Document Intelligence Agent

Most of what governs a business sits in documents nobody can search. Contracts, lease deeds, annexures, fee schedules, service POs, amendments — scanned years ago, filed by vendor, with no text layer and no index. When a question comes up that those documents answer, someone opens them and reads. This agent does that reading.

What it does

You point it at a document and ask a specific question — what does this agreement say the monthly charge should be, which party is liable for this cost, what notice period applies, does this invoice match what was agreed. It locates the answer, reports it with the page it came from, and flags anything adjacent that changes how the answer should be read.

The capability is easiest to see in a real example. Asked to verify a maintenance invoice against its governing agreement, the agent went into a long scanned lease deed with no extractable text, located the specific charge figure printed inside a schedule of maintenance charges, compared it to the invoice amount, and reported the variance. That is a finding a person gets only by reading the entire document, and only if they know to look for it.

Paired with the AP Invoice Agent, this is what performs the contract-and-PO comparison on every incoming invoice. It also runs standalone against a contract repository.

What we don't claim

We'd rather set this out than let a demo imply more than it should.

• —It reads what is printed; it does not compute figures from contracts. In the example above it located a figure the deed states, and compared it. It did not derive that figure from rates and areas, and we don't represent that it can.
• —It does not make the decision. Variances are reported to a person, never enforced, and never used to auto-approve or auto-reject.
• —Coverage on very long documents is bounded and stated. Where a document exceeds what can be read reliably in one pass, we say so and scope the handling rather than assume it worked.

What it needs

• —A location where the relevant documents live, organised so the right one can be found for a given vendor or matter — a folder structure per vendor is enough; a contract-management system is not required.
• —The questions you actually need answered, stated as questions. This is where the accuracy comes from.
• —A handful of documents with known answers, so results can be verified before anything depends on them.

What the output looks like

For each document and question, you get the answer, the page it was found on, and the surrounding text that supports it — so the finding can be checked in seconds rather than taken on faith. Where the agent compares two documents, matches and variances are listed side by side. Where it cannot find an answer, it says so instead of producing a plausible one.