90-day pilot

Pilot command center

Before any live pilot, Eli replays a school's historical cases with no live effect, compares its disposition to what the office did, and has to pass every product-quality gate. Those gates are evaluated separately from procurement, security, contract, identity, and data-access approval. This illustration runs over 182 synthetic cases in scope (the modeled pilot volume is 400).

It starts with a 45-minute workflow interview

Bring one workflow and one de-identified case. No data leaves your office for this meeting.

We walk the workflow the way your office actually runs it, seven questions in 45 minutes. The meeting has one goal: prove Eli understands the work, reproduces your known result, and produces the defense file. That proof is what earns data access for a pilot.

  1. 1Trigger5 min

    What exact event, file, or report tells you there is work? Who notices first? What can be missed?

  2. 2Population and clock5 min

    How do you find every affected record? Which date starts which deadline? What report proves completeness?

  3. 3Tools5 min

    The real SIS screens, portals, spreadsheets, and federal sites. Which one is authoritative at each stage?

  4. 4Happy path10 min

    One real de-identified case, first event to final close: every click, handoff, wait, calculation, and message.

  5. 5Exceptions8 min

    The three cases that consume disproportionate time. What missing or conflicting fact stops the work? Who decides?

  6. 6Execution and reconciliation7 min

    What is posted or transmitted? How do you know it was accepted? Which financial and system records must tie?

  7. 7Audit test5 min

    A different staff member reconstructs the case. What evidence is hard to find? What relies on memory?

Shadow validation, historical
Synthetic product gates pass. External pilot gates remain.
92%
disposition agreement
Disposition agreement
Eli's disposition matches the office's historical decision
Pass
92%target ≥ 90%
Audit-packet completeness
Each case produces a program-review-ready package
Pass
97%target ≥ 95%
Critical compliance misses
A wrong dollar determination or a missed required action
Pass
0target 0
Staff-touch reduction
Share Eli cleared straight-through without a human touch
Pass
34%target ≥ 25%

What the shadow run found

Modeled on the cohort's per-case economics, no live effect

70h
staff-hours saved on the cohort
97%
cases with a complete audit packet
0
critical compliance misses
34%
cleared straight-through
The 90-day path
Phase 1
Week 0
Trust packet

Security review: the model card, data-handling controls, RLS, retention schedule, and the mode-boundary memo. The artifacts a procurement and IT review need.

Phase 2
Weeks 1-2
Data extract

A de-identified historical cohort (300-500 cases) loads into a scoped tenant. Direct identifiers are masked; the distribution that makes the queue realistic is preserved.

Phase 3
Weeks 2-5
Historical shadow

Eli replays the cohort with no live effect. Its disposition is compared to what the office actually did, and the gates below decide whether to advance.

Phase 4
Weeks 5-10
Live-assist (gated)

Only if the gates pass: Eli prepares real cases, your staff approve every call. Straight-through clearing stays off until the office is ready.

Phase 5
Weeks 10+
Managed ops

Optional: Eli's team works the queue on the same software while your office keeps the final sign-off, so the queue scales without new hires.

What the pilot has to prove

Seven dimensions, agreed with your office before any data is exchanged and measured at exit

DimensionAgreed up frontEvidence at exit
PopulationInclusion and exclusion rules, with control totalsYour report counts and dollars match Eli's import
CorrectnessGolden cases and the acceptable toleranceEvery intermediate value and final result agree
ExceptionsNamed unsupported and ambiguous scenariosEach routes to the right owner, with no false confidence
WorkflowRequired human gates and downstream statesApproval, exported action, response, and reconciliation on record
EfficiencyBaseline touches and cycle time, with a targetMeasured before and after on the same cohort
DefenseRequired audit-packet contentsA blind reviewer reproduces the decision
Data boundaryDe-identification, retention, access, and deletionA signed handling protocol and a purge report

Shadow results, sample

Eli's disposition vs the office's historical decision (182 cases replayed)

CaseOffice didEli wouldAgreementAudit packet
Bryce WhitakerRoute to humanRequest docsdocumented diffcomplete
Fatou ParkRoute to humanRoute to human✓ matchcomplete
Liang KowalskiRoute to humanRoute to human✓ matchcomplete
Mohammed HernándezRoute to humanRoute to human✓ matchcomplete
Mateo RossiRequest docsRequest docs✓ matchcomplete
The pilot deliverable kit
MODEL-CARDModel cardWhat the AI does and does not decide; measured accuracy.›RETENTIONData handling & retentionFERPA, IRS Pub. 1075, RLS, the records-retention schedule.›AUDIT.PDFAudit package sampleThe program-review artifact, per case.›TRACE.JSONReasoning trace sampleEvery step Eli took, replayable.›INTEGRATIONSIntegration specThe SIS connector contract and de-identified roster path.
The 5-minute proofSecurity & compliance
Olivia LimCorrectRequest docsdocumented diffcomplete
Sofía RomanoRequest docsRequest docs✓ matchcomplete
Rohan RomanoRoute to humanRoute to human✓ matchcomplete
Dmitri HernándezRequest docsRequest docs✓ matchcomplete
Hana AliRequest docsRequest docs✓ matchcomplete
Liang RomanoRequest docsRequest docs✓ matchcomplete
Rowan TranClearClear✓ matchcomplete
›
SOWDraft SOW & HECVAT response kitPrepared per school for review: scope, gates, mode-boundary memo, and security-questionnaire evidence.