Any document, any department — read, indexed, and answerable with a citation in seconds.

What it is

What is the AI Document Intelligence Engine?

The Document Intelligence Engine reads every document your business runs on — leases, loan files, tax returns, audit reports, claims, contracts — across legal, finance, compliance and tax, and turns them into a private knowledge base you can simply ask. Anyone types a plain-language question and gets an answer in seconds, with every claim cited to the exact source page. It is not a generic OCR tool, and it is not public AI pointed at your files; it is a grounded, cited, private layer over your own documents.

“Safe RAG” is the heart of it. RAG means the AI answers only from a specific set of documents rather than from the open internet; “safe” means those documents are yours, they stay inside your environment, and every answer links back to the page it came from. So the engine cannot make things up, an auditor can trace any answer to its source, and nothing is sent to a public AI service — the three things a compliance, legal or finance team needs before it will trust an AI answer at all.

The reserve study is one proof point. A property firm’s process that took two engineers two weeks — reading handwritten inspection sheets, keying data, running a 30-year model, formatting an 89-page report — now runs in hours, with the one illegible line flagged rather than guessed. The same pipeline reads a lease as easily as a loan file or a tax return, which is why one engine serves legal, finance, compliance and tax instead of four separate tools.

How it works

How the Document Intelligence Engine works, step by step

Six specialised steps take a document from wherever it lives to an answer you can cite. Each model is purpose-built — the handwriting reader is not the classifier; the private retrieval index is not the answer model. Here is exactly what happens at each step, and what a person still controls.

1. 01

   Document Intake

   Every document your business runs on is pulled in — clean PDFs, faxed and scanned pages, phone photos, handwritten forms, email attachments — from Google Drive, SharePoint, Box, or direct upload. Duplicates are detected, and every file gets a timestamped intake record before anything is read. There is no “supported formats” list to fight: if a person can read it, the engine takes it in.

   What you get One deduplicated set of every document — no matter how messy, or how many systems they were scattered across.

   • PDF & scans
   • Phone photos
   • Handwritten
   • Email
   • Google Drive
   • SharePoint
   • Box
2. 02

   AI + OCR Reading

   OCR and AI vision read text, tables, and stamped or handwritten content off documents that legacy OCR tools choke on — a faxed lease, a photographed form, a decades-old scan. Every value carries a confidence score, and the one line it cannot read confidently is flagged for a person rather than silently guessed. In a compliance or legal file, that difference — flag versus guess — is the whole game.

   What you get Clean, structured content from even the worst source files — with the one uncertain line flagged, never invented.

   • OCR
   • Vision models
   • Handwriting model
   • Confidence scoring
3. 03

   Classify across departments

   Each document is recognised for what it is — a commercial lease, a term loan agreement, a Form 1120, a SOC 2 report, a KYC file — and routed to the right department and schema. One engine covers legal, finance, compliance and tax rather than four siloed tools, which is why a single archive becomes searchable across every team at once.

   What you get Every document filed under the right department — so one question can span legal, finance, compliance and tax together.

   • Document classifier
   • Legal
   • Finance
   • Compliance
   • Tax
4. 04

   Private, cited knowledge base

   The content is indexed into a retrieval layer that lives inside your own environment. “Safe RAG” means exactly this: RAG is when the AI answers only from a specific set of documents instead of the open internet, and “safe” means those documents are yours, they never leave your boundary, and every passage stays linked to its exact source page. So the model cannot make things up, an auditor can trace any answer, and nothing is sent to a public AI service.

   What you get A knowledge base private to you, grounded in your own documents, that can cite every source it uses.

   • Secure RAG index
   • Your tenant only
   • Source-page citations
   • No public models
5. 05

   Ask in plain words → cited answer

   This is the payoff. Anyone on the team asks a question the way they would ask a colleague — “which commercial leases auto-renew before December?”, “any KYC files missing a 2024 refresh?”, “which entities filed a 1120 for 2024?” — and gets an answer in seconds, with every claim pinned to the exact document and page. Where a domain model applies, the same layer runs the calculation — a reserve study’s 30-year projection, a loan-covenant total — with every figure traced back to a source document.

