Form Filling Sessions

Upload your documents, let the AI fill the form, review and correct the 2–5 flagged fields, download your PDF — in under 2 minutes

Overview

A form filling session is a workspace that links a PDF form with one or more source documents and runs the AI fill pipeline. You select a form (from your library or from the 44,783+ pre-configured catalog), attach the documents that contain the data — a resume, a prior application, a bank statement, an insurance card — and the AI extracts the relevant information and places it in the correct fields.

The pipeline runs in six stages: initialize the session, retrieve the form's field inventory, fill fields via autofill_db_fields(), generate the filled PDF, calculate usage costs, save and emit events. The field-filling stage dispatches independent field groups concurrently — up to 40 parallel fill tasks — so a 200-field mortgage application doesn't take 10× longer than a 20-field cover sheet. Fields with dependencies (spousal fields conditional on marital status, sub-questions that activate on "Yes") are evaluated in the correct sequence using conditional rules appended to each group's prompt.

Once the AI completes, you review the result in the visual editor. Fields the AI was uncertain about are highlighted — typically 2–5 out of the total. You correct those, optionally capture a signature, and download the final PDF.

Sessions are saved automatically. You can return later, share a session with a teammate for collaborative review, or reference the audit trail (which source document populated which field) for compliance purposes.

Key Capabilities

• Any PDF form: fillable AcroForms and flat/scanned PDFs (converted to fillable on first use); forms up to 200+ pages with 500+ fields supported
• Multiple source types: PDF, Word, Excel, images (with OCR), pasted text, email attachments, webhook payloads, profile data — all usable in the same session
• Parallel group filling: independent field groups fill concurrently; a 100-field form fills in roughly the same wall-clock time as a 10-field form
• Dependency-aware logic: "Fill [Spouse SSN] only if [Marital Status] is 'Married'" resolved before filling — fields that should stay blank remain blank
• Confidence-flagged review: each filled field receives a confidence score; low-confidence fields surface in the visual editor so reviewers know exactly where to focus
• Field explanations: for any filled field, the AI can explain why it chose that value — which source document, which passage, and the reasoning — useful for audit and for training new staff
• Signature capture: draw, type, or upload a signature and embed it in signature fields without leaving the session
• Stateless mode: source documents deleted immediately after filling for maximum privacy compliance
• Export options: filled PDF, flattened (non-editable) PDF, field data as JSON, or webhook delivery of the result

How It Works

1. Select a Form

Upload your PDF form or pick one from the catalog. The form is processed once to extract its complete field inventory: every field's name, label, type (Text, Date, CheckBox, RadioButton, ComboBox), page number, maximum character length, and available dropdown choices. On repeated fills of the same form, this inventory is reused — no reprocessing.

2. Attach Source Documents

Add the documents that contain the data you need:

• Upload PDF, Word, or Excel files
• Paste text directly
• Pull in a saved profile (profile files are async-copied into the session source list before the fill job starts)
• Receive data via email or webhook

Source document text is encrypted with workspace-scoped keys before storage. Processing status (OCR, vector indexing) is trackable while sources are prepared.

3. AI Fill

Source content is mapped to specific form pages before the fill job starts — the AI filling page 3 only receives text relevant to page 3, preventing cross-contamination on long multi-section documents.

Independent field groups are dispatched concurrently. Each group's prompt includes:

• The group's field list (labels, types, max lengths, choices)
• The page-scoped source text
• Any conditional dependency rules for that group

Table and repeating fields (employment history rows, medication lists, address history) go through a dedicated extraction pass that structures source data into rows before populating the table cells.

Field values exceeding maximum character length are reformulated using font-metric estimates — no hard truncation that leaves values cut off mid-sentence.

4. Review in the Visual Editor

The filled form opens in the visual editor with low-confidence fields highlighted. Click any field to see its value. For flagged fields, use the "Explain" option to see which source document the value came from and why the AI selected it. Correct any field by typing a new value — corrections are saved to the session and logged in the audit trail.

5. Generate and Download

• Filled PDF: field values embedded, form remains editable
• Flattened PDF: values baked into the page image, non-editable — standard for form submissions
• JSON export: all field names and values for integration with downstream systems
• Webhook: auto-push filled PDF to a document management system or CRM when the session is marked complete

Use Cases

Healthcare credentialing: A teleradiology group submits the same physician to 12 hospital systems. Each hospital has its own credentialing packet with different field names covering the same information (license, DEA number, malpractice history, education). One session per hospital, same source documents, AI fills each packet in under 60 seconds. 12 packets in 15 minutes instead of 2 hours of manual entry.

Mortgage processing: A 1003 loan application has 200+ fields across borrower information, employment history, assets, liabilities, and property details. The AI pulls from W-2s, bank statements, and pay stubs simultaneously, fills independent sections in parallel, and completes the fill in about 45 seconds. The loan officer reviews and corrects 3–4 fields requiring judgment.

Estate planning: A Massachusetts guardianship form has conditional sections that only apply based on whether the petitioner has prior court history, whether the ward has living relatives who need notification, and whether an emergency appointment is being sought. Dependency handling ensures the applicable sections fill and the inapplicable ones stay blank — preventing the common error of filling in fields that the court will reject.

Insurance claims: An adjuster fills a first notice of loss from a policyholder statement and the declaration page. Claimant contact information from the statement and policy limits from the declaration combine in a single fill pass without manual toggling between documents.

Benefits

• Under 2 minutes from source upload to downloaded PDF for most forms, vs. 15–30 minutes of manual copy-paste
• Parallel fill architecture means field count doesn't proportionally increase fill time — 100 fields in independent sections fills as fast as 10 fields
• Conditional logic prevents a common class of errors: sub-fields that should be blank (because the parent answer was "No") no longer get accidentally filled
• Confidence flags focus the reviewer's attention — most users correct 2–5 fields rather than checking every field on a complex form
• Complete audit trail per session: which source documents, which fields were AI-filled, which were manually corrected, who made corrections, when the PDF was downloaded — sufficient for HIPAA, SOC 2, and legal compliance

Security & Privacy

Session data is encrypted using workspace-scoped keys managed via Azure Key Vault. Source text is encrypted with scope restriction so content from one session cannot be decrypted in another's context. Stateless mode processes source documents and deletes them immediately after PDF generation — no persistent copy of PHI or financial data.

Role-based access controls restrict session visibility to users with workspace permissions. The audit trail records every source document, every field fill event, every user correction, and every download. Sessions and their source documents are automatically purged according to workspace retention policies.

Common Questions