Case study / Applied AI in financial operations
Two billion dollars, reconciled to the penny.
How a bank-backed inventory financing venture went from next-day spreadsheets to a ledger that updates in about thirty minutes, with a person still in charge of every uncertain call.
The problem
Billions in inventory, reconciled by hand, one line at a time.
The venture finances inventory that moves through third-party warehouses. Every day those warehouses email reports describing what came in and what went out.
Each line had to be matched against a purchase order in the ERP to know what the venture owned, what it was owed, and what was at risk. Finance did that by hand: download the spreadsheet, hunt for the order, fix the reference, repeat.
That meant next-day visibility at best, a stack of spreadsheets nobody fully trusted, and a team that would have to keep growing just to keep pace with volume.
What made it hard
Five things a demo never has to deal with.
-
(01)
The inputs were human.
Reports came from several warehouses in different layouts, as spreadsheets and PDFs, sometimes late and sometimes twice. The system had to accept all of it without ever double-counting.
-
(02)
Two systems described the same order differently.
The warehouses' references and the ERP's keys did not line up. Some lines matched exactly once, some never, and some more than once. Each case needed its own path.
-
(03)
It lived inside a bank.
Security review, credentials kept strictly out of the wrong places, an audit record for every message, and an AI model that was never allowed to retain the data it saw.
-
(04)
Money moved on the output.
Anything uncertain had to stop and wait for a person rather than proceed on a best guess. Automation had to know exactly where its authority ended.
-
(05)
The first version outgrew its tools.
It started on a low-code platform and hit hard ceilings on reliability and testing. The parsing and finalization moved into versioned, tested code without pausing the daily flow.
How it works
Seven steps from an email to the ledger.
Inbox
Warehouse reports arrive by email, in whatever shape the sender uses.
Capture
Every attachment is fingerprinted, deduplicated, and logged before anything else happens.
Read
Line items are extracted by rules first. AI assists only where a layout varies.
Match
Each line is matched to its purchase order in the ERP: matched, no match, or ambiguous.
Review
Anything uncertain lands in a side-by-side review: the original document next to what was read from it. One click approves, corrects, or rejects. Nothing ships on a guess.
Send
Approved batches are hash-verified and posted to the bank's message bus with an audit record.
Ledger
The ledger of record updates within about thirty minutes, and reporting follows.
Principles
Four rules that held the whole thing together.
Deterministic first, AI second.
Extract everything that rules can extract. Let the model help only where layouts vary, and never hand it more data than the task needs.
Fail closed.
Unknown formats, missing matches, and ambiguous orders stop and get reviewed. A person's decision is recorded before anything moves.
Every message traceable.
Correlation IDs and an audit trail tie each posted line back to the exact email it came from, so any number can be explained.
Built for the next person.
Versioned contracts, 1,700+ automated tests, and a runbook, so the pipeline runs unattended and survives a handoff.
Outcomes
What changed once it was live.
~90%
Less time on reconciliation. Hours of manual matching each day became a few minutes of review each hour.
1.5 to 2
Full-time hires avoided as the venture scales to new customers on the same team.
$150K+
Estimated operating cost avoided every year, before the next customer is added.
100%
Of posted transactions matched to the penny. Anything uncertain waits for a person instead of a guess.
Thousands of pages of supplier PDFs and spreadsheets now move from inbox to ledger in about thirty minutes, and the people who used to key them in spend their day on exceptions instead. The finance team stopped growing to keep up: new customers are coming onto the pipeline without new hires, and every uncertain line is still decided by a person.