Agentic claims-payment automation layered on a legacy carrier system, 82% touchless
Built an agentic automation layer on top of Halberd Mutual's legacy claims-payment system. 82% of vendor bills process touchless, the rest route to a human payment specialist, and the original software still runs underneath.
To maintain client confidentiality, the company and industry in this case study have been anonymized. The underlying solution is the same.
The problem
Halberd Mutual’s claims-payment team was processing every vendor bill by hand: pulling each bill into their claims financial system, matching it to the right claim file, validating coverage and the applicable fee schedule, and clicking through to authorize payment. The system itself wasn’t broken. It was the right system for their auditors, their statutory reporting structure, and their decades of process. Replacing it would have been a multi-year project for no business reason.
So bill volume just consumed people. The team ran a backlog at every month-end close, error rates climbed under deadline pressure, and there was no good way to scale headcount with claim volume.
What we built
An agentic automation layer that sits on top of the existing claims financial system. The agent reads incoming vendor bills (medical providers, body shops, independent adjusters, defense counsel, restoration contractors), extracts the structured fields (vendor, amount, dates of service, line items, claim reference), matches to the associated claim file, validates against policy coverage and the applicable fee schedule, and then drives the existing UI to process the payment end to end. The underlying system never knows it isn’t a payment specialist at the keyboard.
For bills the agent is confident about (82% of current volume), it processes without escalation. For the remaining 18% (out-of-network providers, unusual procedure codes, line-item discrepancies, ambiguous claim matches, anything outside the rule set), it routes to a human reviewer with context pre-loaded: the matched claim file, the rule that flagged the exception, candidate adjustments. A reviewer who used to spend five minutes per bill now spends under ten seconds.
No system was replaced. Auditors still audit the same software. The general ledger sees the same kind of transactions. Statutory reporting didn’t change. The agent is a layer, not a rewrite.
Results
82% of vendor bills now process without a human in the loop. The remaining 18% clear in a fraction of the time they used to because the reviewer starts with context instead of building it. The system runs around the clock, so the month-end close backlog disappeared in the first month. Error rates dropped substantially because the agent doesn’t get tired and doesn’t skip steps under deadline pressure.
More importantly: Halberd never had to migrate off the claims financial system that already worked for them. Auditors, finance, and the actuarial team all still see the same software they always have. The agentic layer just made that software faster, more accurate, and tireless.