natif.ai
Document ProcessingSubmit documents to natif.ai workflows for invoice extraction, receipt extraction, and OCR on the German intelligent document processing platform. Agents turn inbound paperwork into typed fields ready for approval.
What This Integration Enables
natif.ai is a German intelligent document processing platform, and the thing that makes it unusual is that human verification is part of its data model rather than something you bolt on afterwards. Every extracted field arrives as a uniform wrapper: the value, a confidence score between zero and one, a validation_problem flag, an optional note, and bounding-box references pointing back to where on the page it was read. Then Verify Result marks a processing run as human-verified, recording who confirmed it and when, and Get Verification Status reports that state back so a downstream step can refuse to run until a document has actually been reviewed.
That is a platform designed by people who assumed a human would be in the loop, which is the same assumption FlowRunner is built on. Work is submitted to named workflows rather than to a generic endpoint: invoice_extraction, receipt_extraction, OCR, splitting, and whatever else the account has. The connector exposes the full result surface, so an agent can pull just the business fields, the raw OCR text, the hOCR with bounding boxes, the named entities, the detected language, the page images, or a searchable PDF. The human-in-the-loop step in the middle is not an interruption to that pipeline. It is a field in it.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Straight-through posting with an honest exception path
Scanned supplier post lands in an intake folder. The agent submits each document to the invoice_extraction workflow. Submit Document long-polls, so with the default wait it usually returns the finished result inline rather than handing back a processing ID to chase. The agent reads Get Extractions and splits the fields on their confidence scores. A document where every field clears the threshold and no validation_problem is flagged posts straight into NetSuite. A document with a shaky total or a supplier name the model was unsure of does not. It becomes a review task, and the reviewer sees the extracted value next to the page image the value was read from.
Splitting a scan batch nobody separated
The mailroom scans a stack of paperwork as one continuous PDF. The agent submits it to a splitting workflow, which returns the batch as separate documents, then routes each one to the extraction workflow that fits it. Get Detected Language runs first on the ones that came from overseas suppliers, so a document in French reaches the team that handles French suppliers rather than sitting in an English-language queue for a day. List Page Images and Download Page Image give the reviewer a page-by-page view when a split looks wrong.
Producing an archive that is actually findable
After extraction, the agent calls Download Processed PDF to produce a version with a searchable text layer and stores it in FlowRunner file storage. For records subject to retention, it requests PDF/A and sets DPI and greyscale to keep the file size sane on a large scan. The archived document goes to S3 alongside the extracted fields and the verification record, so a document pulled up three years later comes with the values that were read from it and the name of the person who confirmed them.
Human-in-Loop Highlight
Confidence scores are only useful if something acts on them, and most pipelines that receive them do not. natif.ai gives an agent both halves of the loop. Get Extractions returns a score and a validation_problem flag on every field, so the agent knows exactly which values the model was unsure about, and bounding-box references tell it where on the page each one came from. When a field drops below the threshold the flow sets, the agent does not average it away or accept it because the other fourteen fields were fine. It stops.
What it sends is specific: the reviewer gets the field name, the value natif.ai read, its confidence, and the cropped page image pulled with Download Page Image showing that exact region, with the question stated plainly. "The total on this invoice from Hansen Verpackung read as 4,180.00 with low confidence, and the tax line flagged a validation problem. Is the total 4,180.00 or 4,780.00?" The reviewer answers in Slack. Then the agent does the part that makes this different from a chat message: it calls Verify Result, which records inside natif.ai who confirmed the document and when, and the posting step downstream gates on Get Verification Status rather than on a flow variable. The approval does not live in a chat thread that scrolls away. It lives on the document, in the platform, where an auditor asking who signed off on a corrected total gets an answer.
Agent Capabilities
14 actionsProcessing
2- Submit Document Downloads a document from a URL and submits it to a natif.ai workflow, which decides what gets extracted: invoice fields, receipt fields, plain OCR, splitting and so on. natif.ai accepts no URL input of its own, so the bytes are fetched and forwarded as a multipart upload. The call long-polls, so with the default wait it usually returns finished results inline and only falls back to a processing ID when the document is still running.
- Get Processing Result Retrieves the results of a submitted document by processing ID. natif.ai reports state through the HTTP status rather than a status field: finished, still running, or unknown ID. This route long-polls too, so each poll waits rather than hammering the API.
Results
5- Get Extractions Returns just the extracted business fields, without OCR text or image artifacts. Every field arrives as a uniform wrapper carrying the value, a confidence score between zero and one, a validation_problem flag, an optional note, and bounding-box references back to the page. This is the operation a confidence-routed flow is built on.
- Get OCR Text Returns the raw OCR output with layout and word positions, before any workflow-specific field mapping. The diagnostic for telling a misread word apart from a correctly read word assigned to the wrong field.
- Get hOCR Returns the OCR result as hOCR, the HTML microformat carrying recognised text with bounding boxes and reading order. Used when the text needs rendering back over the page image, or when a downstream tool expects hOCR rather than JSON.
- Get Named Entities Returns the people, organisations, dates, amounts and similar entities natif.ai recognised, independently of the workflow's own field schema. Useful for documents that do not fit a fixed template, such as correspondence, where structure is wanted without defining every field up front.
- Get Detected Language Returns the language detected in the document. The cheapest possible routing signal for multilingual inbound post, applied before any extracted content is looked at.
Artifacts
4- List Page Images Returns the page-image URLs for a processed document, one per page. Used to discover the page count before pulling individual pages, or to build a page-by-page review view.
- Download Page Image Downloads one rendered page and stores it in FlowRunner file storage, returning a URL the flow can pass on. Supplying a width or a height scales it down with aspect ratio preserved. This is how a reviewer sees the actual page a questionable field was read from.
- Download Thumbnail Downloads the document's thumbnail and stores it in FlowRunner file storage. A cheap visual preview to attach to an approval step or a notification without pulling a full-resolution page.
- Download Processed PDF Downloads the processed document as a PDF with a searchable text layer and stores it in FlowRunner file storage. JPEG quality and DPI trade size against fidelity, greyscale or black-and-white shrink scans, physical page dimensions can be fixed in millimetres, and PDF/A output covers long-term archiving. Page width and height must be supplied together.
Verification
2- Verify Result Marks a processing result as human-verified, recording who confirmed it and when. Called after a reviewer has checked low-confidence fields, so natif.ai holds an audit trail of which extractions were signed off rather than accepted automatically.
- Get Verification Status Returns whether a result has been human-verified, with the email of the person or system process that verified it and the timestamp. The gate a posting step reads before it moves money against a document.
Workflows
1- List Workflows Returns the workflows available to the account with ID, display name, description and kind. Workflow IDs use underscores, so invoice_extraction rather than invoice-extraction, and that is the form Submit Document expects. Filtering by kind narrows the list to extraction workflows, splitters and so on. Results are cursor-paginated from the last workflow ID of the previous page.
Frequently Asked Questions
What can FlowRunner do with natif.ai?
FlowRunner agents can run Submit Document, Get Processing Result, and Get Extractions in natif.ai, plus 11 more actions.
Does connecting natif.ai to FlowRunner require OAuth?
No. natif.ai connects to FlowRunner with an API key, no OAuth flow required.
Can natif.ai trigger a FlowRunner workflow automatically?
natif.ai doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with natif.ai
$100 in credits. No card required. Connect in minutes.