AI for Insurance Submission Intake: Costs, Workflow, Limits
How AI submission intake works for insurers and MGAs: the real workflow, per-page costs from Google and Anthropic pricing pages, where extraction breaks, and the compliance rules that apply.

Every commercial insurance submission that lands in your inbox costs 15 to 20 minutes of manual sorting before an underwriter can even look at the risk. AI intake automation cuts that handling to minutes, and the per-page economics now strongly favour doing it: a mid-volume operation spending roughly $20,000 a month on manual intake can run the same volume through an extraction pipeline for $600 to $4,500 a month in machine costs, plus human review of the exceptions. This post breaks down what a submission package actually contains, the three ways to automate intake, real per-page pricing from Google and Anthropic, where extraction still fails, and what regulators expect you to document.
What a submission package actually contains
A commercial-lines submission rarely arrives as one clean form. A broker typically sends a single PDF, often 50 pages or more, stitched together from different sources: the ACORD application, loss runs from two or three prior carriers each with their own layout, supplemental questionnaires, financial statements, and a statement of values for property schedules. Some pages are native digital text. Others are scans, photos of scans, or scans with handwritten annotations in the margins.
This is why intake has resisted automation for so long. The ACORD forms themselves are standardized: ACORD maintains data standards precisely so carriers, brokers, and vendors can exchange data without re-keying it, and those standards are the foundation of straight-through processing. But everything around the ACORD form, the loss runs, SOVs, supplementals, and financials, arrives in whatever format the broker's office produced. The sorting problem, not the form problem, is what eats the 15 to 20 minutes per package that Upstage cites for manual submission preparation.
Three ways to automate intake
Template-based OCR and RPA
The first generation of intake automation used optical character recognition with fixed templates: the system knows where "Named Insured" sits on ACORD form 125, so it reads that box. This works when every document looks the same, and it breaks the moment a broker sends a different form version, a rotated scan, or a loss run from a carrier the template library does not cover. In practice, commercial intake involves too many document variants for templates to hold up. FurtherAI describes the failure mode plainly: rule-based systems break on non-standard layouts, which is most of what intake teams receive.
Template OCR still has one legitimate niche: very high-volume, single-form processes where you control the document source. If brokers fill in your own portal form, you do not need OCR at all.
Managed document AI processors
The second option is a cloud document AI service: Google Document AI, AWS Textract, or Azure AI Document Intelligence. These combine OCR with models that understand forms, tables, and document structure generically, and some offer custom extractors you train on your own documents.
AWS Textract, for example, ships five distinct APIs, including an Analyze Document API with form, table, query, and signature features and a specialized Analyze Lending API pre-trained on mortgage packages, per the Textract pricing page. Google's processors are similarly modular, with separate OCR, form parsing, classification, and splitting components.
These services are strong at OCR quality and structure detection. Their weakness is that extraction is only half the intake job. You still need to classify documents, reconcile conflicting values across forms, check the submission against your appetite rules, and route it. That orchestration layer is yours to build.
LLM-based extraction
The third approach sends pages to a large language model with a structured output schema: "return JSON with named_insured, FEIN, loss_history_years, total_insured_value." Modern models read the page image or its OCR text and fill the schema regardless of layout, because they understand what a loss run is rather than where a field sits. Anthropic positions its smallest model, Claude Haiku 5.5, explicitly for "high-volume, latency-sensitive tasks such as classification, extraction, and routing" on its API pricing page, which describes the intake workload exactly.
The honest trade-off: an LLM will occasionally produce a plausible-looking wrong answer rather than an error, which is why every serious implementation pairs extraction with confidence scoring and human review of low-confidence fields, a pattern we covered in detail for the claims side in our guide to LLM claims triage.
A working intake pipeline, step by step
Whether you use document AI processors, an LLM, or both, a production intake pipeline has the same shape:
- Ingest. Watch a shared mailbox or broker portal. Save the raw email and attachments untouched for audit.
- Split and classify. A 60-page stitched PDF becomes separate documents: ACORD 125, loss run carrier A, loss run carrier B, SOV, supplemental. Classification drives which extractor runs next.
- Extract. Pull structured fields from each document into JSON: named insured, addresses, limits requested, prior carrier, five years of losses with dates and amounts, building schedule rows.
- Validate. Cross-check within and across documents. Does the named insured on the ACORD form match the loss runs? Do the SOV building values sum to the requested TIV? Is the expiration date on the prior policy consistent with the loss run period? Flag every mismatch.
- Enrich and triage. Append external data where licensed (geocoding, catastrophe zones, NAICS classification), then score the submission against appetite: in appetite and complete, in appetite with missing items, out of appetite, or needs a human.
