Documents Don't Make Decisions

Citation-first retrieval over PDFs, specs, manuals, compliance docs. Ask a question, get an answer traced to the exact page and section. No hallucination. No "the AI said so" without a receipt. I built this into a document intelligence system called RAGlue.

It worked. And then nothing happened.

The team still had to read the answer, interpret it, decide what to do, and figure out who needed to act on it. I'd solved the search problem. The search problem was never the bottleneck.

The action problem

Every document-heavy business I've worked with has the same pattern. The knowledge exists. It's in the specs, the manuals, the procedures, the contracts. People don't struggle to find it because it's hidden. They struggle because by the time they find it, interpret it, cross-reference it against three other documents, and work out what it means for the decision in front of them, the day is gone.

Construction is where this gets expensive. Rework, miscommunication, and poor data management cost US construction $177 billion annually, and a significant portion traces to drawing continuity errors. Sheet A3 says the ceiling height is 9'-6". Sheet M7 shows ductwork that won't fit below 9'-6". The spec existed. Someone would have found it eventually. "Eventually" is after the duct is fabricated, delivered, and doesn't fit.

That's not a search problem. That's an action problem. The information was there. The workflow that should have flagged it, routed it, and triggered a review before anyone picked up a wrench — that didn't exist.

Two billion-dollar answers that stop too early

The market has noticed that businesses are drowning in documents. Two responses have emerged.

The search layer helps you find things. Glean, valued at $7.2 billion, built its business on the premise that enterprise search over internal documents is valuable. It is. You ask a question, you get an answer from your company's own data. Useful. Incomplete.

The reasoning layer helps you understand things. Hebbia built an analysis engine that reads complex documents and reasons across them. A $700 million valuation at a 54x revenue multiple says the market agrees. More useful. Still incomplete.

Both stop at the screen. They give you an answer. You figure out what to do with it.

Capture, understand, act

The third move is the one that creates business value. Not "here's what the document says" but "here's the workflow that runs because of what the document says."

Capture. Bring the documents in. One intake point instead of shared drives, email attachments, and filing cabinets that nobody can search.

Understand. Extract intelligence with receipts. Every answer traces to the exact page, section, and revision. Your auditor can check it. Starting August 2026, EU AI Act transparency requirements for high-risk AI deployments make citation-first architecture table stakes, not a feature.

Act. This is where the value lives. The document doesn't answer a question — it triggers the next step. A spec contradiction flags a review. A compliance gap initiates a corrective action. A submittal package gets cross-referenced against the project spec, discrepancies get cited with evidence ("Section 23 05 00 requires Type L copper; submittal specifies Type M — see page 14"), and flagged items route through a structured decision workflow. The project manager makes decisions. The system does the retrieval, the assembly, and the routing.

That's what I'm building with RAGlue — the project that started it all. Not a search engine. Not an analysis tool. An information-to-intelligence pipeline where documents fuel the work instead of sitting beside it.

Services as software

When document intelligence connects to execution, something shifts. Services that used to depend on a person reading a document, interpreting it, and manually triggering the next step can run as software. The document becomes the operator.

I proved this pattern first with May Belle, a service automation platform that converts manual service delivery into software. RAGlue takes it further. The documents your business already owns — the specs, the manuals, the compliance frameworks — become the fuel for workflows that run themselves.

Most document AI asks you to be impressed that it found the answer. That's the demo that proves AI can do something without proving it will. I'm more interested in what happens after the answer. That's where the decisions get made. And documents don't make decisions. Systems that connect intelligence to action do.