Automating the close after defining what you count.

Three concrete uses

1. 01Matching invoices, orders and receipts before payment

   AI Act level · Minimal risk

   1. 01Invoices receivedApproved platform, email, paper
   ↓→
   2. 02Reading and extractionAmounts, references, due dates
   ↓→
   3. 03Proposed matchOrder, receipt, tolerances
   ↓→
   4. CheckApproval by the accounts payable clerk
   ↓→
   5. 04Posting and paymentOnce discrepancies are resolved

   The problem

   Supplier invoices arrive through several channels and matching them with the order and the goods receipt is done by hand. Price or quantity discrepancies are spotted late, sometimes after payment.

   What the AI does

   AI reads each invoice, matches it with the purchase order and the goods receipt, and flags discrepancies with their reason. It approves no invoice and triggers no payment.

   Data needed

   Invoices received, purchase orders, goods receipts, the supplier master file and tolerance thresholds set by finance.

   Who signs off

   The accounts payable clerk handles each flagged discrepancy; the authorised manager approves payment.

   Minimal under the AI Act (EU): checking accounting documents, with no decision about natural persons.

   • AI agent
   • Evaluation
   • Human oversight

2. 02Explaining budget variances with verifiable sources

   AI Act level · Minimal risk

   1. 01Variances detectedAbove the set threshold
   ↓→
   2. 02Linked dataEntries, budget, forecasts
   ↓→
   3. 03Proposed explanationSources cited, unknowns flagged
   ↓→
   4. CheckApproval by the controller
   ↓→
   5. 04Report issuedAfter the controller’s review

   The problem

   Every month, controllers spend days explaining variances against budget, and two departments give two different explanations for the same figure.

   What the AI does

   An agent queries management data through the semantic layer, proposes an explanation for each material variance and cites the entries concerned. It changes neither the budget nor the forecasts, and it flags what it cannot explain.

   Data needed

   General ledger, budget, forecasts, analytical dimensions and the materiality threshold set by management control.

   Who signs off

   The controller approves each comment before it goes to the executive committee.

   Minimal: analysis of aggregated financial data, with no effect on individuals.

   • Semantic layer
   • AI agent
   • Hallucination

3. 03Prioritising collections and preparing customer reminders

   AI Act level · Minimal risk

   1. 01Overdue receivablesUp-to-date aged balance
   ↓→
   2. 02Priority rulesWritten and dated by finance
   ↓→
   3. 03Reminders preparedTone and documents tailored
   ↓→
   4. CheckApproval by the collections officer
   ↓→
   5. 04Sending and follow-upAfter approval, history kept

   The problem

   Late payments pile up and reminders go out in order of arrival, regardless of amounts, payment history or open disputes.

   What the AI does

   AI ranks overdue receivables using rules written by finance and prepares reminders suited to each situation. It sends nothing on its own, sets no credit limit and blocks no account.

   Data needed

   Aged balance, payment history, open disputes, contractual terms and billing contacts.

   Who signs off

   The collections officer approves each reminder; the credit manager decides on blocks and limits.

   Minimal for business customers that are companies. Assessing the creditworthiness of natural persons would be high-risk (EU · AI Act, Annex III).

   • AI agent
   • Guardrails
   • Logging