Skip to content

Work

Agent orchestration for the back office

A working prototype: five role-bounded AI agents process accounts-payable invoices with duplicate detection, fraud signals and confidence-based escalation to a human.

What it proves
Multi-agent orchestration on an accounts-payable process, with a human kept in the loop by design.
Method stages
Built from
  • Claude tool-calling
  • Per-agent, role-bounded toolkits
  • Duplicate-invoice detection against system memory
  • Fraud and anomaly signals: IBAN country mismatch, sender domain age, writing-style deviation
  • Side-by-side translation of non-English invoices
  • Confidence-based escalation to a human decision request
Sectors

The problem

Accounts payable combines high volume, direct financial exposure and a fraud surface. Most automation attempts build one system that reads, matches, checks and decides. Nobody trusts that black box enough to take the human out.

What we built

Five agents share one workflow, and each has a narrow, role-bounded toolkit.

  • Intake reads the invoice and structures it.
  • Matching checks it against purchase orders and prior invoices held in system memory, which catches duplicates.
  • Fraud runs three checks: whether the payment IBAN’s country matches the vendor’s known country, how old the sending domain is, and whether the writing style matches this vendor’s past correspondence.
  • Translation shows any invoice not in the operator’s working language side by side with the original.
  • Decision combines the other four outputs into one recommendation with a confidence score.

No single agent makes a final payment decision. Anything below the confidence threshold goes to a human as a decision request: an unfamiliar vendor, a mismatched IBAN, a duplicate flag. The request carries the full reasoning trail, showing which agent flagged what and why.

Role boundaries also limit damage. A translation error stays a translation error, because the agent that translates is not the agent that decides.

What it proves for a client

Agents on a high-stakes process can be bounded by role, scored for confidence and routed to a person. The human review path is part of the first architecture decision.

Method stages

It sits at Stage 02 · Prove the value. Stage 03 · Engineer the system would harden it with monitoring, access controls and an agreed approval policy. See the KARPINITY Method.