Case study · Energy & utilities · Finance operations
A regional utility providerReconciliation the finance team used to run by hand, now run by agents.
Every cycle, the utility's finance team reconciled billing, metering, payment, and bank records line by line, then chased down whatever didn't tie out. We built a set of AI agents that pull the source records together, match what lines up, and route only the real exceptions to a person, with the evidence attached.
The problem
Every billing cycle, records for the same transactions live in different systems: the billing platform, the meter data, the payment processor, and the bank. Someone in finance had to open each one and match them by hand, line by line, before the books could close.
Most lines matched cleanly. The work was in the ones that didn't: a payment posted a day late, a partial amount, a fee the bank took out, a meter adjustment that never made it to billing. Each break had to be found, explained, and resolved by a person reading across screens.
The matches were mechanical. The exceptions were where the real work, and the real risk, lived.
Reconciliation is a high volume, error sensitive queue that sits directly in the path of the close. Exactly the kind of workflow custom AI should own end to end.
What we shipped
A set of reconciliation agents built around the utility's own systems and the rules its finance team already used.
Source ingestion
Pulls the records for each cycle out of billing, metering, payments, and the bank into one place the agents can work over.
Matching agents
Ties transactions together across systems, tolerant of the timing gaps and partial amounts that break naive matching.
Exception routing
Surfaces only the items that don't reconcile, each with the candidate matches and the source records attached, for a person to decide.
How we built it
We started from how the team already reconciled, not from a generic template, so the agents matched the way finance actually thinks about these accounts.
Phase 01
Map the reconciliations
Cataloged which accounts get reconciled, against which sources, and what "matched" means for each one.
Phase 02
Build the matching agents
Wired the agents into each system and tuned matching to the real tolerances, timing gaps, and fees the team handled manually.
Phase 03
Route the exceptions
Stood up the exception queue with evidence attached, and folded each human decision back in so the agents keep improving.
Outcomes
What changed after the rollout.
Finance team refocused
Nobody matches line by line anymore. The agents clear what ties out, and the team spends its time on the exceptions and judgment calls that actually need a person.
A clean audit trail
Every match and every exception carries its source records and the reasoning behind it, so the close and the audit start from evidence instead of a spreadsheet someone has to reconstruct.