Fax did not go away. Every prediction since 2005 has had it on the way out, and every year the volume metrics have ticked up. In healthcare, fax remains the dominant channel for prior authorization, referral exchange, and medical records release. In government, fax is the channel of last resort for claim submission, document evidence, and inter-agency exchange. In legal, fax is still acceptable service in a number of jurisdictions, and is often preferred.
The reasons fax persists are durable: healthcare data recognizes T.38 fax as a compliant transmission method, jurisdictional acceptance has long pre-dated email-with-attachment, and the cost of forcing a counterparty off fax is higher than the cost of accepting it. The traditional fax pain — manual sorting, lost documents, no audit — is real, and it is the pain that AI is finally well-positioned to address.
This post is a practical walkthrough of what we built into Fax AI and why. It's an engineering post; the architecture matters because regulated fax processing fails in specific, predictable ways and the architecture has to address each one.
The traditional fax pain
A traditional fax workflow looks like this: a fax arrives at a multifunction printer or a fax server. Someone walks over, picks up the page (or downloads the PDF), looks at it, and decides what it is — a referral, a claim, an authorization request. They route it to the right team mailbox or shared drive. The team picks it up, processes it, and updates whatever system of record they're using. Each step has manual handling. Each handoff is a place where the document gets misclassified, lost, or delayed.
The audit story is worse. The original fax was a TIFF or PDF image. The classification was a human decision with no recorded reasoning. The routing was a directory move. The processing was a screen-scrape into another system. If a regulator asks "what happened to the fax that arrived at 9:47am on Tuesday," the answer is somewhere in someone's email and someone else's memory.
What AI can actually do for fax
Three things, in this order: classify, extract, route.
Classification assigns the inbound fax to a known document type. For a healthcare payer, the categories might be prior authorization request, referral, claim attachment, medical records release, appeal, complaint. For a state agency, the categories might be initial claim, appeal, supporting documentation, request for benefits review. The classification is the precondition for everything downstream — without it, the fax can't be routed correctly.
Extraction pulls structured fields from the classified document. For a prior authorization, that's member ID, requesting provider NPI, procedure code, supporting diagnosis, urgency. For a claim, it's claimant identifier, claim number, document type, signature presence. The extraction goes into the system of record without retyping.
Routing sends the classified, extracted document to the right queue, with the right priority, to the right team. The routing is rule-based on the classification and extraction outputs.
Done well, this collapses what was a multi-day, multi-handoff process into something that completes in minutes for the routine cases and surfaces only the genuinely ambiguous ones for human review.
The architecture
The Fax AI pipeline has five stages. Each stage exists because skipping it produced specific, recurring failures in early deployments.
1. T.38 ingest with carrier-grade redundancy
An inbound fax design can use fax-over-IP and customer-controlled document storage. Carrier selection, retry behavior, handling of incomplete transmissions and access permissions must be configured and tested for the implementation.
2. OCR and structure recognition
Once a fax is fully received, it goes through Document AI for OCR and structure recognition. We use a layout-aware OCR pipeline rather than plain text extraction; the structure (form fields, tables, signature blocks, fax cover sheet identifiers) carries information that pure text loses. For multi-page faxes, structural continuity (a table that runs across pages) is reconstructed before the next stage sees the content.
3. Classification
A hybrid classification design can combine a smaller model for familiar document types with retrieval-assisted reasoning for less common cases. Confidence thresholds and review queues should be validated against representative documents.
4. Extraction
Field extraction runs against a per-document-type schema. Extraction also runs a confidence score per field. Required fields with low confidence trigger human review of just that field, not the entire document. Low-confidence extraction is the most common failure mode — a smudged member ID, a hand-written diagnosis — and the review path is designed to be fast (a focused panel showing just the field in context).
5. Routing and audit
The classified, extracted document is routed to the destination queue (a payer's prior auth team, an agency's claims processing group, etc.) along with a structured payload the destination system can ingest. Every stage's output is logged to the customer-owned audit sink: the inbound bytes, the OCR result, the classification with confidence, the extraction with per-field confidence, the routing decision, and the timestamp of each. The audit record is the answer to "what happened to the fax that arrived at 9:47am on Tuesday."
What human review actually looks like
The pipeline routes high-confidence faxes through automatically; low-confidence faxes go to a review interface. The review interface is the unsexy half of the product, and it's where most of the operational value lives.
The interface shows the original fax, the classification with the model's confidence and top alternatives, and the extracted fields with per-field confidence. The reviewer can accept the classification (one click), correct it (drop-down), or reject the document entirely (with a reason code). They can edit individual extracted fields without re-keying the whole record. Their decisions are logged and feed back into the model's evaluation set — over time, the confidence threshold for auto-routing can be relaxed for document types with high reviewer agreement.
Why audit pays for itself
The audit chain on a fax pipeline is not a compliance line item — it's an operational asset. When a counterparty calls and says "we faxed you the prior authorization at 11am, where is it," the support team can answer in seconds, with the document, the classification, the extracted fields, and the routing destination. When a regulator audits the agency's claim-handling, the answer to "show me how you processed inbound documents during the audit window" is a single query.
This is the point that gets lost in most fax automation pitches. The point isn't "fax is annoying, let AI handle it." The point is that AI gives you a fax channel that's morereviewable through the original document, extracted fields and recorded staff actions. The required coverage, access and retention depend on the implementation.
Fax is not going anywhere. Make the fax channel work like every other channel: classified, structured, audit-clean. The pain it causes goes away when the pipeline behind it is solid.
Where to start
If you're sitting on a fax channel that's costing your team manual processing time, the first conversation is about volume and document mix. Fifty faxes a day with three document types is a different deployment than five thousand a day with thirty. The architecture above scales across that range, but the deployment shape — fully managed vs. customer-operated, on-premise gateway vs. cloud T.38, full automation vs. AI-assisted-review — depends on the specifics.
We do free 30-minute walkthroughs of fax workflows, looking at your inbound mix and your downstream systems, and producing a reasoned recommendation. Schedule one here →.