   What you get Answers to plain-language questions across your whole archive — every one cited to the source page, ready to defend in an audit.

   • Plain-language questions
   • Claude reasoning
   • Cross-document
   • Source-page citations
6. 06

   Human review & deliver

   Nothing is filed, sent, or acted on automatically. Flagged lines and drafted answers go to a named person — a compliance lead, a partner, an analyst — who confirms or corrects before anything leaves the system, and every extraction, answer and calculation keeps a full audit trail back to the source document. Where you want a finished document out — a reserve study, an adjuster summary, an underwriting memo — it is produced in your own template.

   What you get A human sign-off on every judgment call, a full audit trail, and finished output in your own format.

   • Reviewer sign-off
   • Flagged items
   • Your report template
   • Full audit trail

The problem

The document processing problem it solves

Most businesses already have every answer they need — it is just trapped in documents. Contracts, filings, claims, returns and reports pile up across departments, and finding one fact means a person opening files one at a time. The bottleneck is not judgment; it is the hours of reading and searching before judgment can begin.

• ✕A compliance or legal team hunts through hundreds of contracts by hand to answer one question — which agreements auto-renew, which are missing a clause.
• ✕Finance and underwriting teams re-key data from appraisals, tax returns and bank statements — hours per file before any analysis starts.
• ✕Anything handwritten, faxed, or badly scanned falls outside what legacy OCR tools read reliably, so it stays manual.
• ✕Documents are scattered across Drive, SharePoint, Box and email, with no single place to ask a question across all of them.
• ✕Generic AI tools can answer, but they make things up, cannot cite a source, and send your confidential files to a public model — a non-starter in a regulated workflow.

The engine does not replace professional judgment. It reads and indexes everything first, so a plain-language question returns a cited answer in seconds — and a person still signs off before anything is filed.

Time to value

How fast you go live

Most document sets are live and answerable in 2–4 weeks.

1. Week 1Connect your documentsPoint the engine at where your documents already live — Google Drive, SharePoint, Box, email, or direct upload. It ingests and deduplicates across all of them; nothing has to be moved or re-filed.
2. Week 1–2Tune reading & classificationRun the reader on 20–50 real documents from your workflow — including your worst scans and handwriting — and validate accuracy against your own ground truth, department by department.
3. Week 2–3Build the private, cited indexStand up the retrieval index inside your own environment, so every answer is grounded in your documents and linked to its source page, with nothing sent to public AI. Where a domain model applies, we wire it in — even if it lives in a spreadsheet today.
4. Week 3–4Set questions, reviewers & go-liveConfirm the everyday questions each team will ask, who reviews flagged items before anything is filed, and any output templates you need. Run live documents end-to-end, sign off, and go live.

What you need to start

• ✓A representative sample of 20–50 documents — including your messiest scans and handwriting.
• ✓Read access to where documents live today — Google Drive, SharePoint, Box, or email.
• ✓The everyday questions each team needs answered, and who signs off on the answers.
• ✓Any output template or domain model you already use — even if it lives in a spreadsheet today.

Your documents never leave your environment. We validate accuracy on your actual files before go-live, and the retrieval index runs inside a tenant you control — which is what compliance and client-confidentiality agreements require.

ROI

The return on a Document Intelligence Engine

The cost is not the software — it is the hours your team spends reading and searching, and the risk of a missed clause or a filing that slips a deadline. When any question against your whole archive returns a cited answer in seconds, a compliance review that took days becomes an afternoon, and an auditor’s request is answered on the call. The reserve-study example — two engineers and two weeks compressed into hours — is the same pattern applied to one vertical: read everything once, then ask it anything, with a person signing off on the judgment calls.

Want your team’s number instead of the benchmark? Run the document processing cost calculator — your volume, your touch time, your staff cost, in ten seconds.

Proof

How we prove it — before you commit

FAQ

Document Intelligence Engine FAQ