Agentic AI · OpenAI · Insurance
Five AI agents.
One claims pipeline.
Zero manual triage.
Clariva Group processes over 14,000 insurance claims per month. Every claim was being manually reviewed, routed, and responded to. Amroar built a five-agent OpenAI system that handles intake, document analysis, fraud investigation, decision-making, and claimant communication — end to end.
Project overview
Client
Clariva Group
Sector
Insurance · Claims Processing
Volume
14,000+ claims / month
AI Model
OpenAI GPT-4o
Agents
5 autonomous agents
Scope
End-to-end claims automation
GPT-4o
Assistants API
Embeddings
Whisper
Vision
Function Calling
73%
Of straightforward claims resolved without human involvement
4.2min
Average claim resolution time — down from 3.8 days
91%
Document extraction accuracy across PDFs, images, and handwritten forms
0
Failed deliverables across the full agentic AI engagement
The Challenge
14,000 claims a month. Every one of them manually triaged.
Clariva's claims team was spending the majority of their time on intake, document sorting, and initial routing — work that required judgment but not expertise. The experts were buried in administrative process, and claimants were waiting days for a first response on claims that should have been resolved in minutes. Scaling meant hiring, and hiring wasn't keeping up with claim volume.
01
Manual claim intake and routing
Every incoming claim — email, web form, or phone call — was manually read, categorised, and assigned to a team. No structured extraction, no automatic routing logic, no consistent classification across handlers.
02
Document review consuming adjuster time
Medical reports, police records, contractor quotes, and photographic evidence were reviewed manually. Adjusters spent hours reading documents to extract the same five data points needed for every claim decision.
03
No systematic fraud detection at intake
Fraud identification relied on individual adjuster experience. Patterns that would be obvious across thousands of claims were invisible to the person reviewing one case at a time. Fraud slipped through or flagged too late.
04
Claimant communication entirely manual
Every status update, information request, and resolution notice was individually drafted and sent by a team member. Response times were inconsistent, communication quality varied, and follow-ups depended on whoever was least busy.
Agent Architecture
How the five agents work together.
Each agent handles one stage of the claims pipeline. They pass structured context between them — not raw data. Every decision is logged and auditable. Humans stay in the loop for complex claims.
Claim Arrives
Email · Web Form · Phone (Whisper transcription) · API
▼
Agent 01
Intake Agent
GPT-4o · Function Calling
Reads claim, extracts structured fields, validates policy number, classifies claim type, routes to appropriate pipeline
Agent 02
Document Agent
GPT-4o Vision · Embeddings
Analyses PDFs, photos, medical reports. Extracts key data. Semantic search against policy wording. Flags coverage gaps.
Agent 03
Investigation Agent
GPT-4o · Embeddings · RAG
Cross-references claim against historical patterns via vector search. Scores fraud risk. Flags inconsistencies. Recommends human review threshold.
Agent 04
Decision Agent
GPT-4o · Function Calling
Generates settlement recommendation for simple claims. Drafts adjuster brief for complex claims. Triggers payment function or escalation workflow.
Agent 05
Comms Agent
GPT-4o · Assistants API
Sends status updates, drafts resolution letters, answers claimant queries in natural language. Escalates to human when tone or complexity warrants.
▼
Auto-Settled Claim
Payment triggered, letter sent, case closed
Human Adjuster Brief
Full context prepared, decision deferred
Fraud Investigation Referral
High-risk case routed with evidence summary
Powered by
OpenAI GPT-4o
Assistants API
text-embedding-3-large
Whisper
GPT-4o Vision
Function Calling
Pinecone Vector DB
Each Agent Explained
What each agent does — and how it does it.
Every agent has a single responsibility, a specific OpenAI capability, and a defined output that the next agent in the pipeline can act on. No agent is making decisions beyond its scope. Human escalation is built into the logic of agents 03 and 04.
