Urgent.News

What's breaking now, across hundreds of outlets.

AI

Agentic AI for Payment Exceptions in LangGraph: Deadline Triage, Sanctions Hits & a Release the LLM Must Never Make

Payments is the domain where "the happy path is already automated" is most literally true — straight-through processing has existed for decades, and a clean payment never touches a human. All the cost lives in the exceptions : the malformed beneficiary reference, the sanctions screening hit that's probably a false positive, the duplicate suspicion, the failed downstream leg — each needing context…

In the payments domain, straight-through processing (STP) is highly automated, leaving exceptions as the primary source of manual work. These exceptions include malformed beneficiary references, sanctions screening hits, duplicate suspicions, and failed downstream legs. Each exception requires context from three systems, a judgment call, and often sign-off, all under a hard deadline.

This post introduces an agent within the exception queue using LangGraph, with the strict rule that release, hold, or reject decisions are made by a human, every single time.

The agent classifies exception types and urgency by a small model while an arithmetic calculation determines how much time remains before a cut-off deadline, never relying on the model for urgency. If the remaining time is less than 30 minutes and the payment amount is considered high value, the agent escalates the exception to a human.

Gathering evidence for each exception is done using read-only tools that return sourced Evidence with an ID that must be cited later in the explanation. This evidence includes beneficiary history, sanctions hit detail, duplicate scan results, and scheme rules for IBANs, BICs, and purpose codes. For example, mentioning prior payments to a beneficiary with a similar malformed reference helps to review judgments.

Sanctions hits are a particular challenge, as screening engines tend to over-sensitize, flagging legitimate payments as false positives. The agent drafts a discrimination narrative for these cases, but it cannot clear the false positive, as that decision is reserved for a human. Similarly, the agent proposes field repairs, but these must be re-validated against deterministic scheme rules before being accepted.

The agent's output is a disposition memo, not a payment instruction. Once approved, a human will turn this memo into money movement. This process ensures that the integrity and compliance of payment exceptions are maintained, avoiding costly errors and regulatory issues.

Rewritten by urgent.news; may contain errors — check the original.

Read the original at dev.to →

Report a problem with this story

More in AI

Upscaling AI-Generated Video to 4K: Resolution, Bitrate, Frame Rate and Artifacts

Disclosure: I'm writing this as part of the VideoUpscalerAI team. The product-specific numbers below come from our own published docs, and the rest is general video engineering that applies to any…

  • 4K upscaling requires consideration of resolution, bitrate, frame rate and artifacts
  • Integer scale factors (2×) are cleaner than higher fixed factors
  • Generated footage often shows artifacts like smeared textures and blocky compression

Your AI Agent Wrote Code. Where Does It Run? A Sandbox Runner in Under 50 Lines of Python

Your AI agent just wrote some Python and wants to run it. Where does that code actually execute? For a lot of agent products the honest answer is "on the same server as everything else, with the same…

  • Sandbox runner implemented in under 50 lines of Python
  • Limits CPU usage, memory usage, file size, wall clock time, output size
  • Executes code in separate process group with restricted environment variables

More from Saturday 21 November →