Building a Claims Assistant Custom GPT

A custom GPT is a configured version of ChatGPT with a specific persona, instructions, knowledge files, and behaviors. For insurance professionals, this means you can build a "Claims Assistant" that knows your line of business, your jurisdiction's rules, your tone of voice, and your file documentation standards. Then you reach for it the way you reach for a calculator.

What You'll Learn

• How custom GPTs work and what they cost
• The components of a great Claims Assistant GPT
• Step-by-step build of a personal injury claims assistant
• How to keep your custom GPT compliant and accurate

What Is a Custom GPT?

Custom GPTs are a feature of ChatGPT Plus, Team, and Enterprise plans (around $20/month for Plus as of 2026). You can build one in 10 minutes through the GPT Builder interface, with no coding required. A custom GPT consists of:

• A name and description (so you and your team can find it)
• Instructions (the system prompt — the rules the GPT always follows)
• Knowledge files (PDFs and docs the GPT can reference)
• Conversation starters (suggested first prompts)
• Capabilities (web browsing, code interpreter, file generation)
• Privacy settings (private, link-only, or public)

For insurance work, you almost always want private GPTs — your team can use them, but they are not in the public GPT store.

Claude has a similar feature called Projects, and Gemini has Gems. The principles are the same.

Why Custom GPTs Beat Repeated Prompts

If you find yourself pasting the same instructions ("act as a senior bodily injury claims examiner in California, follow our reservation of rights template, do not invent specials") into every chat, a custom GPT lets you set those instructions once and forget them.

A claims adjuster who builds three or four custom GPTs (one for FNOL triage, one for evaluation memos, one for customer communications) typically saves 30-45 minutes a day after the first week.

Components of a Great Claims Assistant GPT

Across the carriers and brokers building these now, the strongest custom GPTs share a few traits:

1. Tight scope. "Personal injury claims examiner for New York auto" beats "Insurance assistant."
2. Clear safety rails. Explicit rules about not inventing facts, not making coverage decisions, and not exposing PII.
3. Templated outputs. Standardized memo structures, letter formats, and bullet points.
4. Knowledge grounding. Uploaded reference documents (carrier claim manual excerpts, ROR template, evaluation memo template).
5. Persona. "You are a senior, conservative, plain-spoken claims examiner who would rather ask one more question than miss a detail."

Step-by-Step Build: PI Claims Assistant for NY Auto

Open ChatGPT Plus, click "Explore GPTs" then "Create."

Step 1: Name and Description

Name: NY Auto PI Claims Assistant

Description: Helps NY-licensed adjusters draft evaluation memos, file notes, and reservation of rights letters for personal injury auto claims.

Step 2: Instructions

Paste this into the Instructions box, customizing for your carrier and rules:

You are a senior personal injury claims examiner with 15
years of experience handling New York auto claims. You
work for a regional carrier. You always follow these rules:

1. You only provide guidance on auto bodily injury claims
   in New York State. If asked about another state or line,
   politely redirect.

2. You never make coverage determinations. You provide
   analysis and structured drafts; the licensed adjuster
   decides.

3. You never invent medical specials, repair estimates,
   policy form numbers, statute numbers, case citations,
   or jury verdict comparators. If a fact is missing, ask
   for it or mark it "OPEN ITEM".

4. You always use plain English. Define legal and medical
   acronyms on first use.

5. You always acknowledge what you do not know.

6. You produce three primary outputs on request:
   - File note (past tense, third person, factual)
   - Evaluation memo (with sections: facts, liability,
     damages, reserve recommendation, open items)
   - Reservation of rights letter draft (NY-compliant,
     with placeholders for compliance review)

7. You never include real-person PII (SSN, DOB, full
   addresses) in your output even if it appears in the
   input. If PII appears, redact it as [REDACTED] in your
   output.

8. You always end every response with the disclaimer:
   "Draft only — requires licensed adjuster review."

You may ask clarifying questions before producing output
if material facts are missing.

Step 3: Knowledge Files

Upload (only if your carrier permits):

• A copy of your carrier's reservation of rights letter template (de-identified)
• A copy of your carrier's evaluation memo template
• An NY no-fault primer document (publicly available from many sources)

The GPT will reference these documents when answering relevant questions.

Step 4: Conversation Starters

Add suggestions like:

• "Draft a file note from these inspection notes."
• "Build an evaluation memo from the facts I am about to paste."
• "Draft a reservation of rights letter for late notice."
• "Pull the open items from this fact pattern."

Step 5: Capabilities

For a claims GPT, you usually want:

• Web browsing off (to prevent the GPT from pulling unverified online facts)
• Code interpreter off
• File generation off

Step 6: Privacy

Set to Only me during testing. Once it works well, share via link with your team. Never set to Public for a carrier-internal GPT.

Testing Your GPT

Before relying on a custom GPT, test it with five or six realistic but de-identified scenarios. Watch for:

• Does it follow the rules in your instructions?
• Does it stay in scope?
• Does it acknowledge missing facts rather than invent?
• Does it always end with the required disclaimer?
• Does it handle edge cases gracefully (out-of-state, non-auto, criminal acts) by redirecting?

If any test fails, edit the instructions and re-test.

Other Useful Custom GPTs for Insurance

A few high-value GPTs to build:

• FNOL Triage GPT. Coverage tier, severity, routing, reserve recommendation.
• Underwriting Submission Reviewer. Industry-specific exposure analysis, COPE worksheet, follow-up questions.
• Renewal Letter Drafter. Customer-facing renewal communications with tone and length controls.
• Coverage Letter Drafter. Acknowledgments, status updates, denials with strict language constraints.
• Discovery Response Drafter. For litigated claims, structured responses to common discovery requests (with attorney review).
• Training Scenario Generator. Practice claim scenarios for new adjuster training.

Maintaining Your GPT

A custom GPT is not "set and forget." Plan for:

• Quarterly review of instructions for new statutes, regulations, and carrier guidelines
• Annual refresh of knowledge files
• Audit trail of which GPTs are in use, by whom, for what purposes (NAIC AI Bulletin compliance)
• Sunset and replacement when carrier appetite or regulatory environment changes

A Reality Check

Custom GPTs are powerful, but they are still ChatGPT under the hood. They can hallucinate, they can drift from instructions in long conversations, and they cannot replace licensed judgment. Treat them like a strong junior colleague: useful, fast, and always supervised.

Key Takeaways

• A custom GPT lets you bake your scope, rules, templates, and persona into ChatGPT once and reuse them daily.
• The strongest claims GPTs have tight scope, clear safety rails, templated outputs, knowledge grounding, and a defined persona.
• Build, test, and iterate. Treat custom GPTs as living tools that need quarterly maintenance.
• A custom GPT is still an LLM. Licensed adjusters review every output before it leaves the carrier.