What Is Intelligent Document Processing? IDP vs Plain OCR
By Schema LawPublished Updated
Intelligent document processing, or IDP, is a vendor term for software that goes past reading a document to deciding what it is and pulling chosen facts out of it. OCR turns a page image into text. Intelligent document processing adds classification, extraction and validation on top, so the output is labeled data instead of a page of words.
It is different from the OCR you already use in the way a filing clerk differs from a typist. The typist gives you the words. The clerk tells you it is a deed of trust, notes the lender and the amount, and flags that the date looks wrong.
What is intelligent document processing?
Intelligent document processing is a label for systems that take in documents and return structured information: the type of each document and the values of named fields within it. It is a marketing category. Neither vendor page quoted in this section points to a standard that says which features a product must have to use the name.
The fair way to define such a term is to quote those who sell under it.
Amazon's explainer, titled What is Intelligent Document Processing?, puts it this way: "Intelligent document processing (IDP) automates the process of manual data entry from paper-based documents or document images to digital form, for integration with other digital business processes."
Google does not use the three-letter label in the overview of its product, but its Document AI overview describes the same job. It calls the product "a document processing and understanding platform that takes unstructured data from documents and transforms it into structured data (specific fields, suitable for a database)".
Two things stand out in both descriptions. The goal is data in fields, not text on a page. And the thing being replaced is a person keying values from a document into a system. The trade-offs of that replacement are the subject of the post on OCR vs manual data entry.
Open one complaint and one invoice after the software assigns labels. Search the recognized words for the plaintiff, the claim, and the damages figure. A wrong label can park the damages number in a date field. The page picture still shows the amount that was printed.
Where the term comes from in a law office
A lawyer may meet "IDP" on the website of a contract review tool, an intake product, an e-billing system or a document management add-on. In those settings it is a signal that the product claims to do something with the content of a document beyond making it searchable. What exactly it does is for the vendor to show you.
How is IDP different from plain OCR?
Plain OCR reads. It answers "which characters are on this page?" and stops. IDP reads and then interprets. It asks what kind of document the page belongs to, which characters are the values someone wants, and whether those values make sense. Each added step is a judgment, and each can be wrong on its own.
Google's overview states the two layers in one sentence. Document workflows, it says, "involve getting the raw text from documents, then extracting specific text from that which corresponds to the data needed (the fields or entities)."
Take a scanned closing binder from a fictional refinancing by Marguerite Achterberg with Calloway Savings Bank. It holds a promissory note, a deed of trust, a settlement statement and a title commitment, in one 140-page PDF.
| Plain OCR | A system sold as IDP | |
|---|---|---|
| What it returns | 140 pages of text and tables | A list: four documents, their types, their page ranges, and fields from each |
| Example of output | "...the principal sum of $412,000.00..." | Document type: promissory note. Principal: 412,000.00. Borrower: Marguerite Achterberg |
| What it decided | Which characters are on the page | Which document this is, which number is the principal, whether it matches the deed of trust |
| An error it can make | $412,000.00 read as $417,000.00 | The escrow total recorded as the principal |
| How you catch it | Compare text with the image | Read the document and ask whether this is the right value |
The reading layer is still there underneath. Amazon's explainer lists OCR first among the technologies used in IDP and says it "converts an image of text into a machine-readable text format." A wrong character at that layer flows into every later step. The post on contract data extraction explains why OCR and extraction are two different steps and how to tell which one failed.
Where the line blurs
Some capabilities sit between the two. Reading a table as rows and columns is more than character recognition and less than extraction. It recovers structure without deciding what any cell means. The post on table OCR covers that middle ground. Detecting whether a checkbox is ticked is similar, and has its own name, covered in the post on optical mark recognition.
Schema Law sits on the reading side of this line. It reads pages into text and tables and exports them in the formats listed on the homepage. It does not sort documents by type or hand you a list of named fields, and it is not an IDP product.
