Document data extraction — Pull the data out of PDFs, photos and scans instead of retyping it
Document data extraction uses AI to read documents and pull out specific fields, like names, dates, totals, part numbers or sizes. It works on PDFs, scans, phone photos and many handwritten notes. Staff review the results, especially anything the system is unsure of, instead of typing every field.
By Julia Ja, Owner
Read from a supplier invoice
PDF · 2 pages| Item | Qty | Unit price |
|---|---|---|
| Copper fitting 3/4" | 40 | $2.15 |
| PEX pipe 100 ft | 6 | $58.00 |
| Shutoff valve | 12 | $9.40 |
Matched to purchase order 5521. One price differs from the order and is flagged.
Send to the bookkeeper to approveWhy it matters
The problem it solves
A lot of business data arrives as documents, not data. Invoices, purchase orders, bills of lading and tax forms come as PDFs, scans and phone photos. Someone has to read each one and type the fields in.
The typing is slow, and the errors are quiet. A wrong quantity or total often isn't noticed until a shipment, an invoice or a return is wrong. Photos and handwriting take longest, so they sink to the bottom of the pile.
Example · illustrative
A freight broker's billing inbox
- By 3 pm a freight broker's billing inbox holds bills of lading and delivery photos from a dozen drivers.
- Each file is read, and the load number, delivery date, piece count and signature are pulled out.
- One photo was taken at an angle, so its piece count is marked unsure.
- The billing clerk checks that one field against the photo and confirms it.
- The other loads are updated and ready to invoice.
How it works
How document data extraction works
- Starts whenA document arrives by email, upload, WhatsApp or scanner
- 1Step 1Read the documentThe file is read whatever its form: a typed PDF, a phone photo at an angle, or a handwritten list.
- 2Step 2Pull out the fieldsThe fields you care about are picked out and put in a fixed structure, such as a line-item table.
- 3Step 3Check the resultsTotals are added up, dates and codes are checked against your records, and unclear readings are marked.
- A person decidesConfirm what's unsureA person checks the marked fields against the original, side by side, and corrects them.
- 5Step 5Send it onClean data goes into the accounting, order or case system, with the original file attached.
Before · by hand
- Staff open each document and type the fields into a system.
- Typos in a quantity or a total are found much later, if at all.
- Photos and handwriting are the slowest to deal with, so they wait.
With document data extraction
- Fields are filled in before anyone opens the document.
- Totals and codes are checked automatically.
- Staff only look closely at the fields marked as unsure.
- The original stays attached to the record for checking later.
What we build
The parts of document data extraction
Your systems
- QuickBooks Online
- NetSuite
- Google Drive
- Microsoft 365
- Airtable
- Gmail and Outlook
What we build
1Document intake
Collects files from the email inbox, upload page, chat or scanner folder where they arrive today.
2Field definitions
The exact fields to pull from each document type, agreed with you, such as line items, totals, dates or load numbers.
3AI reading
Reads typed, scanned, photographed and handwritten documents and puts the fields into a fixed structure.
4Checks against your data
Adds up totals, checks dates and codes against your records, and marks readings it's unsure of.
5Side-by-side review screen
Shows each marked field next to the original document so a person can confirm or correct it.
6System hand-off
Sends confirmed data to your accounting, order or case system, with the original file attached.
Your team decides
- A person confirms any field the system marks as unsure.
- Staff decide what to do when a document doesn't match an order or record.
- The business decides which fields matter and which checks to run.
Industries it fits
What document data extraction looks like in each industry
Glass processing & fabrication
Glass sizes, quantities and edge work are read from customer photos, sketches and screenshots into an order. Sizes that look wrong for the glass type are marked for the office to check.
Everything for glass processing & fabricationYour industry isn’t listed, or works differently? Describe it and get an answer in a minute.
Getting started
What you need first
We go through these with you in the audit, before anything is built.
Before we start
4 things to have ready- Real sample documents of each type, including the messy ones, so testing reflects what you actually get.
- A list of the fields you need from each document, and where each field should go.
- Access to the system the data goes into, such as your accounting, order or case software.
- A person who reviews marked fields, and a rule for when a document may skip review.
How a rollout runs
- 1Collect samplesWe gather recent documents of each type, including the hard cases, and agree on the fields to pull from each.
- 2Build and scoreWe set up the reading and checks, run them on your samples, and compare the results field by field with the right answers.
- 3Connect review and systemsWe connect the inbox or upload source, the review screen and the target system, and test the full path.
- 4Run alongside manual entryFor a period, staff keep typing while the system runs, so differences show where it still makes mistakes.
- 5Switch and tuneStaff move from typing to reviewing, and checks are tightened for the fields that cause trouble.
