AI Document Processing

Two different jobs called document processing

When people ask for “AI for our documents” they usually mean one of two things, and the build is different for each.

	Pull fields out	Find answers
The question	What are the renewal date, notice period and liability cap in this contract?	Which of our supplier contracts let us terminate with 30 days notice?
What you get	Structured data written into your CRM, ERP or register	An answer in plain language with page citations
Works on	One document at a time, as it arrives	The whole archive at once
Typical use	Contract registers, certificate tracking, forms, delivery notes	Legal and procurement questions, policy lookups, due diligence
Built with	A workflow and a language model	A search index plus a language model (private RAG)

Most teams start with the first job because it removes typing every day. The second job pays off once the documents are already flowing through the first, because the index is built as they arrive.

What the output looks like

Here is what the workflow returns for one supplier agreement. The supplier is made up; the shape is what lands in the review queue:

{
  "document_type": "supplier_agreement",
  "supplier": "Brightline Logistics Inc",
  "effective_date": { "value": "2025-03-01", "page": 1 },
  "initial_term_months": { "value": 24, "page": 2 },
  "auto_renewal": {
    "value": true,
    "page": 2,
    "quote": "shall renew automatically for successive twelve (12) month periods"
  },
  "notice_period_days": { "value": 90, "page": 2 },
  "renewal_notice_deadline": "2026-12-01",
  "payment_terms_days": { "value": 45, "page": 4 },
  "liability_cap": { "value": "fees paid in the prior 12 months", "page": 7 },
  "governing_law": { "value": "Delaware", "page": 9 },
  "flags": ["liability cap below your standard of 2x annual fees"],
  "route": "legal review: non-standard liability cap"
}

Two things matter here. Every value points to a page, so the reviewer checks it in seconds instead of rereading the contract. And the workflow worked out the date you care about, the last day to give notice, and set a reminder for it.

Which documents work well, and which do not

Good candidates, because they are typed and follow a pattern:

• Contracts, amendments and order forms
• Supplier certificates: insurance, ISO, compliance declarations, with their expiry dates
• Delivery notes, packing lists and bills of lading
• Onboarding and application forms
• Bank and card statements
• Claims and incident forms

Hard cases, where you should expect a person to stay closely involved:

• Handwriting, including handwritten notes on typed forms
• Dark or blurred phone photos of crumpled paper
• Anything where validity depends on a stamp or signature. The model can tell that a signature box is filled in; it cannot tell you the signature is genuine.
• Long contracts changed by later amendments. Working out which clause is current needs the retrieval job and, for anything important, a lawyer.

Accuracy: what to expect and how it is controlled

Vendor pages quote figures like 99% accuracy. Ignore them, including any we might be tempted to give you. Accuracy depends on your documents, your fields and how clean your scans are, and it differs from field to field: a supplier name is easy, a liability cap buried in a long clause is harder.

So we control accuracy through the design of the workflow:

• Every field carries its page reference, so checking is fast.
• Rules catch impossible values: an end date before a start date, a missing required field, a supplier that is not in your master data.
• Nothing is written to your systems until a person has approved it, at least for the first months.
• Before go live we run the pipeline on a sample of a few hundred of your real documents and report accuracy field by field, so you decide with your own numbers.

A worked example of the hours

Measure your own numbers first. Here is the arithmetic for one example operations team handling delivery notes and supplier certificates.

• 1,200 documents a month, about 6 minutes each to open, read and type up: 120 hours a month.
• With the workflow, a clean document takes about 1.5 minutes to check. Say 20% are flagged and still take 6 minutes.
• That is about 24 hours for clean documents and 24 hours for flagged ones: 48 hours a month.

The team gets back about 70 hours a month. At a loaded cost of $40 an hour that is close to $2,900 a month, so a $6,000 build pays back in a little over two months. The missed renewals and expired certificates the reminders catch are not in that number, and for some teams they are worth more than the hours.

Keeping documents private

Contracts, HR files and patient or client records often cannot go to a public model. In that case the language model and the search index run on hardware you control, and the pipeline is the same. This is the private RAG setup, and it is common in legal work and regulated industries.

How we build it

One document type into one system is a fixed-scope AI automation project: $3,500 to $7,000, live in two to three weeks with monitoring and a runbook. Several document types, private models or a search assistant across the whole archive is a larger build in the $8,000 to $18,000 range. If you are not sure which documents to start with, a readiness audit picks the one with the fastest payback.

Invoices follow the same pattern with accounting rules on top. That is covered in AI for accounts payable and the invoice automation setup guide. For delivery notes and supplier paperwork, see supply chain, or browse every process on the processes overview.