A 700-person back office paid $2.5B a year in utility bills on a 17-year-old system. Half of the extracted invoices hit an exception, and 4 in 5 QA flags were wrong. A prior automation had already failed for want of one view of paid, in transit and usage.
The check reads the account’s own state before the exception fires. Most of the queue never forms. The template builds itself from the invoice and the master data. 290 keyers became 65 pattern-builders and exception analysts.
Onboarding became the gate. Account templates come first, and go-live is at 75 to 80% readiness. 90 days became 30. Coverage on day one was traded for an operation that clears 1,900 payments a day, with 1 bill in 100 reaching a person.
Cost came first, $0.58 to $0.21 a bill. Revenue came next, because the same data layer feeds procurement, tariff and forecasting. The operation that pays $2.5B a year can buy on its own numbers.
The same reconciled record carries the climb to 1M.
Exceptions were half the volume. Now one invoice stops, with its evidence attached, before the due date.
Keyers became pattern-builders and exception analysts.
Each one is audited against the record and sent on the day.
The record was built for this operation.
The keyers became analysts, and the desk reads exceptions instead of invoices.
GOVERN · THE GATE
ACCOUNT TEMPLATE FIRST
The account template is created before the first bill arrives. It is validated on save, so the state row exists on day one. Go-live is at 75 to 80% readiness. A change of address takes a vendor 30 to 90 days, outside anyone’s control.
THE DESK · 290 KEYERS
BUILD · THE DATA LAYER
250,000 ACCOUNTS · 20,000+ VENDORS · EVERY BILL NORMALIZED
Procurement, tariff, forecasting and Scope 2 read the rows that paid the bill.
BUILD · THE TWIN
ONE IDENTITY
The twin models the operation. Every vendor, site and meter reconciles to one identity before a bill is read. The agentic system runs on top of it, over the data layer and the client’s own system layers. A record that will not reconcile is held and never published.
Forecasting and accruals, rate optimization and benchmarking read the same record. None of them can be cut from an extracted bill.
BUILD · ONE ACCOUNT RECORD
250,000 ACCOUNTS · 14 FIELDS
The people were the memory. She cannot tell whether this one was paid. The bill says a balance is owed. Nothing on her screen says whether last month’s payment went out. 32 of her 40 flags this hour will turn out to be nothing.
Nothing on the desk says it was paid. The flag stays. She cannot approve what she cannot see.
Nothing on the desk says it was not paid. The flag stays. A wrong reject is a vendor on the phone.
It goes to exceptions. It waits a day for the night shift. Half of 700,000 bills a month end here.
4 WINDOWS TO CLOSE ONE FLAG
Last bill, last payment and days since the cheque went out are written once, per account. A previous balance that matches a payment in transit now closes on the record.
The system knows 14 fields for every account. 65 people remain for the bill the record cannot answer. One of the 14 fields sends it to them.
A BILL NEEDS 6 FIELDS · FINDS NONE · STOPS FOR THE NIGHT SHIFTA BILL NEEDS 6 · FINDS 5 · THE METER IS THE ONE LEFTA BILL NEEDS 6 · FINDS 6 · PASSES BEFORE THE DUE DATE
700,000 A MONTH · ONE STOPS, WITH A REASON, BEFORE THE DUE DATE
GOVERN · THE SIGNATURE
EXCEPTION DESK · 65 ANALYSTS
99 bills in 100 close on the account’s own state. This one carries a charge name the master has never seen. The system will not price it and will not guess. It stops, with its evidence attached, 6 days before the money is due.
Nothing posts below this line without a person.
The charge is mapped and the bill posts. The mapping is written back into the charge master, so the next bill with this name runs clean. The exception does not come back, and the analyst spent her afternoon on the ones that do.
Nothing posts. The bill is held for the morning team with its evidence attached, 6 days before the due date.