What to measure
How to know it's working
Weekly report · document data extraction
IllustrationField accuracy by document type
Shows where reading is reliable and where review must stay in place.
Fields marked unsure
Shows how much review work is left, and which fields cause it.
Corrections after data was saved
Errors found later are the costly ones. This shows whether checks and review catch enough.
Documents waiting for review
A growing queue means the review step needs more time or clearer rules.
Documents that can't be read
Very poor scans or new formats may need a different way in, or a request for a better copy.
We measure the baseline before building, so the change can be compared afterwards.
When it’s not a fit
- It isn't perfect. Very blurry photos, faded faxes and messy handwriting need more human checking, and sometimes a call to the sender.
- If documents vary wildly and arrive rarely, setup may not pay off. Start with the document types you handle most.
- Dropping the review step too early is the common mistake. Keep a person on marked fields until results hold up on your own documents.
- Checks need clean reference data. If your item or customer list is out of date, matching will be too.
Tool or custom build
Do you need a custom build?
Option A · a ready-made tool
An off-the-shelf tool may be enough
Tools like Dext for receipts and bills, Rossum for invoices, or Microsoft's Azure AI Document Intelligence handle standard business documents well. They suit one common document type flowing into one system.
Option B · built around your systems
When a custom build is worth it
A custom build makes sense for photos, handwriting, mixed languages or documents specific to your trade. It also fits when data must be checked against your records before saving.
Not sure which fits?
What it typically costs
Typical ranges as of October 2026. Not a quote.
Ready-made tool, set up by us
An existing platform on your own account, configured for your business.
- Setup, one timepaid to Lorryn
- $1,000–$2,000
- Tool subscriptionpaid to the provider
- $25–$100 / month
Fits when your process matches what the platform already does.
Custom build
Built around how you work, with no platform subscription. You own it.
- Build, one timebasic version to full version
- $2,000–$9,000
- Running costshosting and usage, on your accounts
- ~$10–$40 / month
Fits when your systems, rules or exceptions don't match a ready tool.
Optional Automation Care: $125 · $250 · $375 / month
1, 2 or 3 hours of support a month, with monitoring and alerts included. Without it, you run it yourself after handover.
Typically works with QuickBooks Online, NetSuite, Google Drive, Microsoft 365.
Booking into or writing to your field-service, practice or shop software adds up to $500 to the ready-made tool (often a built-in feature) and $500–$1,500 to a custom build.
Your price depends on volume, the systems involved and how many exceptions need a person. Our team confirms a fixed price before any work starts.
Proof and who stays in charge
From our results · Glass processing plantLiveA glass processing plant reads order sizes from WhatsApp photos, handwriting and screenshots, in Azerbaijani, Russian and Turkish.Read the case studyWho does what
The automation does the routine steps. A person makes the call.
- 1Read the documentAutomated
- 2Pull out the fieldsAutomated
- 3Check the resultsAutomated
- 4Confirm what's unsureA person
- 5Send it onAutomated
Consent and privacy
Documents with patient data should be processed only by tools whose vendors sign a HIPAA business associate agreement (BAA).
FAQ
Questions about document data extraction
Often, yes. Clear handwriting and reasonably sharp photos usually read well. Messy notes may need more fields checked by a person.
Checks catch many errors, like totals that don't add up. Fields it's unsure of are marked for a person to confirm before anything is saved.
Usually not. Modern AI reads layouts it hasn't seen before. We test on your real documents and tune the rules for the formats you get most.
Yes, many languages work. We test with your actual documents to make sure the results are reliable.
Into the system you already use, like your accounting, order or case software, or a spreadsheet if that's where the work happens today.
We set it up on business plans that don't use your documents for training. Files stay in your storage, and access is limited to the people who need it.
Yes. Most businesses start with the type they type in most, then add others once the results hold up.
Related use cases
Often built together
- Intake & documentsOrder entry automationTurns orders from email, PDFs, chat and photos into entered orders, matched to your items and prices.
- Billing & paymentsSupplier invoice processingReads supplier invoices from email, matches them to orders, and drafts bills in your accounting software.
- Intake & documentsDocument classificationWorks out what each incoming document is and who it belongs to, then files and routes it.
Want document data extraction for your business?
Pick the step that suits you. Every path starts with how the work runs today, and we'll say honestly whether it's worth automating.
- Right now · any hour
Ask Lorryn AI
Describe your process and get a first answer in a minute: what to automate, what it connects to, and a rough estimate.
- Free · 15 or 30 minutes
Talk to our team
Bring one process to a video call. Leave knowing whether it’s worth automating and how.
- Free · 6 of 15 spots left
Get a free written plan
Free for businesses with 3 or more employees, while spots last (normally $250).
Claim a free audit9 of 15 claimed