Home›Blog›AI for Handwritten and Faxed Enrollment Forms
Published: November 20, 2026Last updated: November 20, 2026

AI for Handwritten and Faxed Enrollment Forms

KC
Kaustubh ChandrabhanCo-founder & CTO, Zelthy
Share:
AI for Faxed Enrollment Forms: a handwritten form passes a check gate, with approved fields and one flagged for review

Most patient support enrollments still begin as a handwritten form that a prescriber's office faxes to a hub, and every field on that page has to become clean data before anything else can happen. Fax has not gone away: in 2025, a majority of US hospitals still used mail or fax to send summary of care records at least sometimes (40% often, 34% sometimes), and 81% used it to receive them, according to ASTP/ONC hospital interoperability data.

Every misread field turns into a callback, and callbacks cost days. Specialty pharmacies in a 2022 JMCP focus-group study reported roughly 2 to 3 days to fill a clean prescription and about 5 days when one needed intervention. Intake is the first step of the patient journey described in Zelthy's complete guide to patient support programs, and errors made there carry into every later step.

Zelthy is an AI-native platform-as-a-service, built on the open-source Zango framework, that pharmaceutical teams use to configure patient services workflows such as enrollment, consent and eligibility checks for programs deployed on the platform. This post covers why faxed handwriting is hard for machines, how an intelligent document processing (IDP) pipeline works, what the 2026 accuracy evidence actually shows, and how to design a review step around the fields where an error does the most harm.

Why are faxed, handwritten enrollment forms hard for AI to read?

Faxed enrollment forms combine low-resolution images, mixed handwriting and print, checkboxes, signatures and multi-page layouts that change with every form version. Each of those breaks a different step of automated reading, so no single model setting solves all of them.

The main problems:

  • Resolution. The Group 3 fax standard specifies a normal vertical resolution of about 3.85 lines per millimetre, with 7.7 lines per millimetre as the optional finer mode, and 1,728 picture elements across the page (CCITT T.4, published as RFC 804). In practice that is roughly 200 dots per inch across and as little as 100 lines per inch down, well below a typical document scan.
  • Skew and noise. Pages fed crooked, speckle from the transmission and stamped headers sit on top of the handwriting.
  • Checkboxes and marks. A tick, a cross, a filled square and a circled option all mean "selected", and a scribbled-out box means the opposite.
  • Signatures and dates. Consent depends on a signature being present and dated, not on reading the name.
  • Multi-page packets. A form, an insurance card copy, chart notes and a cover sheet often arrive as one transmission and must be split and matched to one patient.
  • Tables. Researchers extracting values from scanned clinical reports found that tables were a major failure source because column headers were lost during text segmentation (Hsu et al., JAMIA Open, 2022).

Missing information is a separate problem from misreading. An IDP pipeline can flag a blank field instantly. It cannot invent the missing value, and it must not try.

How does intelligent document processing read an enrollment form?

Intelligent document processing turns a fax image into validated, structured fields through a chain of steps: clean the image, classify the pages, read the text and handwriting, extract named fields, check them against rules and outside sources, and score each field so that uncertain values go to a person.

IDP pipeline for faxed enrollment forms: clean, classify, read and validate steps, with flagged fields routed to human review
A faxed packet moves through cleanup, classification, reading and validation before any field reaches the case record.
  1. Image cleanup. De-skew, de-speckle and contrast adjustment.
  2. Classification and splitting. Identify which pages are the enrollment form (and which version), the insurance card, the prescription and the clinical notes.
  3. Reading. Optical character recognition for print and handwriting recognition (sometimes called ICR) for handwritten entries. Many 2026 pipelines use a vision-language model that reads and extracts in one pass.
  4. Field extraction. Map what was read to a schema: patient name, date of birth, prescriber NPI, drug, dose, insurance member ID, consent checkboxes.
  5. Validation. Format rules, check digits, cross-page consistency and lookups against reference data.
  6. Confidence and grounding. Score each field and link it to the region of the image it came from, so a reviewer can see the source next to the value.
  7. Human review and audit. Route fields below threshold, or above a risk tier, to a reviewer, and log who accepted or changed what.

