Financial aid document review
A financial aid office handles about 9,000 verification documents a term. An AI module extracts key fields and flags missing items. Staff approve every file.
Illustrative example. Figures are invented to show the method.
Outcome hypothesis
If we use AI field extraction in the verification workflow, we expect median time to a complete file to fall from 12 days to 6 because staff stop keying data by hand, provided key-field accuracy stays above 98% and every file still gets human sign-off.
What was measured
- Vendor module: $48,000 a year
- Integration: 160 IT hours
- Exception handling for documents the AI cannot read
- Exit cost: data export terms negotiated in the contract
- All documents routed through extraction
- Workflow penetration: 70%. Handwritten and non-English forms go to a manual queue
- Median days to complete file: 12 to 7
- Rework: rose from 4% to 6% in month one, then fell to 3%
- “Where is my aid?” inquiries down 20%
Lens gates
| Lens | Result | Finding |
|---|---|---|
| Quality | Pass | Key-field accuracy 98.6%. Lower on handwritten forms. |
| Capability and readiness | Pass | Staff trained on exception handling before launch. |
| Equity and accessibility | Fail | First-generation students land in the manual queue more often, so their files take longer. |
| Governance and risk | Pass | Data agreement and FERPA terms reviewed. Access logs on. |
| Evidence quality | Pass | Phased rollout: half the caseload first, compared against the other half. |
What can be claimed
Capacity redeployed: about 300 more counseling appointments a term, documented in the scheduling system. Avoided cost: $22,000 in temporary staffing.
Headcount savings. No positions were cut, and the module costs more than the avoided temp staffing.
Value is plausible, but the equity gate failed. Fix the manual-queue gap before expanding to other document types.
Re-check subgroup results at the end of next term.
The redesigned form alone captured about a third of the time gain. Without the non-AI comparator, the AI would have been credited for all of it.