ACORD Form Data for Insurance Submission Intake Models

Why there is no public ACORD training set

Public search for ACORD extraction data returns extraction products, not datasets, because the vendors behind them train on private collections of real forms [1][3][4]. Vendors list ACORD 25 certificates and ACORD 125 commercial applications among the most-used forms, and advertise models for ACORD 25, 125, 126, 130 and 140 [1][3]. Their accuracy figures rest on form sets you cannot inspect, so they are not a benchmark you can reproduce.

General form benchmarks do not fill the gap. FUNSD contains 199 noisy scanned forms from domains such as marketing and scientific reports, annotated for OCR, layout and entity linking [5]. It is useful for pretraining a key-value linker, but it has no insurance vocabulary, no multi-page applications and no checkbox-heavy coverage grids. For the broader landscape of public and licensed sets, see the document AI datasets hub.

What a usable submission packet contains

A commercial submission is a packet, not a form, and models trained on isolated ACORD pages fail on real broker inboxes. A typical packet mixes the ACORD 125 with line sections (126 general liability, 140 property, 127 business auto, 130 workers compensation), carrier supplemental applications, a schedule of values spreadsheet, three to five years of loss runs, driver and vehicle lists, and the broker's cover email with its attachments.

That mix drives three distinct tasks, and each needs its own labels:

• Page-stream segmentation and classification: where each document starts and what it is, covered in depth on page-stream segmentation data.
• Field extraction: named insured, mailing and premises addresses, FEIN, SIC or NAICS code, entity type, years in business, requested effective date, limits and deductibles, prior carrier and premium history, and loss summaries.
• Triage: appetite fit, completeness and missing items, which links to underwriting decision rationale data for agent workflows.

Loss runs are their own extraction problem with carrier-specific layouts; buy them alongside the ACORD forms using the guidance on loss run extraction data.

Edition dates and layout drift

Record the form number and edition date for every document, because extraction vendors support several ACORD revision years and templates trained on one edition break on another [2]. The edition stamp printed in the form footer is the cheapest feature for routing a page to the correct schema. Ask suppliers for the edition distribution before you agree to anything, since an archive dominated by one revision will overstate accuracy.

Rendering format matters as much as edition. Fillable PDFs exported from agency management systems carry clean text layers, while scanned, faxed or photographed forms carry skew, stamps and handwriting. Checkbox grids for coverages and entity type deserve dedicated labels; see checkbox and selection mark data.

Label schema buyers should specify

Ask for field-level labels that tie each value to its page, bounding box, form number and edition, with a separate flag for values a human corrected. Labels derived from the policy admin or rating system ("what the underwriter actually keyed") are stronger than fresh annotation, because they reflect production truth; the pattern is covered on documents paired with system-of-record entries.

Illustrative example: invented to show structure; it does not describe an available dataset.

Use a held-out set split by broker and by submission date, not by page, so the same insured never appears in both train and test. Field-level scoring methods are covered on document extraction evaluation sets.

Rights, confidentiality and personal data in submissions

Treat the form template and the filled data as separate rights questions: the layouts belong to the forms publisher (ACORD), while use of the answers is governed by the agreements among the insured, the broker and the carrier. Confirm both before you license, and ask whether the supplier's agreements with brokers allow use of submissions beyond underwriting.

Most submission content is commercially confidential business data such as revenue, payroll, loss history and pricing. Personal data still appears: sole proprietors who enter a Social Security number in the FEIN field, driver lists with license numbers and dates of birth, and owner contact details. Gramm-Leach-Bliley privacy rules cover nonpublic personal information about consumers held by financial institutions [6]; insurers comply under state insurance regulation, and most commercial submission data concerns businesses rather than consumers, so confirm which records are in scope. California defines deidentified data with specific conditions the holder must meet [7].

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Buyer request checklist for ACORD submission data

A precise request shortens assessment because suppliers can check their archives against it.

Illustrative example: invented to show structure; it does not describe an available dataset.

• Lines of business and forms: for example ACORD 125 with 126 and 140, plus supplementals.
• Edition coverage: which revision years, and the share of each.
• Channel and format: broker email with attachments, portal upload, fax, or scan; native PDF vs image.
• Packet completeness: loss runs, schedules of values, driver lists included or not.
• Labels: field-level values with bounding boxes, source of truth, correction flags.
• Outcomes: quoted, declined or bound, if triage models are in scope.
• De-identification: which fields are masked or replaced, and how a sample is checked.
• Allowed uses: training, fine-tuning, evaluation, and term, written into the license.

How SourceX sources ACORD submission packets

SourceX sources operational datasets from US companies on request, including document and finance workflows, and manages licensing and ongoing purchases. Nothing is held in stock: you describe the forms and labels you need, SourceX looks for US businesses that hold them, and a request does not guarantee a match. You can start that description on the SourceX buyers page.

The process runs Find, Assess (the data and licensing permissions), Agree (pricing and allowed uses in a license), Transact and Manage, and nothing is contracted until a supplier agrees. Every dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Names, emails, phone numbers and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect.

Related SourceX pages cover underwriting files for AI training, insurance claims datasets and buyers in insurance brokerages.

Request ACORD form extraction data

SourceX sources filled submission documents from US businesses on request, under a license that defines records, uses, term and delivery, with every release approved by the supplying company. Describe the forms, editions and labels you need on the SourceX buyers page.