Automated + manual payment-to-invoice matching
A complete payment-to-invoice reconciliation workflow that combines automated and manual matching. Mode 1 (Full Reconciliation) runs the 6-pass reconciliation agent first — exact amount, reference number, cross-reference, approximate FX, amount grouping, and Haiku fuzzy matching — then analyzes what remains and proposes manual matches for complex patterns like N:N splits and partial payments. Mode 2 (Quick Manual Match) lets you directly apply a payment to specific invoices with split amounts and rounding write-offs. Mode 3 (Health Check) provides a read-only status dashboard showing matched, partial, and unmatched counts with aging analysis and next-action recommendations.
Assess current state — load unmatched payment and open invoice counts for AP and AR
Run automated reconciliation agent (6-pass matching algorithm)
Review agent results — how many matched, what remains unmatched
Group remaining unmatched items by party (vendor/customer)
For each party: load payments + invoices, propose manual matches using 6 strategies
Present match proposals as tables with amounts, logic explanation, and differences
User reviews and confirms matches (one confirmation per party batch)
Apply confirmed matches via submit("reconciliation", "match", {...})
Categorize unresolvable items (missing bills, unknown parties, overpayments)
Show before/after summary with total match rate
Identify the payment (by ID or search)
List open invoices for the same party with amounts due
User specifies which invoice(s) and amounts to apply
Preview the match — show status changes for payment and each invoice
Apply after confirmation
If any checkpoint fails, the workflow stops and reports the issue. Nothing posts to the ledger until every check passes.
Formulas, templates, regulatory rules — not hardcoded values. Forkable per firm, per client.
Designed to work out-of-the-box with the Arfiti ledger. Claude fetches your data, runs calculations, validates results, and submits entries automatically through MCP.
Or fork it. Skills, commands, validation logic, and reference data are all MIT-licensed. Swap the MCP tool calls for your own backend, or use the reference files as standalone Claude project resources.
Get the next post in your inbox.
Practitioner notes on AI-native finance. One email when something new ships. Unsubscribe any time.