The stages vendors group under IDP: capture, classify, extract, validate
The vendor descriptions cited in this post can be mapped onto four stages. Capture brings the document in and reads it. Classification decides what kind of document it is. Extraction pulls out named values. Validation checks those values against rules or other data. Names, order and coverage vary by product.
| Stage | Question | Input | Output |
|---|---|---|---|
| Capture | What does this file say? | Scans, PDFs, images, emails | Text, layout, tables |
| Classify | What kind of document is this? | Text and layout | A type label, sometimes page ranges |
| Extract | What are the values we want? | Text, layout and the type | Named fields with values |
| Validate | Do the values hold up? | Fields, rules, reference data | Accepted values, or exceptions for review |
Capture
Capture covers getting the file into the system and reading it. OCR lives here, along with whatever image cleanup and layout analysis the product runs. Google's overview lists this first among the things its platform does: "Digitize documents using OCR to get text, layout, and various add ons".
Classify
Amazon's explainer joins capture and classification into one step: "The first step in IDP is capturing and classifying documents." It gives the reason classification comes early: the type determines the later processing steps for each document. A promissory note and a settlement statement need different fields pulled from them, so the system has to know which it has.
Extract
Amazon again: "After classification, the system extracts relevant data from the documents", identifying "specific information such as dates, amounts, or names." Extraction can be built in more than one way. Google's Document AI documentation, for example, lists template-based extraction, custom-based extraction and a custom extractor with generative AI as separate mechanisms.
Validate
The same page describes validation by example: "the system might cross-reference extracted data with existing databases or use predefined rules to check for errors." A rule might say that a principal amount must be a positive number, that a maturity date must fall after a signing date, or that a party name must match a client record.
The four-stage list is a simplification
The two vendors quoted here do not draw the boxes identically. Google's overview sorts its components into three categories and puts them in a different order: "Digitize: OCR", "Extract", and "Classify". Amazon's explainer continues past validation to data processing, continuous learning and reporting. When you read a vendor's diagram, map its boxes onto these four questions and note which ones are missing.
Document classification: sorting pages by type before anything is extracted
Document classification assigns each document, or each page, to one of a fixed list of types, such as lease, invoice, court order or correspondence. It runs before extraction because the type decides which fields to look for. Where one file holds several documents, classification may also mark where each begins and ends.
What the classifier is choosing from
A classifier does not discover types. It picks from a list someone gave it. Microsoft's documentation for its custom classification model says training one "requires at least two distinct classes and a minimum of five document samples per class."
That has a consequence for legal files, which are varied. A document that fits none of the listed types still has to go somewhere: into the nearest wrong type, or into a catch-all class if the system has one. Ask which.
Page by page, and splitting
A scanned closing binder or a production volume is one file with many documents in it. Microsoft's page says its models "perform classification of an input file one page at a time to identify the documents within and can also identify multiple documents or multiple instances of a single document within an input file."
Finding the boundaries is called splitting. Google's documentation on document splitters adds a detail that is easy to miss: "Splitters identify page boundaries, but don't actually split the input document for you." The output is a list of page ranges with types. Cutting the PDF is a separate act.
What classification hands on
Google's page on its custom classifier describes the handoff: "You typically would use a custom classifier on documents that are different types, then use the identification to pass the documents to an extraction processor to extract the entities."
So a classification error is not cosmetic. If the settlement statement in the Achterberg binder is labeled a promissory note, the system looks for a principal amount and a maturity date on a page of closing costs. It may find numbers and dates that fit the pattern and report them with a straight face.
Signs of a classification problem
- Fields that are empty for a whole document, because the wrong set of fields was looked for.
- Values of the right kind from the wrong place, such as a recording fee reported as a loan amount.
- A document count that does not match the binder's index.
- A boundary that falls mid-document, so the signature page of one instrument is attached to the next.
- Exhibits and riders treated as separate documents, or swallowed by the one before.
Microsoft's classifier returns a confidence score with its result, and the page tells developers: "To set the threshold for your application, use the confidence score from the response." In a product built on a score like that, someone chose the threshold. It decides how many doubtful documents are passed through and how many are held for a person.
Where does a person still check the result in an IDP workflow?
A person checks at the points where the system is unsure or a rule fails, and on a sample of everything else. The vendors' own documentation describes human review as part of the design. In legal work the reviewer also has to catch what the system was sure of and got wrong, which no score will flag.
