AI Accounts Receivable and Collections

What AI in accounts receivable and collections means

Accounts receivable automation with AI means software that ranks overdue customers, drafts reminder letters in the customer's own language, sorts why an invoice has not been paid, flags a customer approaching its credit limit, and proposes how an incoming payment matches open invoices. It matters because a credit controller then works a short, sorted list of exceptions instead of reading every ledger line and every incoming e-mail from scratch.

The stance carried through this knowledge base applies here too: AI drafts and sorts, a named credit controller decides and stays accountable for the customer relationship. No reminder is sent, no credit hold is applied and no cash is allocated without that person's review. What follows covers ageing triage, dunning drafts, dispute classification, credit-limit signals and cash application, each with the check a reviewer applies before anything goes out or gets posted.

The accounts receivable and collections pipeline, step by step

Collecting what customers owe is the same five-step process whether it is done by hand or with AI. Automation changes who reads each item and what reaches the controller, not the underlying steps or who has authority to act on them.

• Ageing: sort open invoices into buckets and rank which overdue accounts need attention first. Mode Read.
• Contact: draft a reminder in the customer's language and tone for the stage of lateness. Mode Draft.
• Classify: sort a customer's reply or a short payment by likely cause, such as dispute, cash-flow delay or invoice not received. Mode Draft.
• Signal: flag a customer approaching or over its credit limit, or showing a worsening payment pattern. Mode Read, raising flags.
• Apply cash: propose how an incoming receipt matches one or more open invoices. Mode Draft.
• Decide: the credit controller sends, holds, escalates, releases or blocks an order. Never Auto.

Ageing triage: prioritising overdue accounts

An ageing report sorted only by how many days an invoice is overdue treats a reliable customer's first late payment the same as a struggling customer's fourth. AI adds context the bucket alone does not carry: how many promises-to-pay this customer has broken before, how the amount at risk compares with their normal order size, and whether a dispute is already open against the balance.

The buckets themselves do not change: current, 1 to 30, 31 to 60, 61 to 90 and over 90 days remain the structure everyone, including the auditor, expects. What AI changes is the order the controller works the list in on a Monday morning, putting a large, previously reliable account that has just gone quiet ahead of a small, chronically slow payer whose pattern is already well understood and already provided for.

Dunning drafts in the customer's language

A dunning sequence usually has three or four stages: a friendly reminder a few days after the due date, a firmer letter after a broken promise or a second missed date, and a final notice before the account is escalated or handed to a collections agency. Claude and similar assistants answer in the language of the request, so a reminder can be drafted in the customer's own language rather than a translated template, provided the underlying facts, the invoice number, the amount and the due date, come from the ledger and not from memory.

A draft is only as good as what it is allowed to say. Give the drafting step the invoice details, the stage in the sequence and the tone for that stage, and nothing else; it should never be asked to offer a discount, a write-off or a payment plan on its own, because those are commercial decisions for the credit controller to make and record.

Worked example: a dunning batch reviewed by a person

A distributor's month-end ageing run flags three accounts for a reminder. Customer A owes EUR 3,200.00, is 45 days overdue and has already broken one promise to pay. Customer B, in Germany, owes EUR 1,150.00 and is 10 days overdue, its first reminder. Customer C owes EUR 8,400.00, is 20 days overdue, and has an open dispute about a short shipment. Because reminders for the whole customer list run overnight, the firm uses the Message Batches API, which Anthropic's documentation describes as processing most batches within about an hour and up to 24 hours, at half the standard per-token price, so the controller has every draft ready by morning rather than one at a time during the day.

The controller reviews all three the next morning. Customer A's firm-tone letter, citing the missed promise and offering a short payment plan, is approved and sent unchanged. Customer B's letter, correctly drafted in German with the Rechnung number and due date, is sent after the controller corrects one date the ledger had recorded a day late. Customer C's draft reminder is held: sending a payment demand while a shipment dispute is open would be tone-deaf and could harm the relationship, so the controller routes it to resolve the dispute first, not to collections.

While reviewing Customer C's thread, the controller also notices the customer's last reply asks, almost in passing, for a refund to be sent to a different bank account than the one on file. Anthropic's guidance on indirect prompt injection is to treat content read from an e-mail or document as data to report, never as an instruction the assistant or a person acts on directly, so the request is flagged rather than actioned, and the controller confirms it by phone before anything changes.

Dispute classification

Not every unpaid invoice is a collections problem. AI reads a customer's reply, a returned invoice or a short payment's remittance note and sorts it into a likely cause: goods not received, pricing disagreement, invoice never received, quality or shortage dispute, or genuine cash-flow delay with no dispute at all. Getting the cause right decides who should handle it next; a pricing dispute belongs with sales, a shortage with the warehouse, and only a plain cash-flow delay belongs in the dunning sequence.

