What an AI expense audit can and can’t read on an invoice

Automated invoice reading is good at the part people skip: reading every line of every invoice the same way, extracting amounts, dates and descriptions, and applying the same rules each time. It is weak where the answer isn’t on the page: whether a service was needed, what a contract clause means, or what happened at the dock. A useful audit uses software for the reading and checking, and keeps a person on the findings.

What can be read reliably?

• Typed line items: descriptions, quantities, rates and amounts on digital or cleanly scanned invoices.
• Repeated structures: the same vendor’s invoices month after month, so changes stand out.
• Percentages and totals: a fee as a share of the base, a surcharge rate, a running total.
• Dates and identifiers: service dates, pro numbers, tracking numbers, so duplicates can be found.

What is hard to read?

• Handwritten notes and stamps on tickets.
• Skewed, faint or photographed scans, especially tables that break across pages.
• Vendor shorthand: codes that mean something only in that vendor’s system. A rule can’t flag what it can’t name.

Any honest tool should say when it couldn’t read something rather than guess. The free invoice check on this site is deliberately simple: you type or paste the lines, and it tells you when it couldn’t find amounts.

What can’t software know from the invoice alone?

• Whether a liftgate was needed. That depends on the delivery site.
• Whether a fee is allowed. That depends on the contract’s wording.
• Whether a dumpster was full. That depends on someone looking.
• Whether a cheaper option is better. That depends on service, damage and the people involved.

These are why an audit asks for contracts and operating facts, not only invoices, and why each finding states its evidence and confidence.

Rules or models?

Two techniques get mixed together under “AI audit”. Extraction (turning a document into structured lines) is where machine-learning models help most. Checking (is this fuel percentage what the table gives at this week’s price?) is best done with plain rules that anyone can read and test. Using a model to decide whether a charge is wrong, without showing the rule, makes findings hard to verify. OpEx Agent’s approach is to keep checks as written rules, show the method, and attach the evidence line to every finding.

What about the files themselves?

Invoices carry account numbers, addresses and pricing you may consider confidential. Before sending invoices to any audit service, ask what is kept, for how long, who can see it, and whether it is used to train models. See what we keep for this site’s answers today. The site does not accept file uploads.

Where this leaves an audit

Reading is cheap; judgment is not. The right split is software for consistent reading and rule checks across every line, and a person to confirm findings with evidence, weigh them against operations and present them as decisions. That is how the first audits are being run during early access.