Data Protection

What AI does in Airparser

Airparser is not a thin wrapper around a chat model. Extraction combines several stages — preprocessing, OCR, structured-extraction models, and post-processing/validation — and AI is used within that pipeline where it adds the most value: interpreting content that varies in layout or format and mapping it to the fields in your schema. There's no generic chat interface and no open-ended prompting of your data beyond the extraction task you've configured.

How a document is processed, step by step

• Ingestion — a document arrives by upload, email, API, or an integration into your Airparser mailbox.
• Preprocessing — the file is normalized, and OCR is run on scanned files or images to produce a text layer.
• Structured extraction — the document is matched against your schema using our extraction models and logic: a mix of our own models and enhanced/specialized models — built on top of the providers listed on our subprocessors page — tuned specifically for document data extraction.
• Post-processing — the extracted output is validated and normalized against your schema's field types before it's finalized.
• Delivery — the result is stored in your account and available via the app, API, or your configured export/webhook.

Why we combine OCR, AI, and post-processing

Each stage covers a different weakness. OCR alone gives you raw text, but no understanding of what that text means or how it maps to your schema. AI models are good at interpreting meaning and handling documents whose layout or wording varies — which fixed templates handle poorly — but on their own don't guarantee a well-typed, validated result. Post-processing closes that gap by checking types, formats, and consistency before data reaches your account. Combining the three gives more reliable results than relying on any single method.

Embeddings and long documents

Embeddings are numerical representations of text that let a system identify which parts of a document are most relevant to a given field or query, without having to process the entire document at once for every step.

We've used this technique to help process very long documents more efficiently. It's currently applied selectively and remains somewhat experimental rather than a default part of every extraction — we may expand its use going forward as it matures.

Our extraction models

Our extraction pipeline uses a mix of our own models and enhanced, specialized models built on top of the LLM providers listed on our subprocessors page — tuned specifically for document data extraction, not general-purpose chat.

Training and data retention

• We never train or fine-tune any model on your data, and we never sell your data. This is a standing commitment, not a plan-dependent feature.
• Our agreements with our AI subprocessors prohibit them from using data submitted for extraction to train their models.

Encryption and transport

All communication with our AI subprocessors happens over encrypted connections (HTTPS/TLS 1.2+), the same standard we apply to every other part of our infrastructure.

Data segregation

Each Airparser account's documents, schemas, and results are logically isolated from every other account. Processing one customer's document never exposes or mixes in another customer's data.

What stays under your control

• You are the Data Controller; Airparser and its AI subprocessors act only as Processors, on your instructions.
• You can delete any document, schema, or your entire account at any time.
• Configurable retention policies let you automatically delete processed data between 1 and 180 days.