ComPDF AI
Document ProcessingComPDF AI is the document-understanding side of ComPDF's Cloud API. Agents extract named values and table columns from invoices and forms, run OCR over scans, detect tables and stamps, and straighten photographed pages before filing them.
What This Integration Enables
Most document connectors assume a clean file born from a system. ComPDF AI assumes the opposite: that the page was photographed on a dock, that it is curved and lit badly, and that a seal in the corner may matter more than any sentence on it. That is why its first useful move is to fix the input rather than read it. Straighten Scan squares up a photographed page and Clean Up Scan corrects the color and contrast, and running those first makes everything after them more accurate than pointing recognition at a curved page and hoping.
The second thing it does differently is skip the schema. Extract Fields takes the names a person would use, the words actually printed on the form, and returns those values without a template anyone has to define or maintain. Parse Document is the broad counterpart, returning the whole structure of a page as headings, paragraphs, tables and images, and Recognize Tables returns a table's cells as data when the table only ever existed as a picture. FlowRunner agents chain those into ordinary work: correct the page, ask for the values, check the numbers against a system of record, and hand a person the small pile of automation exceptions that needs a decision rather than the whole stack.
Without FlowRunner
With FlowRunner
Use Case Scenarios
-
Trade paperwork that arrived as a phone photo
An inbound container's paperwork reaches the mailbox as four images taken on a warehouse floor. The agent runs Straighten Scan on each, then Clean Up Scan, and stores the corrected pages with Download Result so the readable version is the one that gets filed. Extract Fields then asks for Exporter, Consignee, Country of Origin, Invoice Number and Gross Weight by name, with no per-exporter template involved. The values reconcile against the open purchase order in Acumatica, and the receiving record in ShipBob picks up the confirmed weight. Nothing here needed anyone to define what an invoice from this particular supplier looks like.
-
The table that only exists as an image
A packing list arrives as a scanned page with a twenty-row table on it, and the rows are the whole point. Recognize Tables finds the table and returns its cells, which the agent writes straight into Google Sheets as a reconciliation tab against what the purchase order said would ship. Line quantities that disagree get flagged in Slack with the row and the source page, so a coordinator argues with a supplier about a real discrepancy rather than retyping twenty rows to find it. Extract Text handles the same page when what is wanted is not a table but the position of every piece of text, since its JSON output carries where each fragment sits.
-
Documents that are supposed to carry a seal
A compliance sweep runs over the documents filed against each shipment in the last week. For each one that is supposed to be stamped, the agent runs Detect Stamps and records whether a seal was found, where it sits, and the confidence reported. Documents with a clear seal pass. Documents with no seal, or with a weak detection, are corrected with Straighten Scan and retried, and if the answer does not change they go to a person along with the page image. The sweep costs a few credits, which Get Credit Balance is checked against before the batch runs, and it replaces a person opening every file to look at a corner of it.
Human-in-Loop Highlight
Detect Stamps reports where a seal is and how confident it is. It has no way to report that a document should have one. So when it comes back empty, the agent is not holding a result, it is holding the beginnings of an accusation: this certificate is unsealed, reject the paperwork. That is worse than a wrong number, because the correction operations change what the detector sees, and the same photographed page can produce a detection after Straighten Scan that it did not produce before. FlowRunner's human-in-the-loop step is where that claim stops. The agent posts to the trade compliance channel: "Certificate of origin for container MSKU 4417732 returned no stamp detection, before and after correction. The corrected page is attached. Confirm the document is unsealed and hold the container, or mark it sealed and release for filing." The compliance lead is looking at the page, which is the only place the answer was ever going to be.
Agent Capabilities
11 actionsUnderstanding
4- Parse Document Returns a document's whole structure as machine-readable data, including headings, paragraphs, tables and images. The broad first move when a flow does not yet know what kind of document it is holding.
- Extract Fields Pulls named values and table columns out of a document with no schema and no template. Name the values the way the page prints them, and name table headings separately.
- Extract Text Reads the text off a scan or an image, either as JSON carrying where each piece of text sits on the page, or as plain text when only the words matter.
- Recognize Tables Finds the tables on a page and returns their cells, as data and as HTML. Used when the rows are the payload and the table only ever existed as a picture.
Image Correction
3- Straighten Scan Flattens and squares up a photographed or crooked page so it reads like a scan. Worth running before anything else on a page shot by hand.
- Clean Up Scan Corrects the color and contrast of a scan so the text stands off the page. Pairs with straightening on documents photographed under poor lighting.
- Detect Stamps Finds seals and stamps on a page and reports where each one sits and how confident the detection is. The operation this page's human gate exists for.
Results
4- Download Result Downloads a processed file and stores it in file storage. ComPDF returns a download URL that does not last, so this is how a corrected page is kept.
- Get Task Reads a task, including the download URL of each file once it has finished.
- List Tasks Lists recent tasks on the account. Used for reconciliation and for tracing what a flow actually submitted.
- Get Credit Balance Reads how much credit is left on the account. Needs no identifiers, so it doubles as a connection check and as the guard before a large batch runs.
Frequently Asked Questions
What can FlowRunner do with ComPDF AI?
FlowRunner agents can run Parse Document, Extract Fields, and Extract Text in ComPDF AI, plus 8 more actions.
Does connecting ComPDF AI to FlowRunner require OAuth?
No. ComPDF AI connects to FlowRunner with an API key, no OAuth flow required.
Can ComPDF AI trigger a FlowRunner workflow automatically?
ComPDF AI doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Other Document Processing integrations
Start building with ComPDF AI
Free plan, no card required. Connect in minutes.