Controller, CFO, finance operations, internal audit
How finance teams can clear routine expense reports overnight—without automating exceptions
Can we reimburse the clean reports by morning without giving the model payment authority?
Controller, CFO, finance operations, internal audit
Can we reimburse the clean reports by morning without giving the model payment authority?
Counts are from the walkthrough file below — a sample Monday morning, not a promised customer result.
By 8:00 a.m., policy-clean expense reports can be ready for reimbursement. Exceptions arrive with the original receipt, the relevant policy, the reason they were flagged, and the action that requires a manager. Overnight Engine connects to the systems finance already uses, checks each submission against company policy, and prepares the routine work. It does not remove financial controls: payments and designated exceptions remain subject to the organization’s configured approval rules.
Overnight Engine
41 reports are policy-clean and waiting on the configured payment run. Seven exceptions include the original receipt, the policy clause, and the action that still needs a person.
Exceptions
Reimburse
Exceptions
Policy
Northline Industrial, Monday 8:04 a.m. Open EXP-18442. The Copper Room receipt, the T&E clause, and the payment hold are the same objects Priya reviews. Approve or send a line back — that click is training data, not a policy change.
The walkthrough follows EXP-18442, filed Sunday night by Daniel Okonkwo after a Chicago service trip.
Employee submission
A Concur report with itemized receipts, a hotel folio, and a business-purpose note.
Card feed
Amex 4104 posts the same lines. The engine matches card to receipt; orphans stay exceptions.
Policy and employee record
T&E Rev 12 from the policy store, Daniel’s cost center and manager from Workday, payment authority from NetSuite.
Ingest
Receipts are read. Line items, merchant, tax, tip, alcohol, and guest count are extracted and tied to the card feed.
Policy check
Each line is compared to written policy: meal caps, city hotel bands, alcohol rules, late-submit windows, required attendee lists.
Split the queue
Lines that match policy and clear the configured confidence threshold are packaged for the payment run. Everything else is an exception with evidence attached.
System action, not payment
The engine can draft a request for a missing attendee list and write a hold into the payment preview. It does not release ACH.
Priya Shah opens Overnight Engine on the office machine. Forty-one reports are marked ready. Seven need her. EXP-18442 is one of them.
The hotel folio and the Uber are already in Friday’s payment file, under her $2,500 authority. The Copper Room dinner is not. She sees the $186.40 receipt, clause 4.2 and 4.4, the missing attendee list, and the $38 alcohol line with no director code.
She does not re-key the report. She decides whether to reject the meal, accept it with a director code, or wait for Daniel’s names. That decision, and the reason, stay on the file.
Automation boundaries are configured per company. Northline’s expense workflow is set this way.
Can prepare
Can execute
Always needs approval
If Priya rejects the meal because client dinners always need names, that override is captured with the receipt, the clause, and the outcome. It is not silently written into next week’s model.
Overnight the engine proposes hypotheses — raise the meal tolerance, require attendee lists, route hotel incidentals. Each one is tested against Northline’s own labeled history. A 15% meal-tolerance change fails that history and does not ship. The attendee-list rule passes, merges with the incidental rule, and the winning model is released for the next night.
The previous model remains the rollback. Live disagreement rate is monitored. A correction that does not beat history never becomes policy.
These are proposed pilot metrics, not claimed customer results. Baselines are taken from the entity’s last 60 days before hardware is live.
| Metric | How it is measured | Pilot target |
|---|---|---|
| Time to close a clean report | Submit → reimbursed, policy-clean only | Cut in half vs. baseline; many close overnight |
| Controller review rate | % of reports that still need a person | Fall as the model wins history tests; never forced to a number |
| Exception precision | % of flagged items Priya agrees were exceptions | High enough that the queue is trusted, not noisy |
| Cost per report | Fully loaded finance minutes + payment-file rework | Down on clean volume; exceptions stay human-priced |
| Override capture | % of controller decisions stored with reason + evidence | 100% of supervised actions in the pilot entity |
One legal entity, one card program, one policy revision. Expand after the pilot, not during it.
The engine is a machine on your floor. Receipt images, card data, and employee records are processed inside that perimeter. Integrations call Concur and the card network from your network, with credentials in the local enclave.
Learning stays on the encrypted drive. Weight updates are not shared with other customers. Telemetry off-box is hardware health — temperature, uptime, error rates — not T&E contents.
Ask for the private-deployment architecture if security or internal audit needs the network diagram before a pilot.
Flagship
Payment-ready overnight
Vendor match, three-way match, duplicates held for AP.
Claims
Claims prepared overnight
Complete files for adjusters. Coverage stays with them.
Healthcare
Coding without the backlog
Pre-coded charts with citations. Coder still signs.
Research
Protocol-amendment review
Redlines and eTMF gaps. RA still files.
Credit
Cleaner underwriting files
Docs, DTI, exceptions. No credit recommendation.
Next step
A 30-day supervised pilot on one legal entity. Hardware stays on your floor.