# Intelligent document processing for invoices and contracts

> Intelligent document processing (IDP) uses OCR and language models to classify documents, extract fields and line items, validate them against your records and post the result to your systems. We build one document family, such as supplier invoices, to production with a human review queue for uncertain fields. Projects take 6 to 12 weeks and cost $12,000 to $40,000, plus per-page software fees.

- Page: https://methodcrew.com/services/intelligent-document-processing
- Last updated: 2026-10-28
- Publisher: [The Method Crew](https://methodcrew.com/)
- Service: Intelligent document processing
- Category: Automation

We build pipelines that read the documents your team keys in by hand, check every extracted field against your own records and send only the doubtful ones to a person.

## Key facts

- Price: $12,000 to $40,000 per project. Covers one document family to production, such as supplier invoices or one contract type. Per-page fees for the extraction service are separate and billed to you at cost.
- Timeline: 6 to 12 weeks
- Typical tools: Azure Document Intelligence, Amazon Textract, Google Document AI, Mistral OCR, OpenAI API, Anthropic API, Python, n8n, Postgres, Retool, QuickBooks Online, Xero, NetSuite
- Change requests: $175 an hour, quoted in writing first
- Model and tool usage: billed to your own accounts at cost, no markup
- Ownership: code, workflows, prompts, accounts and data belong to you

## Who it is for

- Finance teams keying 500 or more supplier invoices a month into QuickBooks, Xero, NetSuite or Sage
- Operations teams processing delivery notes, bills of lading, claims or application forms
- Legal and procurement teams that need key terms pulled from contracts into a register
- Companies whose document volume is growing faster than they want to hire

## Signals you need it

- **People retype what is already on the page.** Staff open a PDF, read a number and type it into another system. Error checking means someone glancing at it twice.
- **Month-end waits on a stack of documents.** Invoices or claims pile up in an inbox and get processed in batches, so your numbers always trail reality by a week.
- **A template-based OCR tool keeps breaking.** Every new supplier layout needs a new template, and someone spends hours a month maintaining them.

## What we hand over

- **Document inventory and labeled sample.** A count of document types, layouts and monthly volumes, plus 100 to 300 of your real documents with the correct value for every field, used to measure accuracy.
- **Extraction schema.** The fields to capture for each document family, with formats and validation rules: lines must sum to the subtotal, the tax ID must match the supplier record, dates cannot be in the future.
- **Production pipeline.** Intake from email, upload or a scanner folder, then classification, extraction, validation and posting to your accounting, ERP or contract system, running in your cloud account.
- **Review queue.** A screen where a reviewer sees the page beside the extracted fields, with low-confidence values marked, and approves or corrects them.
- **Accuracy report.** Field-level accuracy and straight-through rate measured on your labeled sample before launch, then tracked monthly in production.
- **Runbook and reviewer training.** How to add a supplier, change a field, handle a rejected document and read the accuracy report, plus a training session for reviewers.

## How the engagement runs

1. **Map** (1 to 2 weeks): We count your documents by type and layout, agree the fields that matter downstream and collect a labeled sample. You get the extraction schema and a target accuracy per field, which together become the acceptance test for the build.
2. **Prove** (2 to 3 weeks): We run competing extraction approaches against the labeled sample, typically a document AI service, a vision-capable language model and the two combined, then report field accuracy and cost per page for each. You choose on the numbers.
3. **Build** (3 to 7 weeks): We build intake, validation, the review queue and the write into your system of record. The pipeline then runs in parallel with your team on live documents for 2 weeks before it posts anything on its own.
4. **Run** (Monthly, optional): New suppliers and layouts arrive every month. On a care plan or AI ops retainer we track accuracy, review correction patterns and adjust prompts or rules before errors pile up.

Every engagement follows [The Method](https://methodcrew.com/approach): Map, Prove, Build, Run.

## What is intelligent document processing, and how is it different from OCR?

Intelligent document processing is the full pipeline that turns a document into checked, usable data: classify it, extract the fields, validate them and post them where they belong. OCR is one step inside it. OCR turns an image of text into characters. It does not know that the number at the bottom right is the invoice total, or that the total should equal the sum of the lines.

| | Basic OCR | Template extraction | Intelligent document processing |
|---|---|---|---|
| Output | Raw text | Fields from fixed positions | Named fields and line items with confidence |
| New layouts | Not relevant | A new template each time | Usually handled without a template |
| Validation | None | Format checks | Business rules and lookups against your records |
| Uncertain values | Not flagged | Often wrong without warning | Sent to a review queue |
| Best for | Making scans searchable | A few stable forms | Many suppliers, layouts or document types |

Current IDP pairs a document AI service, such as Azure Document Intelligence, Amazon Textract or Google Document AI, with language models that read a page much as a person would. The model handles the variety. The validation rules decide whether to trust it.