- Route. Write the clean records to the policy admin system or rating workbench, send out-of-appetite submissions to a decline workflow with a draft broker response, and queue missing-information requests automatically.
- Human review. Show a reviewer only the flagged fields with the source page highlighted. Their correction feeds back as a labeled example for evaluation.
FurtherAI's walkthrough of an automated intake flow follows essentially this sequence, and reports one MGA cutting per-submission processing from about 32 minutes to about one minute. Treat that as a vendor-reported best case; the underlying time saving from eliminating manual sort-and-key work is real and independently corroborated by the manual-baseline numbers above.
What it costs: a worked example
Take a regional commercial insurer or MGA handling 2,000 submissions a month at an average of 60 pages per package: 120,000 pages a month.
Manual baseline. At 20 minutes of sorting and keying per package (the top of the range Upstage reports), that is roughly 667 staff hours a month. At $30 an hour fully loaded, intake handling costs about $20,000 a month before an underwriter has evaluated a single risk.
Managed document AI route. Using Google's published rates from the Document AI pricing page: a custom splitter and classifier at $5 per 1,000 pages plus Form Parser extraction at $30 per 1,000 pages gives 120 x $5 + 120 x $30 = $4,200 a month. Add integration engineering and the validation layer.
LLM route. Using Claude's published rates: assume roughly 1,200 input tokens per page (page image plus instructions) and 300 output tokens per page of structured JSON. These token counts are planning estimates; measure them on your own documents in week one. At those assumptions, 120,000 pages is 144 million input tokens and 36 million output tokens. On Claude Sonnet 5.5 at $2 per million input and $10 per million output, that is $288 + $360 = $648 a month. On Claude Haiku 5.5 at $0.10 and $0.50, the same volume is about $32 a month, which is why teams typically route clean pages through the cheap model and reserve the stronger one for flagged pages.
The number that stays. Human exception review does not disappear. If reviewers spend five minutes per package checking flagged fields, that is 167 hours, about $5,000 a month at the same rate. Total automated intake: roughly $5,700 to $9,200 a month all-in, against $20,000 manual, and the underwriters get a structured, validated submission instead of a PDF.
Two more cost notes. Per-page prices on the specialized processors reward scale: Google's OCR tier drops from $1.50 to $0.60 per 1,000 pages above five million pages a month, and Form Parser drops from $30 to $20 per 1,000 pages above a million. And LLM batch processing halves token costs for work that does not need an immediate answer, which fits overnight re-extraction and backfile conversion.
Where extraction breaks
Accuracy limits deserve more honesty than vendor pages give them. The failure modes discussed consistently across primary sources:
- Handwriting. Annotations on scanned ACORD forms are the classic failure case. OCR engines handle neat handwriting; hurried broker notes are another matter. Flag, do not guess.
- Ambiguous checkboxes. A checked box is easy; a faint scan where two boxes look half-marked is not. Low confidence here should always route to a human, because the answer often changes coverage.
- Conflicting values across documents. The ACORD form says 2018 construction; the SOV says 1985. Neither OCR nor an LLM knows which is right. The pipeline's job is to detect the conflict and surface it, not to pick one silently.
- Novel document types. A supplemental questionnaire your extractor has never seen will get plausible but unreliable output until you add it to the evaluation set and tune the schema.
- Hostile content inside the package. A submission PDF is untrusted input, and text in a document can be written to look like instructions to the model. Treat everything extracted from a broker package as data, never as commands, and validate before anything downstream acts on it. Our prompt injection defenses guide covers the layered controls.
The NAIC itself makes the underlying point on its artificial intelligence topic page: LLMs can generate information that sounds accurate but is incorrect, so AI output used for important decisions should be reviewed carefully. Build your evaluation set from real submissions on day one: a few hundred labeled packages, re-run on every model or prompt change, measuring field-level accuracy. Our guide to evaluating AI systems before deployment in a bank or insurer covers the adjacent discipline of keeping personal data out of the pipeline while you do it.
What regulators expect
Intake automation is one of the lower-risk uses of AI in insurance because it prepares decisions rather than making them, but it is not unregulated territory.
In the United States, the NAIC's Model Bulletin on the Use of AI Systems by Insurers, adopted in December 2023, expects insurers to maintain a written AIS Program covering governance, risk management and internal controls, third-party vendor oversight, and documentation for AI systems that support regulated insurance practices. The NAIC's adoption map shows 29 states and jurisdictions had issued the bulletin as of November 2026, with four more (California, Colorado, New York and Texas) applying their own insurance-specific AI rules or guidance. Adoption matters for intake too: if extracted data feeds rating or underwriting decisions, regulators can ask how that data was produced and validated. The NAIC has also run an AI Systems Evaluation Tool pilot with 12 states from March through September 2026, with a revised version of the tool slated for adoption consideration at the 2026 Fall National Meeting, so expect examination requests to get more specific, not less.
