be650a2cf625471684bd94d18122d17b   

UC Tech News Content Survey

Do you visit the UC Tech News site regularly and/or get the monthly newsletter? If so, please take our short content survey: https://survey.alchemer.com/s3/9080340/2026-UC-Tech-News-Content-Survey