Checkability is why finance works
The four shapes of AI that work reliably all share one property: a human can verify the output quickly. Finance is full of those, because numbers reconcile.
That makes it a comparatively low-risk area to start, provided the system suggests and a rule or a person decides.
Document extraction and matching
Purchase invoices, remittances, statements and receipts extracted into structured data, then matched against orders and receipts by rule with tolerances.
Extraction is the AI part; matching should be rules, because you want the matching logic to be explainable and adjustable by finance rather than probabilistic.
Anomaly detection that finance trusts
- Payments outside the normal pattern for that supplier
- Duplicate invoice detection across slightly different references
- Expense claims outside policy or outside a person's own pattern
- Journal entries at unusual times or amounts
- Supplier bank detail changes, flagged for verification
The bank detail change flag is the one with the clearest financial return. It catches the most common and most expensive fraud pattern affecting small businesses.
Cash forecasting
Predicting when invoices will actually be paid, based on each customer's history rather than on payment terms. Most businesses forecast on terms and are consistently wrong in the same direction.
A model trained on your own payment history is usually noticeably better and requires only data you already have.
Keep the ledger rule-based
Coding and posting decisions should follow rules that finance can read and change. AI can suggest a category with a confidence score; a rule should decide whether that suggestion posts automatically or waits for review.
This distinction is what lets an auditor understand the process, and it is what keeps a finance director comfortable.
Controls to build in
- Every automated action logged with its inputs and the version that produced it
- Thresholds above which a human must approve, set by finance
- Segregation maintained — the system should not both create and approve
- A sample reviewed monthly regardless of confidence scores