How to Extract Data from ACORD Forms with AI

Why ACORD forms are hard to read at scale

The ACORD form is the closest thing commercial insurance has to a common language, yet the data inside it still arrives as documents a person has to read and rekey. A single submission can travel as an emailed PDF proposal form, a scanned and signed certificate, a photographed page from a broker, and a spreadsheet of schedules, all describing the same risk. Before anyone can price it, someone has to find the named insured, the FEIN, the requested limits, the effective dates, and the loss history across those files, key them into the core or the rating engine, and reconcile the contradictions between them. Gartner has estimated that corporate teams lose 20 to 30 percent of their time organizing unstructured data instead of doing analytical work, and insurance intake is exactly where that loss lands.

AI extraction attacks this at the intake stage of the quotation and underwriting journey. Instead of asking which character sits in a pixel region, a reading model asks what a value means, so it can pull the same field from a layout it has never seen in that exact form. That shift, from transcription to interpretation, is what makes ACORD extraction viable across the hundreds of broker and carrier variants a real book of business produces.

Why traditional OCR and RPA under-deliver on ACORD

Most insurers have already tried to automate this, usually with classic OCR and RPA, and both fell short for the same structural reason. Traditional OCR turns pixels into characters but does not understand the document, so it is template bound: it works only when a field sits in the same place on the same form every time. ACORD reality is the opposite, because every broker, every line of business, and every insured sends a slightly different arrangement, a different attachment order, or a marked-up prior form. RPA scripts the clicks a human would make, so it breaks the moment a portal field, a form version, or an upstream layout changes. Both tools treated reading as mechanical, when the real task is judgment about what each value is.

An AI reading layer replaces that brittleness with three properties. First, layout independence, because models trained on insurance documents generalize across forms, emails, certificates, and spreadsheets, so a new broker template does not require a new rule. Second, a confidence score on every field, so high-confidence values flow straight through while low-confidence ones route to a person for a quick check. Third, validation, where each value is checked for completeness against the requirements of that line, for format on fields like FEIN, ZIP, dates, and currency, and for consistency across the documents in the same submission. If you are weighing this against your current stack, the deeper comparison lives in why AI reading beats RPA and OCR for insurance.

How the extraction pipeline actually works

A production pipeline treats the whole submission as one object, not as isolated files. The layer ingests everything that arrives for a risk, the email body and each attachment together, and reads it as a single unit. It reads scanned and photographed ACORD forms reliably, and it attempts handwritten and free-text fields as well, which is where realism matters: handwriting recognition in 2026 is materially less accurate than print, so the pipeline does not pretend those fields are solved. Instead it scores each extracted value and sends anything uncertain, including most handwriting and marginal notes, to a reviewer. That inverts the old OCR model where a person checked everything; here a person checks only the genuinely ambiguous fields.

The output is not raw text. It is a structured, validated submission object the rest of the workflow can consume, which is what makes downstream stages fast: broker enrichment, risk and fraud scoring, pricing, and the final decision all depend on clean intake. Clean, structured input is also the single biggest lever on your straight-through processing rate, because every field that arrives validated is a field no underwriter has to touch. The broader challenge of turning messy documents into decisions is covered in handling unstructured data in insurance with AI, and the intake-specific mechanics in insurance submission intake automation.

Deploying it without replacing your core

The architectural point that matters most is that this does not require a core migration. WIR is an external AI layer that sits on top of the systems the insurer already runs, and it never replaces the policy admin core, the underwriting workbench, or the rating engine. It ingests ACORD submissions from the channels above, produces structured validated data, and writes that data back into the system of record through APIs or files. The core stays intact. This matters because IT constraints are one of the most cited blockers to insurance innovation, and an overlay model lets an insurer modernize intake without betting the company on a multi-year replacement program.

The work runs as a defined project rather than an open-ended IT effort. A one-time setup phase covers scope, integration with the existing core, calibration to the underwriting manual and risk appetite, testing, and go-live, with a fixed scope and KPIs agreed before the work starts. After go-live the layer moves to continuous operation in production, staying fully external with no load on the insurer's IT. The extracted data then feeds the desk directly, which is where an AI underwriting workbench turns clean intake into faster, more consistent decisions.

Governance, explainability, and Brazil context

Reading ACORD forms touches personal and corporate data, so governance is part of the design rather than an afterthought. For Brazilian operations, submission data often contains personal data under the LGPD (Lei Geral de Proteção de Dados, Lei 13.709/2018), so the layer must process it on a lawful basis, with data minimization, encryption at every step, and a human in the loop where an automated decision affects a person, with the ANPD as supervisory authority. SUSEP supervises the market, so automated reading and the underwriting that follows must stay consistent with the registered product terms and the underwriting manual.

Explainability is where the confidence score stops being only an accuracy feature and becomes a governance mechanism. Every extraction, every confidence level, every validation outcome, and every human override is logged, so the insurer can prove which values were machine-extracted with high precision and which a person checked. Decisions are explainable and return a full audit trail, never presented as infallible. That provenance is exactly what makes automated ACORD extraction defensible to internal audit and to the regulator.

How WIR reads ACORD submissions

WIR is the AI layer for insurance, on top of the systems the insurer already runs, never in their place. It automates the quotation and underwriting journey according to the insurer's own risk-acceptance policy, with Machine Learning calibrated to the risk appetite and the underwriting manual. ACORD reading is the intake stage of that flow: multichannel ingestion and validation, document reading, broker enrichment and scoring, a multi-factor risk and fraud engine, pricing, and a final decision written back to the core with its explanation and audit trail. WIR Innovation was founded in 2025 from accumulated operational experience, built with Mahway, a Venture Builder in California, and Avante, a Venture Studio in Brazil, and its first proof of concept is in execution with a global insurer in the Transport line. To see how ACORD extraction would map to your own forms and core, talk to WIR.