Look closely at invoice INV-88180. Its three line items add up to $8,240.00. The total printed at the bottom says $8,420.00. Two digits have swapped places, and the difference is $180.
A busy reviewer might miss it. An automated process that trusts whatever it extracts will miss it every time — and pass the wrong number on as if it were right.
Our new video, Checked, not trusted, follows that single invoice through DocReveal: from the moment it is read, through the check that flags it, to the moment a person decides what happens next. It runs three and a half minutes, you can watch it at the top of this page, and it is made for the people who carry the responsibility — finance directors, accounts payable teams, operations and audit — rather than for developers. Here is what you will see, and why it matters.
Reading is not the same as trusting
AI has become very good at reading documents. It will extract every line of INV-88180 correctly — including the wrong total, because the wrong total is what is printed. A perfect copy of a flawed document is still a flawed result.
That is the gap most document automation leaves open. Processing gets faster, and so does the speed at which errors travel into ledgers, reports and payments. The answer is not to slow down. It is to check every document against the rules your business already works by, automatically, before anyone relies on the data.
Read: every invoice becomes data
The video opens in Reveal Manager, where six September invoices are processed in a single run. Each file is parsed, mapped to the invoice model, checked against the rules and prepared for analytics, and the board shows where every file stands as the run progresses. There are no templates to maintain: suppliers use different layouts and languages, and the invoices are read as they come.
Five invoices pass. One, from a supplier called Stavlund, does not: its rules checks have failed.
Check: rules your auditors would write
The check behind that failure comes from Rules Manager. There, a business user puts rules together by clicking fields, conditions and values — no code, no scripts. In the video, one rule says that the line items must add up to the invoice total. Another makes sure the same invoice number never appears on two documents.
These checks are deterministic. They do not ask an AI for its opinion; they calculate and compare, the same way every time. Rules are grouped, versioned and published on purpose, and the audit log shows who created, changed and published each version. When someone asks months later why an invoice was flagged, the answer is on record.
The outcome is a verdict with a reason. In Business Rules Validation, INV-88180 is marked as failed, with the exact rule that failed and the values behind it: line items of $8,240.00 against a printed total of $8,420.00. Not “something looks wrong”, but what, where and by how much.
A failed verdict is not a dead end. It is stored with the document and published as an event, so it can start a flow in Integration Manager or wake an AI agent to take the next step.
Decide: an agent that knows when to ask
In the second half of the video, that event wakes an AI agent for invoice triage in Agent Manager — one of the digital colleagues we described in Agent Pool. Nobody starts it by hand; it runs on its own whenever a document is validated. Its first move is not to act but to read: the rules verdict and the extracted invoice data.
The agent works within limits its owner has set: a defined set of tools, a maximum number of steps per run and an autonomy profile that decides which actions need a person’s approval. Its instructions are clear about invoices that fail their checks — they go to a person. So the agent does what a good colleague would do: it asks. The run pauses in the approvals inbox with the question and the possible answers laid out.
A person reads the question and chooses to hold the invoice and request a corrected one. The agent resumes, keeps a note of the decision and finishes. Its trace shows every step — what it read, what it asked, the answer it received and how the run ended. The inbox is empty again, and nothing happened that someone cannot explain.
Why this matters beyond one invoice
INV-88180 is a small story about $180. The principle behind it applies to every document your company processes, and it fits in one line: AI proposes. Rules check. People decide.
- Treat extracted data as a proposal, not a fact. Find the places where extracted data is used without anyone checking it — that is where your first rules belong.
- Write down the checks your people already do by habit. Totals that must add up, numbers that must be unique, values that must stay within limits. If a reviewer checks it, a rule can check it on every document.
- Give automation a way to stop and ask. An agent that hands over what it should not decide is worth more than one that decides everything.
- Keep the record. Versioned rules, verdicts with reasons and agent traces with human answers turn “the system did it” into a history you can audit.
All companies and figures in the video are fictional demo data.
Checked, not trusted. Every document. That is how automation earns trust.