FERPA Risk Audit: Map an AI Workflow
An AI workflow is only as compliant as its weakest link. In this lab you'll take a real-looking agent workflow, decide whether each tool may even hold student data, then map its FERPA exposure and attach a control to every step — a risk matrix you could hand your department.
Work through it
Check each box as you go — your progress is saved on this device.
The scenario
Maple Valley College pilots an "Advising Assistant" built from an agentic browser:
- Node A — logs into the SIS and reads each advisee's transcript and contact info.
- Node B — combines GPA, major, and directory info (name, email) into one profile.
- Node C — drafts a "stay or switch majors" recommendation for each student.
- Node D — emails the recommendation directly to the student.
Important: it runs on a faculty member's personal AI account — no institutional agreement.
Your risk matrix
Fill in each row. Saved on this device until you reset.
| Node | FERPA exposure | Risk (L/M/H) | Mitigation control |
|---|---|---|---|
| A — reads SIS records into a personal AI account | |||
| B — combines GPA + directory info | |||
| C — drafts a major recommendation | |||
| D — emails the recommendation to the student |
Saved automatically on this device.
Reveal model answer
| Node | Exposure | Risk | Control |
|---|---|---|---|
| A — personal account | Third-party disclosure; passive reading | High | STOP — tool not covered by an agreement. Use an approved tool with a DPA, or de-identify. Least privilege. |
| B — combines data | Recombination into a protected composite | Medium | Don't merge directory info with protected records unless authorized. |
| C — drafts recommendation | Autonomous record creation | Medium | Treat the draft as an education record; don't finalize without review. |
| D — emails student | Decision shaping | High | A human advisor reviews and approves before it reaches the student. |
Key takeaway: the whole workflow fails at Node A — an uncovered personal account is the disqualifying problem. Fix the tool first; the node-by-node controls matter only once the tool is allowed to hold student data.