Cobalt Ridge Specialty and the submission queue
Cobalt Ridge Specialty (made up) writes commercial property and casualty through brokers. Submissions arrive as emails with a mix of ACORD forms, statements of values, loss runs and free-text notes. Underwriters spend much of their day re-keying data and hunting for the facts that decide whether a risk is in appetite. This is how the Insight products fit that process. The company and its numbers are fictional.
The problem in one sentence
The underwriting decision needs a few dozen facts, and those facts are spread across five documents in four formats that arrive by email.
1. From email to validated data: Insight Extract
Insight Extract runs agentic pipelines rather than single-step OCR: classify each document, extract it to a schema, run business rules, cross-check against systems of record, and escalate when something doesn't fit. Extraction is layout-aware across structured forms, semi-structured loss runs and free-text emails, with a confidence score for every field. Its out-of-the-box pipelines include submission intake, claims first notice of loss, renewal data refresh and bordereaux processing.
Whether a submission goes straight through or to a person is configured per line of business, per document type and per confidence threshold. Reviewers correct fields and classifications with a reason, and accumulated corrections are turned into suggested changes to prompts, rules and class descriptions, for a reviewer to approve.
2. Assessment: Insight UW
Insight UW reads each submission, checks it against the carrier's guidelines, appetite and existing book, and gives the underwriter a recommendation with the signals behind it, before they open the file. How much it does on its own is a setting, line by line; humans hold the authority, and the underwriter can override every recommendation. It covers the underwriting workflow from submission and triage through research, rating, quote and bind, with integrations to the policy administration system and enrichment sources.
3. The evidence trail
Each step is recorded: which agent version extracted which field from which page, what UW recommended and why, and what the underwriter decided. For carriers that also run Insight Lense, agent runs and authority decisions sit in the same kind of thread as everywhere else in the programme.
What changes for the team
| Role | Before | With Extract + UW |
|---|---|---|
| Underwriting assistant | Re-keys every form | Reviews only low-confidence fields |
| Underwriter | Hunts for appetite signals | Starts from a recommendation with cited signals |
| Underwriting manager | Samples files for consistency | Sees overrides and their reasons |
| Compliance | Reconstructs decisions after the fact | Reads them from the audit trail |
How Cobalt Ridge starts
- One line of business (commercial property) and one broker channel.
- Extract on submission intake, with a high confidence threshold so most fields get reviewed at first.
- UW recommendations shown to underwriters, with no automatic actions.
- Lower the threshold and widen the scope as review data shows where extraction is reliable.
Regulation and model risk. Use of AI in underwriting is subject to insurance regulation and model-risk expectations that vary by jurisdiction. This example is general information; confirm requirements with your compliance and legal teams.
Questions to ask in the first working session
- Which lines and document types? Extraction agents are built per line of business; start with the one with the most volume.
- Where is the truth? Which system of record each extracted field should be cross-checked against.
- What are the guidelines? Appetite and referral rules written down clearly enough to check a submission against.
- Who reviews exceptions? And how quickly, so low-confidence submissions don't become a new queue.
- What must the audit show? For underwriting, model-risk and compliance sign-off.
The UW site suggests seeing it run against a real broker email, a real submission and your own lines and policy system; that's also the fastest way to find which of these questions need more work.
Related: running AI inside your environment.