Steps 5 and 6 are where most of the safety comes from. Grounding makes each value checkable against the image; confidence decides which values are worth checking.

How accurate is AI at reading handwritten medical forms?

Accuracy on real handwritten clinical forms is good but not good enough to skip review on high-risk fields. Published 2026 studies report strong averages that hide much weaker results on dosing instructions, formatted values like dates and numbers, and messy handwriting, and the models rarely signal when they are unsure.

Three pieces of evidence are worth knowing:

  • Handwritten medication lists. In a May 2026 study of 33 handwritten lists, the best GPT-based model scored a median 38.5 out of 40 elements. But that model misread dosing and frequency in 35% of "as needed" entries and 27% of entries with timing instructions, against 3% for simple entries, and ciprofloxacin 500 mg was occasionally misread as 30 mg. The authors concluded that "human oversight remains essential to ensure medication safety."
  • Real handwritten forms. An April 2026 benchmark (a preprint, not yet peer reviewed) tested frontier vision models on 49 real maternity case records. The best median field accuracy was 0.85, hallucination rates ranged from 0.06 to 0.23, and accuracy on formatted fields such as dates and numbers was as low as 0.27 for some models. The authors noted that "models do not flag uncertainty" when a field cannot be interpreted.
  • The precision trade-off. An older JAMIA pipeline for handwritten form fields reached 94.6% positive predictive value but only 13.5% sensitivity. High precision is easy if the system is allowed to skip most fields.

Treat any vendor's single accuracy number as a starting question. Ask for field-level accuracy on your own forms, broken out by handwritten versus printed entries, with the error rate measured on fields the system auto-accepted.

How should you design the human review step for enrollment fields?

Design the review step by field risk, not by document. Decide in advance which fields can be accepted automatically when they pass an independent check, which always need a person, and which only need sampling, then measure the error rate on what was auto-accepted.

The table below is a starting framework for a US hub enrollment form. Thresholds should be calibrated on your own labelled forms.

Field-risk gate: low-risk fields take the auto-accept path, while identity, dose and consent fields go to human review
Low-risk fields pass on a calibrated threshold. Identity, dose and consent fields go through a person.

Four design rules make the table work in practice:

  • Validate against something the model did not produce. An NPI is a 10-digit identifier with a check digit in the 10th position, so a single misread digit usually fails arithmetic before it reaches a registry lookup. Drug codes can be checked against the FDA NDC Directory; note that FDA has finalized a uniform 12-digit NDC format taking effect March 7, 2033, so build validation that can handle both.
  • Treat free text as a safety channel. Notes scribbled in a margin can describe a reaction. Zelthy's post on pharmacovigilant PSP architecture explains why AI should flag those records for review and never close them.
  • Measure what you auto-accepted. Sample auto-accepted fields every week and track the field-level error rate, the share of forms needing any human touch, and time from fax receipt to a complete case.
  • HIPAA applies to the whole pipeline. HHS guidance treats a cloud provider that maintains ePHI as a business associate even if it cannot view the data, so any external OCR or model service processing enrollment forms needs a business associate agreement. The Security Rule also requires audit controls and integrity protections, which is why reviewer corrections should be logged with the original extracted value. Zelthy's HIPAA Security Rule guide for patient portals covers those safeguards in more depth.

Benefits verification and copay checks depend on the same clean fields.

What this looks like on a configurable platform

On Zelthy, teams can configure enrollment intake as a workflow: field schemas per form version, validation rules, review queues by risk tier, consent capture and an audit trail, connected to the rest of the patient services record. Because programs deployed on Zelthy keep ownership of their application code, review rules and thresholds can be changed by the program team as evidence accumulates rather than waiting on a vendor roadmap.

