AI Training Example on Accounting Records
A customer sends $6,000. How much of it belongs to the $10,000 invoice on your screen? In one version of this example, the invoice has $2,500 left to collect. In the other, it has $8,500. Every dollar amount in the records is identical.
If you work in accounts receivable, you probably know which record to open next. That choice is exactly what makes this a useful AI training example. Reading three amounts is easy. Deciding which ones belong in the calculation requires knowing how the business records a payment.
Watch the invoice reconciliation walkthrough
Follow the documents, then do the math
These are two fictional versions of the same ledger. In each, invoice INV-1042 was issued for $10,000. A $1,500 credit memo for returned goods was posted and applied to that invoice. The customer also paid $6,000. The question is how much remains open on INV-1042 at October 31, after all entries shown have posted. Assume there were no other adjustments.
A · Payment applied to this invoice
RCPT-77 → INV-1042B · Payment applied to another invoice
RCPT-77 → INV-1043The receipt proves that cash came in. The payment application is the ledger entry that says which invoice received the cash. In Case B, the $0 in the equation means zero dollars of that receipt went to INV-1042. All $6,000 went to INV-1043, another invoice for the same customer.
What changes on the three financial statements?
Take the seller’s view. Assume the sale was earned on delivery, the $1,500 return reduced sales this period, and the $6,000 receipt cleared by October 31. These amounts exclude sales tax. The accounting is the same whether the receipt is posted against INV-1042 or INV-1043.
- Income statement
- $8,500 in net sales. The $10,000 sale less the $1,500 return gives us the sales figure. Collecting the $6,000 payment adds no new revenue. To calculate gross profit, we would also need the cost of the goods sold and the cost of any returned inventory put back into stock.
- Balance sheet
- Cash increases $6,000. Total accounts receivable increases $2,500. Across the three entries, receivables rise $10,000 for the sale, fall $1,500 for the credit, and fall $6,000 for the collection. That $2,500 is the change in total receivables from these transactions. The open amount on a particular invoice depends on where the cash was applied.
- Cash flow statement
- $6,000 of operating cash inflow. Cash came in once, when the customer paid. The invoice and the credit memo moved the receivable without moving cash. You can check the collection by subtracting the $2,500 increase in receivables from $8,500 in net sales.
Can the financial statements tell you whether the payment went to the right invoice? Suppose INV-1043 already had $6,000 open. Case A leaves $2,500 on INV-1042 and $6,000 on INV-1043. Case B leaves $8,500 on INV-1042 and $0 on INV-1043. Both add to $8,500 due from the customer. If an AI assistant tells collections to chase INV-1043 in Case B, the error will never show up in a check of total receivables. The posted payment application is what proves that invoice is settled.
Use the accountant’s work as a training example
For a worked training example, give the model the invoice, the linked credit, the receipt, and its posted application. Ask for the open balance on a named invoice as of a specific date. Pair those inputs with the accountant-reviewed answer, including which entries support it.
A worked answer for Case B: “INV-1042 has $8,500 open at October 31. CM-18 reduces it by $1,500. RCPT-77 was applied to INV-1043, so I applied $0 of that payment to INV-1042.”
A spreadsheet of ending balances alone would omit the decision the accountant made. The linked records show how someone can reproduce it. A buyer developing an accounting assistant could use examples like these to teach the assistant to follow an application through the ledger and explain the result.
Turn the same task into a test the model can fail
For evaluation, give the model a fresh document packet and the question. Keep the reviewed answer separate. A model that subtracts every payment from the invoice in front of it gets Case A right by coincidence. If it returns $2,500 again in Case B, the second case tells you exactly what it missed.
I’d score this pair on three things:
- Balance. Does it return $2,500 for A and $8,500 for B, both for INV-1042 at the same cutoff?
- Evidence. Does it cite CM-18 and correctly identify where RCPT-77 was applied in each case?
- Explanation. Can it explain why changing one invoice ID changes the answer by $6,000?
Keep both versions together when splitting your data. If these two cases are the test, teach with different transactions. Also remove any ending-balance field or reviewer note that gives away the answer from the packet the model sees. Two cases can expose this particular mistake. To measure performance across a business, test more cases, including exceptions.
Try a third case: the payment application is missing
You can verify the $10,000 charge, the $1,500 credit, and receipt of the $6,000. You cannot determine how much of that payment reached INV-1042. A good answer gives the $8,500 balance before any cash application and asks for the posted allocation before stating a final balance. A split payment could put that balance anywhere between $2,500 and $8,500.
There is a real parallel in AI testing. Artificial Analysis’s AA-Paperclip benchmark includes 500 receipts and invoices among 1,500 documents. Each document passes only when the model gets all 15 answers right. I’d use the same scoring idea in accounting. A model that reads every amount correctly but assigns a payment to the wrong invoice still leaves you with the wrong balance. Read the benchmark methodology.
What to inventory if your business has these records
Start with one completed reconciliation and work backward. Can you trace the approved balance to the original invoice, its credit memos, and the posted cash applications? Can you do the same for partial payments, returns, and payments split across invoices? Those exceptions give a prospective buyer specific tasks a model could learn or be tested on.
One useful entry in a data manifest might read: “Invoice, credit, receipt, and application records linked by consistent IDs, with posting dates, currency, and reviewed month-end balances. Includes partial payments and credits for returns.” Add the actual period covered, number of complete cases, and any limits on permitted use. If you prepare a sample, use consistent substitute IDs so the reviewer can still follow the transaction.
Use the Business Data Checklist to identify the records you control, then describe one workflow in a data manifest. Once you know what you have and what you can share, explore the data licensing opportunity.