For the short definition and related terms, see the glossary entry on [intelligent document processing](https://methodcrew.com/glossary/intelligent-document-processing). For a worked example on supplier invoices, from inbox to accounting system, read the [invoice processing pipeline blueprint](https://methodcrew.com/blueprints/invoice-processing-pipeline).

## How accurate is AI at extracting data from invoices?

Accurate enough to post most clean invoices without review, and never accurate enough to skip validation. The real answer depends on your documents, so we measure it on them before you commit to a build.

During Prove we run extraction on 100 to 300 of your labeled documents and report accuracy for each field separately, because a single blended score hides the field that matters. Header fields such as supplier name, invoice number, date and total are usually the easiest. Line items are harder, especially multi-page tables, merged cells and descriptions that wrap. Invoices with several tax rates are another common source of errors.

What lowers accuracy:

- Poor scans, such as phone photos taken at an angle, faxes and low-resolution images.
- Handwriting. Current vision models read neat handwriting reasonably well. Scrawl, stamps over text and crossed-out values go to review.
- Unusual layouts, such as sideways pages or invoices with several separate tables.
- Number and date formats from other countries, for example 1.234,56 or 03/04 meaning April.

What raises it: asking for fewer fields, validating against your supplier master and purchase orders, and sending each document type to whichever method scored best on it.

The figure your finance team will care about is the straight-through rate: the share of documents that post with no human touch and every value correct. We report it next to field accuracy and agree a target with you during Map.

## What happens when the AI is not sure about a field?

It stops and asks a person. Every extracted value carries a confidence signal, and every document passes validation rules before anything is posted. If either check fails, the document goes to the review queue with the doubtful fields marked.

Typical validation rules for supplier invoices:

- Line items add up to the subtotal, and subtotal plus tax equals the total.
- The supplier exists in your vendor master, matched on tax ID or bank details.
- The invoice number has not been seen before for that supplier, which catches duplicates.
- Any purchase order number exists, is open and matches the amount within your tolerance.
- Bank details match the ones on file. A change always triggers a hold, because altered bank details are a common sign of invoice fraud.

The reviewer sees the page on one side and the fields on the other. Clicking a field jumps to where it sits on the page, and the reviewer corrects or approves it. Every correction is logged. Once a month we review the log, and where a pattern shows up, such as one supplier's layout tripping the same field, we adjust the prompt or add a rule. That loop is [human-in-the-loop review](https://methodcrew.com/glossary/human-in-the-loop) doing its job.

Thresholds are yours to set. A finance team might let invoices under $2,000 from known suppliers post on their own and review everything above that.

## How much does intelligent document processing cost?

Building one document family to production costs $12,000 to $40,000 and takes 6 to 12 weeks. A document family is one kind of document with one destination, such as supplier invoices into NetSuite or employment contracts into a contract register. Additional families on the same pipeline cost less than the first, because intake, review and monitoring already exist.

The price moves with:

- Layout variety. 20 suppliers with similar invoices sit at the low end. 800 suppliers in 6 languages sit at the top.
- Line items. Header fields alone are simpler than full line-item extraction with product codes matched to your catalog.
- The destination. Posting to Xero or QuickBooks Online through their APIs is straightforward. An on-premise ERP with batch imports takes longer.
- Review design. Using an approval tool you already have costs less than a custom review screen with role-based access.
- Data residency and retention rules.

Running costs are separate and paid by you: the extraction service charges per page, model calls are billed per use and hosting is usually small. Published per-page prices for IDP software run from about $30 to $560 per 1,000 pages, depending on vendor and volume, according to a September 2026 comparison by Parseur, itself an IDP vendor ([Parseur IDP pricing](https://parseur.com/blog/idp-pricing)). We estimate your monthly running cost from your own volumes during Map, so you can set it against the hours spent keying documents today.

## When is IDP the wrong choice, and what is not included?

When the data already exists in structured form. If suppliers can send e-invoices as UBL or Peppol XML, or your customers can fill out a web form, parse that directly. It costs a fraction of an IDP build and has no extraction errors. Ask your 10 largest suppliers before you commission anything.

IDP is also hard to justify when:

- Volume is low. A few dozen documents a month rarely repay a $12,000 build. A shared inbox and a checklist may be enough.
- Every document is unique and needs judgment, such as one-off contracts under negotiation. We can pull key terms into a register, but deciding whether a clause is acceptable is a lawyer's job. The [contract review blueprint](https://methodcrew.com/blueprints/contract-review-assistant) shows where that line sits.
- People mainly want to ask questions across thousands of policies or manuals. That calls for a [knowledge system](https://methodcrew.com/services/rag-knowledge-systems) that answers with citations.

Not included in an IDP project:

- Scanning hardware, mailroom services or digitizing a paper archive.
- Records management and retention policy design.
- Legal advice on contract terms.
- Per-page software fees and model usage, which you pay directly at cost.
- Cleanup of your vendor master or chart of accounts beyond what validation needs.