In Kenya, submission packages contain personal data: names, ID numbers, financial statements, claims histories. The Data Protection Act applies to how you process and store that data, including which cloud regions your extraction pipeline runs in. Our guide to the Kenya Data Protection Act and AI in banking and insurance walks through the obligations. Wherever you operate, keep raw submissions in your own storage, log every extraction, and treat vendor selection as a compliance decision: our AI vendor due diligence guide for banks and insurers has the question set.
When intake automation is the wrong choice
Do not build this if any of these describe you:
- Volume is under roughly 200 submissions a month. At that scale, the manual cost is a few hundred dollars a month and a well-organized shared mailbox with clear naming conventions beats a pipeline project.
- You can fix the problem at the source. If most of your brokers will fill in a structured web form or portal, do that instead. Extracting data you could have collected structured is permanent complexity for a solvable problem.
- Your book is mostly non-standard specialty risks. Intake automation pays back on repeatable document types. If every submission is a bespoke manuscript package, the evaluation and tuning effort per document type may never pay off.
- You cannot staff the exception queue. Automation that routes 15 percent of fields to human review and nobody reviews them is worse than manual intake, because errors enter your systems with a veneer of machine authority.
Common mistakes
Automating before measuring. Baseline your current cost first: minutes per package, error rate on keyed data, time from receipt to underwriter. Without it you cannot prove the project worked or catch a pipeline that made things worse.
Skipping the splitter. Teams demo extraction on a clean ACORD 125 and declare victory, then discover in production that 80 percent of the effort is splitting stitched PDFs and classifying the pieces. Test on real broker packages, not sample forms.
Treating extraction as the product. The extracted JSON is worthless until it is validated, triaged, and written into the systems underwriters actually use. Budget at least half the project for validation rules, routing, and integration.
No feedback loop. Reviewer corrections are labeled training data. If your pipeline does not capture them, accuracy plateaus at whatever the vendor shipped.
Letting PII flow to places it should not. Submissions contain personal and financial data. Redact or minimize what leaves your environment, as we detail in the PII redaction guide linked above, and confirm your extraction vendor's data retention terms in writing.
Frequently asked questions
How accurate is AI extraction on ACORD forms? On clean, digital-native ACORD forms, a tuned pipeline with validation rules should leave only a small share of fields needing human review. Accuracy drops on scans, handwriting, and non-standard supplements, which is why confidence scoring and human review of flagged fields are part of the design, not an optional extra. Measure field-level accuracy on your own evaluation set rather than trusting a vendor's headline number.
What does AI submission intake cost per month? At 120,000 pages a month, machine costs run from about $32 a month on a small extraction-focused LLM to about $4,200 a month on managed document AI processors at published rates, plus integration engineering and human exception review. The manual baseline for the same volume is around $20,000 a month.
Do we still need intake staff after automating? Yes, but fewer and differently deployed. Reviewers handle flagged fields and exceptions, typically minutes per package instead of twenty. The NAIC's own guidance expects human oversight to remain part of insurance decision-making.
Does this work for loss runs and SOVs, not just ACORD forms? That is precisely where LLM-based extraction earns its keep: loss runs and SOVs arrive in dozens of carrier-specific layouts that defeat templates but are readable by a model that understands the document type. Validate totals and cross-form consistency automatically and route conflicts to a human.
Is submission intake automation compliant with insurance regulation? It can be, and regulators have said how. In the US, the NAIC model bulletin, now issued in 29 states and jurisdictions, expects a written AI governance program, risk controls, vendor oversight, and documentation. In Kenya, the Data Protection Act governs the personal data in submissions. Build logging and audit trails into the pipeline from day one.
Where to go from here
The economics are settled: at any meaningful submission volume, machine extraction plus human exception review is cheaper and faster than manual sort-and-key work, and regulators have told you exactly what documentation they expect around it. The build is not exotic, but it does demand real insurance data handling: evaluation sets from actual submissions, validation rules that match your underwriting appetite, and an audit trail a regulator can read. If you want that pipeline designed and run for you, our enterprise AI work for banks and insurers covers private deployment, evaluation, and governance documentation, and our business automation team handles the mailbox-to-system routing around it.
Related reading
About AI Agents Plus Editorial
AI automation expert and thought leader in business transformation through artificial intelligence.