The results reported for patient support programs deployed on Zelthy include 60% faster onboarding, 30% cost reduction and a 45% improvement in adherence. Those figures reflect configured end-to-end program workflows, not document AI alone, and they should not be read as an extraction accuracy benchmark. Across deployments, the platform supports more than 1 million patients in more than 12 countries. For the wider program context, see the Zelthy patient services overview.

Fax will not disappear soon. CMS has finalized electronic standards for claims attachments, with compliance due by May 26, 2028, and CMS Administrator Dr. Mehmet Oz announced it by saying "The 1980s called, and they want their fax machines back." That rule covers claims documentation. It does not address patient support enrollment, so hubs should plan for faxed forms for years yet. This article is general information, not legal advice.

If your program is still re-keying faxed enrollment forms, book a demo with Zelthy to see how a field-risk review workflow can be configured for your forms.

Frequently asked questions

Can AI accurately read handwritten faxed forms?

Yes, for many fields, but not uniformly. A 2026 benchmark on real handwritten clinical forms found the best models reached about 85% median field accuracy, with much weaker results on dates, numbers and dosing instructions. High-risk fields still need validation or human review.

What is intelligent document processing in healthcare?

Intelligent document processing is software that turns document images such as faxes and scans into structured, validated data. It combines image cleanup, page classification, OCR and handwriting recognition, field extraction, rule-based validation and confidence scoring, with uncertain fields routed to a person.

Which enrollment form fields should always get human review?

Handwritten drug, dose and frequency entries, signatures and consent dates, and any patient identifier that does not match across pages should always be reviewed. Free-text notes should be checked for possible adverse events. Fields like an NPI can be auto-accepted only when they pass an independent check such as the check digit and a registry name match.

Is using AI to process enrollment forms HIPAA compliant?

It can be, if the pipeline meets HIPAA requirements. HHS treats a cloud service that maintains ePHI as a business associate even if it cannot view the data, so a business associate agreement is needed, and the Security Rule requires audit controls and integrity protections. Log every reviewer correction alongside the original extracted value.

How do you set confidence thresholds for document extraction?

Set thresholds per form type and per field, using your own labelled forms, and validate them on documents the system has not seen. Then sample auto-accepted fields regularly and track their error rate. A high overall accuracy figure does not justify a looser threshold on high-risk fields.

Sources

  1. Methods used by hospitals to engage in interoperable exchange. ASTP/ONC, HealthIT.gov Quick Stats, data through 2025.
  2. Hufstader Gabriel M, et al. Specialty pharmacy turnaround time impediments, facilitators, and good practices. J Manag Care Spec Pharm, 2022. PMC
  3. CCITT draft recommendation T.4 (RFC 804). IETF, 1981.
  4. Hsu E, et al. Deep learning-based NLP data pipeline for EHR-scanned document information extraction. JAMIA Open, June 2022.
  5. Jacobs TG, et al. Comparative evaluation of handwriting recognition by large language models in interpreting handwritten medication lists. Exploratory Research in Clinical and Social Pharmacy, May 27, 2026. PMC
  6. Pather N, et al. From Handwriting to Structured Data: Benchmarking AI Digitisation of Handwritten Forms. arXiv preprint 2604.16504, April 14, 2026.
  7. Rasmussen LV, et al. Development of an optical character recognition pipeline for handwritten form fields from an electronic health record. JAMIA, 2012.
  8. NPI overview. Centers for Medicare & Medicaid Services.
  9. National Drug Code Directory. U.S. Food and Drug Administration.
  10. Revising the National Drug Code Format and Drug Label Barcode Requirements, final rule. FDA, Federal Register, March 5, 2026.
  11. Guidance on HIPAA and Cloud Computing. HHS Office for Civil Rights.
  12. 45 CFR 164.312, Technical safeguards. Electronic Code of Federal Regulations.
  13. CMS Rule Phases Out Fax Machines, Snail Mail to Save Taxpayers $781.98 Million a Year. CMS press release, March 20, 2026.

Similar Blogs