All posts
·8 min read

Claude for AR Teams: Invoicing, Collections, Cash

Claudeaccounts receivableAR automationcollectionscash applicationDSOAI finance

Part of our series on Claude for finance teams. The surface map is in Claude Has Four Surfaces.

AR has the strangest incentive structure in finance: the work is mostly polite persistence — invoice on time, remind on time, match the money when it lands — and yet every day of slippage is real cash. DSO doesn't drift because anyone decided to tolerate it; it drifts because reminders are nobody's favorite Tuesday task and cash application is everyone's least favorite Friday one.

Which makes AR nearly ideal agent territory: high-cadence, rule-driven, evidence-based — with one part that must stay human, and we'll get to it.

Scenario 1: Invoicing in a sentence

Invoice Karma Foods for May — the usual retainer plus the 12 extra consulting hours at the contract rate, NET 30.

Claude pulls the customer record, the agreed rates, and the numbering series; drafts the invoice; routes it through the approval lane; and books the AR entry — document attached, audit trail included. Recurring invoices run on schedule as agent work. Multi-entity and multi-currency are native, so the Swedish subsidiary invoicing in SEK isn't a special case.

The same conversational surface answers the daily questions: what's outstanding over 60 days, who's our biggest exposure, what did Hala Digital pay last quarter and how late were they on average?

Scenario 2: The reminders that actually go out

Collections cadence is the clearest DSO lever and the most reliably skipped task. A collections agent (our bills-collector) runs the cadence as scheduled work: polite reminder before due date, firmer at +7, escalation note at +21 — your templates, your tone, your thresholds, written as readable Skills. Every send is logged against the invoice.

The agent also knows when not to send: disputed invoices, customers flagged for a human touch, or accounts where a payment just landed. One operational lesson from running this in production: the fallback for "who gets the escalation email" must be configured — early on, missing owner records meant skipped sends, which is the failure mode you don't see until you look. The fix shipped; the lesson stands: silent skips are worse than wrong sends, so the agent reports what it didn't do, too.

What stays human in collections — permanently, in our view: the judgment call. When to pick up the phone, when to offer a payment plan, when to stop work for a client. Claude prepares the context for that call (history, promises made, pattern of lateness) better than any CRM note. It doesn't make the call.

Scenario 3: Cash application that doesn't eat Fridays

Money lands; someone has to figure out what it pays. The reconciliation agent does this as background work: exact matches (amount, window, reference) confirm automatically; near-matches — partial payments, two invoices in one transfer, the customer who always rounds down the FX — queue with confidence scores and candidate invoices attached; true mysteries surface with context. Stripe receipts flow in through the connector and match against their invoices the same way.

The result isn't "no human ever touches cash application." It's that the human touches the genuinely ambiguous tenth of it, with the candidates pre-assembled. (Watch a real matching run.)

Scenario 4: AR analytics without the export

Aging with collection-priority ranking, DSO trend, concentration risk ("two customers are 61% of receivables — here's the exposure if either slips"), payment-behavior patterns per customer, expected receipts feeding the cash view your treasury colleague reads — all live queries against the ledger, exportable to Excel artifacts when the board wants the chart.

What's real today — and what isn't

Real, shipping now: conversational and scheduled invoicing through approval lanes; the collections agent with configurable cadence, templates, and logged sends; 3-pass cash application with confidence-scored review queues; Stripe receipt matching; aging, DSO, and concentration views on demand; multi-entity, multi-currency AR on a full double-entry ledger with per-row audit.

Not there yet — and we'd rather tell you: a customer-facing payment portal ("click to pay this invoice") doesn't exist — invoices carry your bank details and the matching handles the rest; credit scoring and credit-limit automation are basic (limits are yours to set; Claude reports against them but doesn't underwrite); dispute management is a workflow you run in conversation, not a structured module; dunning beyond email (SMS, postal, legal handoff) isn't built; and no agent will ever decide to stop serving your customer — pausing work for non-payment is a human decision with a name on it.

How AR teams adopt this

Cash application first — it's measurable within one statement cycle and touches no customer. The collections cadence second, starting with the gentle pre-due reminder (lowest risk, immediate DSO effect). Conversational invoicing last, usually replacing the template-and-attach ritual one customer segment at a time. All of it works in bridge mode beside your existing ledger — here's how the two options work.

The control philosophy behind every send and every match is in The Thinking Behind Artifi. Real runs in the walkthroughs.

SUBSCRIBE · NEW POSTS · NO SPAM

Get the next post in your inbox.

Practitioner notes on AI-native finance. One email when something new ships. Unsubscribe any time.