AI for finance teams: from invoices to month-end, where to start.

At month-end, the same questions fill a controller’s calendar. Why doesn’t this account reconcile? What is behind this variance? Which invoices are still stuck in matching, and why? Most of the answers exist somewhere in the ERP, a delivery note or an email thread. Finding and explaining them is where the days go.

Finance work is full of reading, matching and explaining, which suits AI, and full of numbers that must be exact, which doesn’t. The division of labor follows from that: systems compute and reconcile, AI reads documents, drafts explanations and flags exceptions, and the people who sign keep the approvals. This playbook applies that division from invoices to month-end.

Which finance tasks suit AI, and which don’t?

If the output is a number that will be booked, a system computes it.

If the output is a sentence about that number, AI can draft it. Apply that rule task by task, and the division of labor draws itself.

Where does accounts payable benefit first?

Invoices that match their purchase order and goods receipt already flow. The time goes into the exceptions, and each one starts the same way: someone opens three documents, finds the difference and writes a note explaining it.

That is the first job to hand to AI. For each three-way matching exception, the system collects the evidence and explains the gap in a sentence.

Picture an illustrative invoice that bills the full order quantity at the agreed price, but the goods receipts show two deliveries, and the second delivery note carries a handwritten correction that lowers the quantity. The draft says exactly that, links the invoice line, both receipts and the scanned note, and proposes a next step: pay for the received quantity and request a credit note for the difference. The clerk agrees, edits or overrules.

The explanations cover a short list of exception types, each with its own evidence:

• Quantity differences: partial or split deliveries, returns and corrected delivery notes.
• Price differences: an outdated price on the order, a discount agreed after ordering, a different currency.
• Missing receipts: goods that arrived but were never received in the ERP.
• Possible duplicates: the same invoice number twice, or the same amount and date under a new number.
• No purchase order: services and one-off purchases that need a budget owner’s approval.

Three checks keep this safe. The matching rules and tolerances stay in the ERP. The system never changes an order, a receipt or an invoice to make them agree. And every explanation cites the documents it relies on. For invoices that arrive as PDFs and scans, and for the Türkiye-specific side of e-invoices, see our note on where e-Fatura ends and AI begins.

How can AI help at month-end?

Around the numbers, never inside them.

• Unreconciled items. Bank and intercompany reconciliations leave a residue the rules couldn’t match. AI reads the free-text descriptions, spots the invoice number with a typo or the payment that covers three invoices, and proposes a match with its evidence.
• Open differences. For each difference that remains, the system gathers the relevant documents, such as an invoice in transit or a disputed delivery, and drafts the explanation for the reconciliation file.
• Close status. A daily summary of which close tasks are open, blocked or late, and who owns them, drawn from the close checklist itself.
• Commentary drafts. First drafts of the explanations that accompany balances and movements.

In every case, figures come from the ERP or the BI layer. The model receives balances; it doesn’t produce them.

What about reporting and variance commentary?

Commentary is where AI can save hours of writing, and where a wrong number does real damage. The method is simple to state: figures are computed and checked upstream, the model receives a table rather than raw data, each sentence refers to the exact figure it explains, and the “why” comes from notes people already write, such as sales notes and operations logs. When no recorded reason exists, the draft says so instead of inventing one.

One trap deserves a name: arithmetic inside sentences. “Up sharply on the prior month” is fine when the table shows it; a percentage change the model worked out for itself is not. Differences, ratios and percentages are computed upstream and arrive in the table like every other figure. Our note on management reports with AI walks through it step by step.

Where must approvals stay with people?

Treat AI as a preparer in your segregation of duties. It prepares work but never approves its own, and its service account holds no posting or payment rights. People with authority approve:

• every posting and journal entry;
• payments and payment runs;
• changes to supplier master data, above all bank details;
• estimates that need judgment, such as provisions, write-offs and accruals;
• anything reported outside the finance team, from management packs to lenders;
• changes to tolerances and policies.

Every draft, the sources behind it and the approval are logged, so an auditor can follow a number from the document to the ledger.

How do you pick the first finance workflow?

Use the same tests as for any first AI workflow: frequent, bounded, built on data in reachable systems, and owned by a named person. In finance, that usually points to accounts payable exceptions or policy questions first, and to commentary once the figures feeding it are trusted. Policy questions, such as expense rules or approval limits, make a low-risk start: the answer lives in a document, the assistant cites it, and a wrong answer is caught when the request is approved. Record the baseline before building: exceptions per week, time per exception, items still open at month-end.

Adoption matters as much as the build. The illustrative finance AI champions engagement shows how a finance function can turn tool licenses into a weekly routine. If the list of candidates is long, AI strategy and discovery ranks them on evidence, and the ERP and enterprise workflows page shows what a first build looks like.