OCR form processing: a step-by-step guide to OCR for forms

OCR form processing reads filled-in forms and turns each field into data. OCR, short for optical character recognition, is the part that reads the text. Forms such as loan applications, tax forms and insurance certificates are typical. Plain OCR only turns the page into text. Form processing goes further. It matches each value to the right field and reads handwriting and checkboxes. Then it checks the values and sends them to your other systems.

The sections below sort forms by layout, follow one form through the software, and end with what to ask a vendor.

Structured vs semi-structured forms#

How a document is laid out decides how hard it is to read automatically.

• Structured forms

  Every copy has the same fields in the same place, like Form 1003, Form W-2 and ACORD 25. Templates can read them, as long as the layout and the scan stay put.

• Semi-structured forms

  The same information in a different layout from every sender, like invoices, bank statements and pay stubs. Fields have to be found by their labels, not their position.

• Unstructured documents

  Free text with no fields, like letters and contracts. Values come out of sentences, which takes language models rather than templates.

Even a standard form varies. Employers send Copy A of Form W-2 to the Social Security Administration (SSA). It is printed in red dropout ink, which scanners ignore. A substitute copy must match the layout exactly, because SSA scanners read the form automatically. Borrowers give lenders the copies that employees receive. These copies can leave out unused boxes. They can also move the name and address boxes or use a vertical layout. IRS rules allow all of this.

The four kinds of content on a form#

A single form can mix four kinds of content, and each has its own reader. OCR reads the printed text. Intelligent character recognition (ICR) reads hand-printed characters in boxes. Handwriting recognition (HWR) handles cursive notes. Optical mark recognition (OMR) decides which checkboxes and bubbles are marked. Form processing software uses all four readers on the same page.

Good form design helps: one box per character, a request for capital letters, and boxes in a dropout color. Handwriting still reads less reliably than print, so values the engine isn't sure about should go to a person. Our guide to intelligent character recognition covers the handwritten part in depth.

How OCR form processing works#

Most form processing systems run the same stages. Systems differ in how well each stage handles a form that looks different from the ones it was set up with.

Follow one borrower's packet through the system. It arrives by email as one PDF with a W-2 and page 3 of a Form 1003. One page is a skewed phone photo. The system straightens and cleans each image first. Then the system finds the form type on each page and splits the packet into separate forms.

Next, each page goes to the reader it needs. OCR reads the printed text, ICR the hand-printed entries, handwriting recognition the notes and OMR the checkboxes. Each value is then tied to a field such as borrower name. On a fixed form, the system can find a value by where it sits on the page. On forms that vary, it uses the field name next to the value and the text around it. Template OCR starts to fail at this step.

Validation comes last. Dates must be real dates, tax IDs must have the right format, totals must add up and required fields can't be blank. A value below the confidence threshold goes to a person. The rest go to your system through an API, or as JSON, CSV or Excel files.

Form processing use cases#

These forms come up every day in lending, tax and insurance work, in onboarding and in healthcare. Form-like documents such as pay stubs belong here too.

• Mortgage applications

  Form 1003, the Uniform Residential Loan Application, and Form 1008, the Uniform Underwriting and Transmittal Summary, read alongside the income documents in the loan file.

• IRS tax forms

  W-2s, 1040s and 1099s often arrive in one mixed packet. Form 4506-C is the borrower's signed consent for a lender to get IRS transcripts.

• ACORD insurance forms

  Standard forms from ACORD, the insurance industry's standards body. ACORD 25 is the certificate of liability insurance. ACORD 125 is the commercial insurance application, and 126 is the commercial general liability section.

• Pay stubs

  Gross pay, deductions, net pay and year-to-date totals, in a different layout for each payroll provider. See pay stub extraction.

• Onboarding and KYC forms

  Account opening forms, vendor forms and W-9s, where the name, address and tax ID have to match other records.

• Healthcare and benefits forms

  Patient intake forms, often partly handwritten, plus claim forms and explanation of benefits (EOB) statements. See healthcare.

Template OCR vs AI form processing#

Traditional form OCR, often called zonal or template OCR, reads each field from a fixed spot on the page. It works well on one clean, fixed layout and needs a new template for every other. AI form processing finds each field by its label and context, so one setup covers many layouts.

What to look for in OCR form processing software#

Most form recognition software reads a clean printed form well. The differences show on handwriting, unfamiliar layouts and the values the engine isn't sure about.

• Handwriting and checkboxesReads hand-printed entries, cursive notes and ticked boxes in the same pass as printed text.
• New layouts without new templatesCopes with new form editions, substitute forms and skewed scans.
• Classification and splittingRecognizes each form in a mixed packet and splits it before extraction.
• A confidence score on every fieldLets you set per field how sure the engine must be before a value skips review.
• Validation rulesChecks formats, required fields and totals, and flags every value that fails.
• Review and outputShows reviewers each flagged value next to its place on the page, and sends approved data to your systems by API or webhook.

Test any tool on a sample of your own forms, including the worst scans and the messiest handwriting, before you commit.

How to automate form processing with Docsumo#

Docsumo, our product, is an intelligent document processing platform for this kind of work. It reads printed and handwritten text on any form. On the Business plan it sorts and splits mixed packets. It extracts fields and tables, and joins tables that run across pages into one. Fields it isn't sure about go to your own reviewers. You set the confidence threshold per field, clicking a field highlights its source line on the page, and corrections improve the model.

Checked data goes to your systems through the API and webhooks. Docsumo runs in the cloud only, and it isn't a desktop tool for editing PDFs or making searchable files. See the platform for the full workflow.

• 99%99%field-level accuracy across 250+ document types
• 95%+95%+of documents processed straight through, without manual review
• <5 min<5 minper document, down from 2+ hours

Play video: A 5-minute guide to OCR software

When a form is one document in a larger file#

We'd test form OCR on a full packet, not one clean form. A W-2 can be read perfectly and still disagree with the 1003 in the same file. Checks that compare documents catch that, and they matter more than one error-free page. Cross-document validation is on the Docsumo Enterprise plan.

OCR form processing turns a filled-in form into data your systems can use, with little or no manual data entry. Templates can read structured forms until a new edition comes out or the scan quality changes. Semi-structured documents and handwriting need software that finds each field from its label and the text around it, and flags values it is unsure about. Choose a tool by how it handles your worst forms, and have people review only the values it flags.

Book a demo with a few of your own forms, or start a free trial.

Frequently asked questions#