What is document AI?
Document AI is the use of machine learning and AI vision to read documents the way a person does: find the supplier, the order number, the dates, the line items and the totals, whatever the layout. Analysts call it intelligent document processing (IDP); vendors say AI document processing, document understanding or AI OCR.
Classic OCR turns an image into characters. It does not know which number is the invoice total and which is the phone number, so someone has to draw a template for every layout. AI that understands the page needs no template, so a new supplier or a moved field changes nothing.
But reading was never the whole job. The job is a correct record in the system your business runs on.
Why intelligent document processing stops at extracted data
The output is a file
A typical IDP platform ends with JSON, a CSV or a spreadsheet. The data is extracted, but nothing in your ERP or claims system has changed yet.
The system of record has no API
Many companies run their core work on an AS/400 (IBM i), a mainframe or a custom application with no import door. The only way in is the screens, so people still key the extracted data by hand.
The integration becomes the project
Closing that gap means middleware, new host programs, or RPA bots that break when a screen changes. Reading was the easy part; the last mile costs the most.
Document processing automation, from inbox to confirmed record
legacy bridge is an AI agent, delivered as a cloud service, that runs the whole chain. Automated document processing here means six steps, with a person in control of every submission.
Documents arrive
By email (a forwarding address, IMAP or Gmail), from an SFTP or FTP folder it polls, or by API upload.
AI vision reads them
Fields and line items are extracted with no OCR templates. A file holding several records is split into one record each, and a document your team already keyed is caught as a duplicate instead of being entered twice.
Values are checked
Against the document itself, against your process rules written in plain English, and against what your system already holds.
It fills your screens
It signs on, navigates the menus, fills the right fields, handles lookups and function keys, and reads the host's own messages.
A person approves
By default each prepared transaction waits for an employee to approve it in a web console or from an email link. Once trusted, a workflow can run on its own and people handle only the exceptions.
The record is confirmed
It reads the record back from the host and logs the result.
What document AI reads
legacy bridge reads PDFs, scans and phone photos (JPG and PNG), faxes, Word files (.docx) and Excel files (.xlsx). Typed and hand-filled forms are read by AI vision. The full list of document types legacy bridge processes covers the common ones:
- Invoices and remittances
- Purchase orders
- Sales orders
- Insurance and medical or dental claims
- Work orders and repair orders
- Bills of lading and delivery notes
- Policy changes
- Faxed and scanned forms
- Multi-record batch files
How document extraction AI is checked before anything is entered
A confidence score is not enough. A model can be confident and wrong, and a wrong value in a system of record costs far more than one held for review. So legacy bridge checks what it read against evidence, and asks a person when the evidence does not settle a value.
- The document agrees with itself: line items add up to the printed total, and dates make sense
- Your process rules are applied, written in plain English by your team, not coded by a developer
- Values are compared with what the host already holds, so a duplicate is caught before it is keyed
- Unclear, missing or handwriting-corrected values are flagged, never guessed: the record is held with the exact field and the document image
- If the host returns an error, your employee sees the error and the screen, and nothing is submitted
- A person's correction is remembered, so the next document like it goes through
OCR templates vs IDP software vs legacy bridge
| Criterion | legacy bridge | Typical IDP platform | OCR templates |
|---|---|---|---|
| Layout changes | Reads any layout by meaning, no templates | Pre-trained models; new types may need training | One template per layout; breaks when a field moves |
| Validation | Totals, dates, your plain-English rules and the host's own records | Confidence scores and configured rules | Field formats at most |
| Gets data into the legacy system | Enters it through your existing screens and reads it back | Exports JSON or CSV; the integration is yours to build | Exports text; people re-key it |
| Human review | Approval before submission; unclear fields held with the image | Review queue for low-confidence fields | Manual spot checks |
| Audit | Every keystroke and screen, SHA-256 hash chain, replayable | Extraction logs; entry happens elsewhere | Little or none |
| Pricing model | $99/month plus a fee per record filed and confirmed | Usually per page read, plus the integration project | License plus template upkeep |
The last mile: document AI that finishes the job
The last mile is getting checked data into the system of record. legacy bridge connects the way a terminal emulator does, over TN5250, and types into the same screens a person uses. Nothing is installed on the host. No changes to RPG or COBOL programs, no new APIs, no database access. A host that is not exposed to the internet is reached through a secure SSH tunnel.
It is not traditional RPA either: no desktop bots, no recorded mouse coordinates. It reads screens by meaning, so a moved field does not break it (see AI agents vs RPA). And it is finished document automation, not a toolkit you build workflows in.
In production today it runs on IBM i (AS/400, iSeries, System i), including custom RPG and COBOL applications and packaged ERPs. See AS/400 document processing. The same engine also speaks TN3270 for IBM Z mainframes, VT for OpenVMS, Unix and Pick, and HP 6530 for HP NonStop, and it can drive legacy web portals and desktop applications over Citrix or RDP. Those are scoped per workflow with the same controlled proof; see legacy system automation.
An IDP tool reports how much it read. legacy bridge reports which records your system confirmed, and bills only for those.
Who uses this document automation
Accounts payable
Supplier invoices read, matched to your rules and posted on the host. See invoice processing automation.
Order desks
Customer orders keyed into the order screens with lookups handled. See purchase order automation.
Claims teams
Claim forms, including hand-filled ones, filed with each line checked. See insurance claims processing.
Start with one real document: a controlled proof on your own document and screens that stops before final submission, so your employee approves or rejects. The company behind legacy bridge is part of scandiweb, serves 500+ clients including Fortune 500 companies, and is certified ISO 9001, ISO/IEC 27001 and ISO/IEC 27017.