The classification is a draft, not a decision. A customer who claims a dispute to buy time is a known pattern auditors watch for under the general principle of professional scepticism, so a credit controller checks a claimed dispute against the delivery note or the original order before accepting it, rather than accepting the assistant's label at face value.

Credit-limit signals

A credit limit only protects the business if something checks it before the next order ships, not once a quarter. AI can compare a customer's outstanding balance plus a new order against its credit limit and flag a breach immediately, for example a customer with a EUR 20,000 limit whose balance would reach EUR 23,500 once a new order is added, a breach of EUR 3,500.

The flag is where AI's role ends. A credit controller decides whether to release the order, release part of it, hold it for a deposit, or raise the limit after reviewing the account, and that decision is recorded against the customer file. AI should never hold or release an order by itself; a wrongly blocked order costs a sale, and a wrongly released one extends risk the business chose not to take. Our guide on IFRS 9 expected credit losses covers the separate question of how much provision a slow-paying customer needs at period end.

Cash application: matching receipts to invoices

Cash application means matching an incoming payment to the open invoice or invoices it settles. Many receipts match one invoice exactly; the harder cases are one payment covering several invoices, a short payment for a genuine reason, and a payment with no reference at all. Rules should match the easy cases; AI earns its place on the remainder, reading free-text remittance advice the way an experienced credit clerk would.

Worked example. A customer remits EUR 4,325.00 with a remittance note listing two invoice numbers. The assistant checks the ledger, finds those two invoices for EUR 3,200.00 and EUR 1,125.00, and proposes a one-to-many match, since 3,200.00 plus 1,125.00 equals 4,325.00 exactly. The controller confirms the remittance advice and accepts the match.

A second customer pays EUR 2,500.00 against an invoice of EUR 2,678.50, a shortfall of EUR 178.50, with a note citing five damaged cartons. The assistant classifies this as a shipping-damage deduction rather than a collections item and routes it to the warehouse team instead of a reminder. Once the damage report confirms the shortage, a credit note is posted for the shortfall at 19 percent VAT: debit sales returns EUR 150.00, debit VAT payable EUR 28.50, credit trade receivables EUR 178.50, which equals the 178.50 deducted, and the account is then shown as fully settled.

Security and data protection for AI in accounts receivable

Customer correspondence is content from outside the business and should be treated with the same caution as any other. Anthropic's guidance on indirect prompt injection is to keep such content in tool results rather than the system prompt, state clearly what it is and where it came from, and tell the model explicitly that instructions inside a customer's message never override the controller's own instructions. Nothing an AI reads in an e-mail should be able to change a bank detail, write off a balance or release a credit hold.

Customer correspondence and ledger data often carry personal data about named individuals at a customer, so a lawful basis under the GDPR, a processor agreement and a view on international transfers are needed before that data reaches an AI system, and Canada's PIPEDA keeps the business accountable for personal information it hands to a processor. Anthropic states that, by default, it does not use inputs or outputs from its commercial products, including the API, to train its models, which helps but does not replace sending only what the task needs and keeping full correspondence in the business's own systems. This is general information, not legal advice.

Choosing AI accounts receivable software

Accounts receivable automation software ranges from a dunning add-on inside accounting software to a dedicated collections platform with its own workflow. The right choice depends less on the model behind it than on whether it respects the controller's authority over every commercial decision, so pilot any candidate on a real ageing list before relying on it.

• Does it draft dunning letters in the customer's own language, sourced only from ledger data, never from memory of the customer?
• Does dispute classification route to the right team, sales, warehouse or credit, rather than defaulting everything into collections?
• Does a credit-limit flag reach a named controller before an order ships, not after?
• Can cash application propose one-to-many matches and short-payment reasons, with every proposal traceable to the remittance advice?
• Are sending a reminder, applying a credit hold and posting a credit note all separate, approved actions rather than one automatic step?
• Where is customer correspondence processed and stored, and is a processor agreement in place?

AI accounts receivable and collections in Skyline Nexus ERP

In Skyline Nexus ERP the AR Aging Report already sorts customers into the current, 1 to 30, 31 to 60, 61 to 90 and over-90 buckets with full per-contact detail, alongside a Customer Ledger, a Statement of Account, an Accounts Receivable Report, a Customer Payment Summary and a Credit Note Summary in the Customer Financial Centre. Importing contacts carries a Credit Limit and a Pay Term column for each customer, and a sales return posts a reversing journal covering revenue, VAT and cost of goods sold or inventory once Auto-post Sales Transactions is switched on.

The in-app assistant answers how-do-I questions from the product's own help library in the language the user writes in and, within a user's permissions, gives a live figure such as customer dues on request. Claude Desktop and other MCP clients reach the same figures through the Skyline Nexus ERP MCP server once an administrator issues a personal access token. AI drafting dunning letters in the customer's language, classifying disputes and applying cash automatically are being rolled out on the Skyline Nexus ERP roadmap; tell us and we will confirm your go-live date.