Eli
What we solveHow it worksWhy EliSecurityTrust center
Log inOpen the demo

Built the way a financial aid office thinks.

Financial-aid work is federal work: every decision has a rule, a citation, and an auditor who may ask. Eli takes on the workflows your office already runs, FAFSA fraud and identity review, SAP, professional judgment, R2T4, and verification, and prepares each one the way an aid office needs: cited, logged, and routed to your staff when judgment is required. Eli resolves more of the repeatable work and keeps institutional decisions with the office.

The problem

Exception work is the bottleneck offices can't hire their way out of.

83%of aid offices say they need more automation, the most-cited need (NASFAA 2025)
~57%hit by conflicting-information ('C-code') resolution (NASFAA 2025)
56%of aid professionals say they're likely to leave the field (NASFAA 2025)

The government automated the easy half of this work; Eli takes the half that's left.

The FUTURE Act IRS data exchange (FA-DDX, live 2024–25) now imports most FAFSA tax data and treats it as verified, so the bulk of routine income verification is gone. What remains is stubbornly manual, and it is the work Eli takes on:

  • ›2026–27 FAFSA fraud intake, FAA Fraud Overrides, and V4/V5 identity reporting
  • ›SAP appeals, against your policy and the federal floor
  • ›Professional judgment, bespoke income-change recalculations
  • ›R2T4 returns, the withdrawal math with the funds returned in order
  • ›Conflicting information, a document contradicts the application
  • ›Verification, the routine income cases IRS matching cannot clear

Incumbents (Ellucian/StudentForms, Inceptia, ProVerify) mostly collect & route documents to humans. Eli works the whole exception queue, with a person on every judgment call.

A deterministic engine

Supported compliance calculations rest on unit-tested engines keyed to versioned rules: the $25 correction trigger and all-or-nothing submission, IRS "Per Computer" precedence, the Student Aid Index recomputation, the SAP standards (34 CFR 668.34), the R2T4 return order (668.22), and loan proration and caps. The demo labels directional SAI/Pell previews separately from final federal-table results and stops unsupported cases for review.

The LLM as bounded tools

Models handle bounded extraction, classification, summaries, explanations, and drafts within typed tasks. Every extracted value carries a source and confidence; supported compliance calculations remain in tested, versioned code, with staff review where policy or judgment is required.

A second pair of eyes, then yours

Identity rejects, V4/V5 outcomes, conflicts, SAP and PJ decisions, R2T4, and anything low-confidence route to a reviewer with a full, replayable audit trail. Eli does the legwork and drafts the decision; a person owns the judgment. The current demo only stages simulated or file-based action. Any production SIS or federal-system write requires a scoped integration and named authorization.

An audit-ready trail

Workflow and reviewer actions land in a replayable record: what Eli read, why it recommended what it did, the governing rule, and the person who made the call. Here is one case, start to finish.

A case is not finished at the answer.

Every aid workflow runs the same ten stages, and the calculation sits in the middle. A calculated answer is only the midpoint; a case closes when the posting is acknowledged, the student is told, and the money ties out. Eli tracks each case across the whole spine.

  1. Detect→
  2. Scope→
  3. Gather→
  4. Normalize→
  5. Calculate→
  6. Decide→
  7. Stage→
  8. Transmit→
  9. Acknowledge / reconcile→
  10. Communicate / retain

Calculate is stage five of ten. The five stages after it are why a case stays open.

Live interface, synthetic case

One case, through the engine.

A professional-judgment case: Eli reads the documents, recomputes the Student Aid Index, shows its confidence and its reasoning, then waits for an administrator to decide.

Professional judgmentEli reviewing
Sandra Boyd
2026–2027
Eli recommends

Recompute SAI with documented income loss, then route for a professional-judgment decision.

83%
  • ✓ Income loss documented (termination letter)
  • ✓ SAI recomputed: 4,200 → 0
  • → Awaiting an administrator's decision
A person makes the callApprove & correct
Measured accuracy

Accuracy we measure and publish.

Eli grades itself against a labeled corpus of known-truth documents. Most recent run (2026-06-09, 21 synthetic fixtures, 52 fields):

100%
field accuracy (n=52)
$0
dollar mean-abs error
0
hallucinated fields
2.8s
avg per document
These fixtures are synthetic documents, so the corpus measures extraction on clean and OCR-degraded renders, not scanned IRS originals. A scanned-document benchmark is the next step. Supported deterministic calculations are separately unit-tested; production quality must also be measured on contracted, representative data.

See it work on synthetic cases.

The live demo runs on 100% synthetic data across three institution types. Sign in as "Eli Operations" to see the full cross-institution picture.

Try a demo

100% synthetic data. No real student, IRS, or FAFSA records. FERPA role, FTI scope, and Title IV third-party-servicer status depend on the contracted work and require institution and counsel review.

Eli

The AI operations platform for financial aid offices. Software your team runs, or a managed service we run with you. A person signs every consequential decision.

What we solveHow it worksWhy EliSecurityTrust centerThe 5-minute proofThe 90-day pilotLog in to the demo
Questions? founders@tryeli.comSynthetic-data demo. No real student, IRS, or FAFSA records. © 2026 Eli.