## What changes for your team

- Documents are read within minutes of arriving by email, upload or scanner
- Clean documents post to your system of record without anyone touching them
- Reviewers handle only the fields the pipeline is unsure of, with the source page beside them
- Every posted value can be traced to the page and line it came from
- Field accuracy is measured every month, so you know how far to trust the output

## Tools we work with

Azure Document Intelligence, Amazon Textract, Google Document AI, Mistral OCR, OpenAI API, Anthropic API, Python, n8n, Postgres, Retool, QuickBooks Online, Xero, NetSuite

## What it costs

$12,000 to $40,000 per project. Covers one document family to production, such as supplier invoices or one contract type. Per-page fees for the extraction service are separate and billed to you at cost.

- Change requests: $175 an hour, quoted before work starts
- Model and tool usage: billed at cost, no markup

All 20 price bands: https://methodcrew.com/pricing

## Blueprints that use this service

- [Invoice processing pipeline](https://methodcrew.com/blueprints/invoice-processing-pipeline): built in 6 to 9 weeks, $15,000 to $35,000 per project.
- [Contract review assistant](https://methodcrew.com/blueprints/contract-review-assistant): built in 8 to 12 weeks, $20,000 to $40,000 per project.

## Common questions

### Which document types can you process?

Most business documents a person can read: supplier invoices, receipts, purchase orders, delivery notes, bills of lading, bank statements, insurance claims, application forms and contracts. We build one document family at a time, starting with the one that has the highest volume and the clearest destination system. Email bodies and attachments come in through the same intake.

### Do we need to train a custom model?

Rarely. Prebuilt document AI models and current language models handle most invoices, receipts and forms without training. We use your labeled sample to choose and tune prompts and rules, which is faster to build and cheaper to maintain than a trained model. Custom training can make sense for very high volumes of one unusual form, and we will show the cost difference if it applies.

### Where are our documents stored and processed?

In your own cloud account. Intake, storage and logs run in your Azure, AWS or Google Cloud tenancy, or in your existing document store, in the region you choose. Extraction services and model APIs are called from there under business terms that keep your data out of model training by default. We document the data flow for your security review before go-live.

### How long before invoices post without anyone touching them?

Usually 2 to 4 weeks after go-live. For the first 2 weeks the pipeline runs in parallel with your team and posts nothing on its own. Then we switch on automatic posting for the groups that met the accuracy target, typically known suppliers with clean layouts, and widen it as the correction log shows each group holding up.

### Will it work with our accounting system?

If the system has an API or a reliable import, yes. QuickBooks Online, Xero, NetSuite, Sage Intacct and Microsoft Dynamics 365 Business Central all accept bills through APIs. Older on-premise systems often take a scheduled CSV or XML import, which works but adds a step. We check the posting method before we quote, then confirm it with a test post during Map.

### What about suppliers who send e-invoices?

Then you may not need AI for those suppliers at all. Structured e-invoices in formats such as UBL, Peppol or Factur-X can be parsed directly with no extraction errors. We route them through a simple parser and send only PDFs and scans to the IDP pipeline. As more suppliers move to e-invoicing, the share that needs AI extraction falls, and running costs fall with it.

### Can IDP catch duplicate or fraudulent invoices?

It catches the common cases. Validation checks each invoice number against earlier ones from the same supplier, compares bank details with your vendor master and flags totals that do not match the purchase order. A change in bank details always holds the invoice for a person. It does not replace your payment approval controls, and we do not run security audits.

## Related services

- [AI automation](https://methodcrew.com/services/ai-automation): Workflows where a language model reads and sorts, and your existing systems do the rest. $4,000 to $35,000 per project.
- [RAG and knowledge systems](https://methodcrew.com/services/rag-knowledge-systems): Your documents turned into an assistant that cites its sources. $15,000 to $75,000 per project.
- [AI ops monitoring and improvement](https://methodcrew.com/services/ai-ops-monitoring): Quality checks, cost control and model updates for live AI systems. $1,500 to $5,000 per month.

## Further reading

### Guides

- [How to measure the ROI of AI automation, running costs included](https://methodcrew.com/insights/measure-ai-automation-roi)

### Terms explained

- [Chunking](https://methodcrew.com/glossary/chunking): Chunking is splitting documents into smaller passages before they are embedded and indexed, so a retrieval system can find and return the exact part that answers a question.
- [Fine-tuning](https://methodcrew.com/glossary/fine-tuning): Fine-tuning is further training of an existing language model on your own examples, so it learns a specific format or task without long instructions every time.
- [Human-in-the-loop (HITL)](https://methodcrew.com/glossary/human-in-the-loop): Human-in-the-loop is a workflow design where a person reviews, corrects or approves an AI system's output at set points before it reaches a customer or a system of record.
- [Intelligent document processing (IDP)](https://methodcrew.com/glossary/intelligent-document-processing): Intelligent document processing uses OCR and AI models to read documents such as invoices, contracts and forms, extract the fields you need, check them and post them into your systems.

## Next step

Request a 30-minute scoping call: https://methodcrew.com/contact. Or email contact@methodcrew.com.