The three review points vendors describe
Low-confidence results. Microsoft's guide to accuracy and confidence scores says: "For scenarios where accuracy is critical, confidence can be used to determine whether to automatically accept the prediction or flag it for human review." Its list of tips includes "Consider incorporating human review into your workflows."
Random samples. Amazon's documentation on using Amazon Augmented AI for human review describes "human review of low-confidence predictions or random prediction samples", and says reviewers can step in when a model cannot make a high-confidence prediction "or to audit its predictions on an ongoing basis." The same page gives an example close to legal work: "extracting information from scanned mortgage application forms can require human review due to low-quality scans or poor handwriting." That page also notes the service is no longer open to new customers. It is cited here for its description of the pattern, not as a product to adopt.
Validation exceptions. When a value breaks a rule, such as a date out of order or a total that does not add up, the item is routed to a person.
What those review points miss
All three depend on the system knowing it might be wrong. A value read confidently and wrongly, which also passes the rules, reaches no reviewer. A principal of $417,000.00 in place of $412,000.00 is a positive number and a plausible loan. The post on the OCR confidence score explains why a high score is not a promise.
Random sampling is the only one of the three that can catch such an error, and only at the rate you sample.
A review plan for a legal matter
- Decide which fields carry consequences. Amounts, dates, party names, governing law, notice periods. Review every one of those against the page image, whatever the score.
- Review every exception. A failed rule is the system telling you where to look.
- Review every low-score item, and ask the vendor what the cutoff is and who set it.
- Sample the rest. Pick a fixed share of accepted items at random and check them. Record what you find. If the sample turns up errors, widen it.
- Check classification separately. Compare the list of documents and page ranges with an index or a manual count before trusting any field.
- Keep the link to the source. A reviewer who has to hunt for the page will check less. Prefer tools that open the page at the spot each value came from.
Who reviews is a professional question as much as a technical one. Rules on supervising staff and technology differ by jurisdiction, so check your own before deciding which checks can be delegated.
Reading an IDP claim on a legal technology vendor's website
Treat "intelligent document processing" on a vendor's site as a heading, not a description. The label tells you the product claims to do more than OCR. It does not tell you which stages it performs, on which document types, with what checks. Ask for each stage to be shown on your own documents.
Translate the adjectives into stages
For each claim on the page, ask which of the four questions it answers.
| The site says | Ask |
|---|---|
| "Understands your documents" | Does it classify? Into which types? What happens to a type it has not seen? |
| "Automatically extracts key data" | Which fields, from which document types? Can I define my own? |
| "Works on any document" | Show me a handwritten rider, a faxed order and a two-column lease |
| "Human-level accuracy" | Measured how, on what documents, by whom? Can I run my own test set? |
| "Human in the loop" | Whose human? Mine or the vendor's? What triggers review? |
| "Learns over time" | From whose documents? Are mine used to train a model others use? |
Questions for each stage
Capture
- Can I see the recognized text for a page, separately from the extracted fields?
- What happens with handwriting, stamps and poor faxes?
Classification
- What is the list of document types, and can I change it?
- Is a multi-document PDF split, and can I correct the boundaries?
- Is there an "unknown" class, and how are those documents handled?
Extraction
- Does each value link back to its place on the page?
- When a field is not found, is it left blank or filled with a guess?
- Can the tool tell me that two candidate values were found?
Validation and review
- Which rules run, and can I add my own?
- Is there a score per field, and where is the cutoff?
- Is there a record of who reviewed and changed what?
Confidentiality
- Where are documents sent at each stage, and do other companies process them?
- Are documents or corrections used to train any model?
- How long is everything kept, and can I delete it?
Run a small test
Pick ten documents from a closed matter that you are permitted to use, and build an answer key by hand: the type of each and the correct value of each field you care about. Run them through the product. Score classification and each field separately. Ten documents will not give you a statistic, but they will show you which stage is weak, and whether the product tells you when it is unsure.
A vendor that welcomes this test and shows you the recognized text, the labels and the source of each value is describing a product. A vendor that answers every question with the three-letter label is describing a category.