AR Reconciliation Agent — StaffingAgent.ai

The Problem

Cash arrives. The subledger drifts. Disputes age into write-offs.

Lockbox files reference the wrong invoice number. ACH lump sums cover dozens of invoices. Short-pays look like full payments until month-end. AR specialists burn 6–12 hours a week reconciling cash by hand. Unidentified deposits sit in suspense for weeks, and disputes get discovered too late to defend.

Meanwhile collections chases invoices that were already paid, the aging report can’t be trusted, and the CFO can’t put a real number on DSO.

How It Works

Deterministic-first matching. Claude only on the exceptions.

Fast match (no LLM, no token cost) — exact invoice number + amount + customer, learned-pattern lookup, bundle matching, customer-amount-unique. Resolves the clean majority on rules alone.
LLM match (Claude, exceptions only) — parses messy remittance PDFs, interprets deduction codes (CB-25 chargeback, PR-LATE late penalty, etc.) using per-tenant memory, reasons over ACH lump sums to rank candidate invoice bundles.
Classify and persist — every match gets an outcome (exact / short-pay / overpay / partial / bundle-split / unidentified / dispute-flagged), a match method, a 0–1 confidence, and proposed cash applications that sum to the receipt amount.
Disputes route to Collections — any short-pay the agent flags as a likely dispute hands off cleanly to the Collections Communications agent for outreach drafting.
Per-tenant learned memory — recurring deduction codes, client payment cadences, and short-pay reasons accrue per tenant so the agent gets sharper each cycle.

Doctrine

The LLM never moves money. A human does.

The math (sum of proposed applications equals the cash receipt to the penny) is verified by a deterministic check before anything reaches the AR subledger. Default launch posture is zero auto-apply — AR specialists approve every application. Tenants can opt into a high-confidence auto-apply threshold once they trust the agent on their data.

Every match, every application, every status change writes an audit log entry with a SHA-256 input hash, so any decision is reconstructable months later from `run_id` + audit row alone.

If the LLM provider is down, the run still completes — unresolved receipts simply land as unidentified and route to a specialist. No records are ever lost to a failed model call.

AR Ops is two paired agents. AR Reconciliation hands off short-pays and likely-disputes to the Collections Communications agent for risk-tiered outreach drafting. Together they own the receipt → aging → outreach cycle. Read the AR Ops BRD →

Primary Users

Built for the people who chase the cash today.

AR Specialist / Cash Apps — a pre-matched queue split into auto-applied (do nothing), needs-review (decide quickly), unidentified (investigate). One-click apply / split-apply / reject with notes.
Collections Specialist — receives dispute-flagged matches with the inferred dispute type already filled in, so outreach drafting starts from a real reason, not a guess.
AR Manager / Controller — an aging report you can actually trust. Audit trail per cash applied. Visibility into DSO trend and disputed dollars.
CFO — defensible numbers for the close, days-to-apply benchmarks, and a quarterly ROI line you can point to.

Design Targets

What the Agent Is Built to Deliver

Design targets for the agent at launch. Pulled from the AR Ops BRD; not measured customer results. Actual outcomes will be baselined at implementation and reported in the quarterly ROI readout.

≥ 60%Reduction in AR specialist reconciliation hours (design target)

5–10 daysDSO improvement within two quarters of go-live (design target)

< 24hShort-pays and disputes surfaced after cash arrival (design target)

100%Audit-ready trail per match, application, and status change

Coming Soon — 2026 H2 roadmap. Ships in the Transform tier (paired with Collections in the AR Ops workflow). Start with the Command Center today →

See the Command Center Book a Meeting