Submission ingestion converts a broker's package into structured fields an underwriter can act on. It covers classification, OCR, field and table extraction, normalization, schema mapping, and confidence-based review, and it ends once a structured record exists.

A typical broker package includes ACORD forms, statements of values (SOVs), loss runs, and a covering email. Submission data is also the freshest read a carrier has on a risk, a policy administration system reflects what was true at the last renewal, a broker submission reflects what's true now: current operations, current locations, current exposures. Ingestion is the point where that currency either gets captured or gets lost.

What submission ingestion covers and where it stops

Submission ingestion starts when the carrier receives unstructured broker documents and ends when a structured record exists. Everything after that point belongs to a different stage. Keeping this boundary clear is the only way to interpret what ingestion has actually delivered.

ACORD standardizes the forms and messaging brokers use. Ingestion determines how the information inside those documents becomes data the carrier's systems can read. Any detail left only in the original PDF stays invisible to automated checks and routing.

The stages that follow ingestion

Once a structured record exists, it moves through separate downstream steps with different owners. Conflating them makes it impossible to tell which stage is actually delivering value.

• Clearance checks for duplicate submissions and broker of record conflicts.

• Triage scores the record against appetite and prioritizes it.

• Routing places the record in the right underwriter's queue.

• Pricing runs the actuarial model to produce a technical price.

• The binding decision stays with the underwriter, who quotes, binds, or declines the risk.

How submission extraction works, stage by stage

Extraction runs through five stages, and an error early on can pass silently through the later ones. A misclassified document gets the wrong extraction logic, and a downstream confidence score won't necessarily reveal the mistake.

1. Intake and classification. The system converts scanned PDFs, native PDFs, Excel workbooks, and images into a processable form, then decides whether each file is an ACORD application, a loss run, an SOV, or a cover email.

2. OCR and layout analysis. Optical character recognition converts scanned images into text, and quality drops with poor print or degraded copies. Native digital PDFs skip OCR, removing one source of error. Multi-column ACORD forms and split-column SOVs are common failure points.

3. Key-value and table extraction. Forms yield labeled fields, while SOVs and loss runs arrive as tables. Loss run formats vary by issuing carrier, so extraction often needs separate handling for each carrier family.

4. Normalization and schema mapping. One package might name the insured "ABC Manufacturing Co.," "ABC Mfg.," and "A.B.C. Manufacturing Company." Normalization resolves these to a single insured; schema mapping then translates normalized values into the field names the carrier's systems expect.

5. Confidence scoring and human review. The system assigns a confidence score to each extracted field and routes anything below a configurable threshold to a reviewer, underwriting assistants, a clearance team, or another operations group.

Why submission data is the most current read on a risk

A submission reflects what's true about a risk right now, which makes it different from almost every other record a carrier holds. Nobody re-files a policy administration record when an insured's operations change between renewals, but the next submission shows it.

Russia's 2022 invasion of Ukraine is a stark illustration. Over 600 aircraft leased from foreign lessors were operating inside Russia when the invasion began; by late April that year, only 32 had left, leaving more than 400 aircraft, collectively worth over $10 billion, stranded and reregistered onto Russia's domestic registry. Lessors and insurers had been carrying those assets on records that assumed normal operating location. The mismatch between recorded location and actual location became the subject of a multi-billion-dollar legal dispute that ran for years, including AerCap's $3.5 billion claim against AIG and Lloyd's insurers and a $4.5 billion London Commercial Court ruling in 2025. A year after the invasion, S&P Global Market Intelligence reported insurers were still unable to tally the war's true cost, partly because exposures "far removed from the war itself" kept surfacing as better data came in.

How structured submission data feeds triage, pricing, and the underwriting decision

Structured data creates value only when the next stage can use it without rekeying. Each downstream stage has a different owner, and the handoffs between them determine how much of that value survives into the quote.

Triage scores the structured record against appetite and routes it. Actuaries and pricing teams build the models and set the framework underwriters work within. Individual underwriters use the priced record to quote, bind, or decline, and they can override any triage outcome. The same division of responsibility holds across carriers, reinsurers, Lloyd's syndicates, and MGAs.

When extracted fields don't match what the pricing model expects, someone remaps or rekeys them, and the time saved at intake disappears. Cleaner movement from intake to pricing shortens quote turnaround and improves the quote-to-bind ratio, which supports GWP growth. Speed only pays off when the price is right, as covered in speed without intelligence.

Why extraction accuracy alone isn't the point

Correct extraction is necessary, but a correctly extracted field is still just a value sitting in a schema until something reads it against appetite, pricing, and the rest of the book. That reading is where the real value of ingestion shows up, not in how the field was captured. McKinsey's own outline of an underwriting operating system treats ingestion the same way, as the first link in a connected flow, not a standalone checkpoint to be scored in isolation.

How a shared decision layer changes what ingestion is for

Most ingestion tools stop at the handoff: extract the data, hand it to the next system, and let that system figure out what matters. hx connects ingestion to a shared decision layer instead, the same context that pricing, triage, and portfolio intelligence draw on while they work.

That shared layer is what AI agents inside hx actually operate on. Every agent working a submission, the one extracting data, the one scoring triage, the one preparing a price, reads and writes against the same context rather than passing a file between disconnected systems. A field that would otherwise just be "captured correctly" gets read instead against what the carrier's pricing model expects, what the portfolio already holds, and what the carrier's appetite rules say about this specific combination of factors.

That's what the shared layer enables: signals get picked up not because an extraction model scored them with high confidence, but because the broader context makes clear they matter. A location change might be extracted perfectly by any ingestion tool, but only a system reading it against the existing book knows whether it shifts the risk's price, its profitability, or its fit with other business the carrier already holds. The same underlying question, is this risk winnable, gets a second, separate answer when it's read this way: is it desirable, worth winning at the price on offer, and does it strengthen or dilute the portfolio it joins. Extraction accuracy can't answer either question on its own. A shared decision layer can, because the data isn't just captured correctly, it's read in context.

How hx connects submission ingestion to a shared decision layer

hx is AI for insurance underwriting: an agentic underwriting workbench that executes the work around the underwriting decision inside the carrier's own pricing logic and appetite rules. A traditional workbench manages and routes work, while hx prepares, advances, and connects it. The underwriter keeps the judgment, a split explored in where agents execute.

• Actuaries and pricing teams define the extraction schema from the pricing model itself. Extracted fields arrive in the shape pricing and rating expects, with no rekeying between intake and price.

• Each line of business runs its own submission triage model and schema, keeping intake consistent with that line's appetite and data requirements. An in-appetite risk moves to full pricing in one click.

• Decision Trace records every action, model call, and data touch by default, so governance teams can investigate exceptions without shifting accountability away from the underwriter.

• Carriers can start with email forwarding or manual upload, then phase production use under their own IT, security, and governance controls.

Book a demo to see how hyperexponential connects submission ingestion to the same decision context as pricing, triage, and portfolio intelligence, so what gets captured is also what gets acted on.