01
Intake Agent
GPT-4o · Function Calling
Receives claims from every inbound channel and converts unstructured text into a consistent structured claim object — policy number, claim type, incident date, amount, claimant details — before routing to the appropriate pipeline.
Reads email, form submission, or Whisper-transcribed phone call
Calls policy database via function calling to validate policy
Classifies claim type (motor, home, health, commercial)
Routes to correct sub-pipeline with structured JSON payload
GPT-4o
Function Calling
Whisper
02
Document Agent
GPT-4o Vision · Embeddings
Processes every document attached to the claim — medical reports, police records, contractor invoices, site photographs. Extracts key data, checks it against the claim description, and flags inconsistencies before the claim reaches a human.
GPT-4o Vision reads PDFs, scanned forms, and photographs
Extracts amounts, dates, diagnoses, and supporting evidence
Semantic search against policy wording via embeddings
Outputs structured evidence summary for downstream agents
GPT-4o Vision
Embeddings
RAG
03
Investigation Agent
GPT-4o · Embeddings · RAG
Cross-references the claim against Clariva's full historical claims database using vector embeddings — finding semantic similarities to previously fraudulent claims, unusual patterns, or policy violation signals that individual reviewers would miss.
Embeds claim data and searches Pinecone for similar historical claims
Scores fraud risk on a 0–100 scale with reasoning
Flags specific inconsistencies with evidence references
High-risk score triggers mandatory human review — not auto-settlement
GPT-4o
Embeddings
Pinecone
04
Decision Agent
GPT-4o · Function Calling
Takes the structured output from agents 01–03 and generates a decision recommendation. For straightforward claims below a defined threshold with low fraud score, it triggers payment via function call. For complex or high-value claims, it prepares a briefing document for the human adjuster — not a blank file to start from, a fully evidenced brief with a recommended decision and reasoning.
Applies Clariva's coverage rules to claim data via structured reasoning
For eligible simple claims: calls payment API function, triggers settlement
For complex claims: generates adjuster brief with recommended decision and evidence
For high fraud-risk: routes to SIU with full investigation summary
Every decision logged with full chain-of-reasoning for audit
GPT-4o
Function Calling
Assistants API
05
Comms Agent
GPT-4o · Assistants API
Handles all claimant-facing communication throughout the claims lifecycle. Sends acknowledgements, status updates, information requests, and resolution letters — all in natural language, personalised to the claimant's specific claim. Uses the Assistants API to maintain conversation context so claimants can reply with questions and receive coherent, accurate responses without a human in the loop.
Generates personalised acknowledgement within minutes of claim receipt
Sends information request letters when documentation is missing
Answers claimant status queries via email thread — maintains context
Drafts resolution letters (approved or declined) with clear, plain-English explanations
Detects distressed or complex tone — escalates to human handler
GPT-4o
Assistants API
Function Calling
What Changed
The numbers after six months in production.
These are real operational outcomes measured at the six-month mark across Clariva's full claims volume. The agents handle 73% of claims end-to-end. The other 27% get to human adjusters faster and better-prepared than before.
73%
Claims resolved without human involvement
Straightforward claims — valid policy, clear documentation, low fraud score — are settled and communicated to the claimant entirely by the agent pipeline.
4.2min
Average resolution time for auto-settled claims
Down from 3.8 days. The same claim that previously waited in a queue for a human to open it is now processed, decided, paid, and communicated in under five minutes.
91%
Document extraction accuracy
GPT-4o Vision extracts structured data from PDFs, scanned forms, and photographs at 91% accuracy — validated against a manually reviewed test set of 2,000 documents before go-live.
3.1×
More fraud flags surfaced at intake
The investigation agent surfaces 3.1 times more potential fraud indicators than the previous manual review process — because it can cross-reference every claim against the entire historical dataset simultaneously.
What the client said
Our adjusters were spending most of their day reading documents and sending status emails. That's gone now. The agents handle everything that doesn't actually require expert judgment — and for the complex claims, they arrive with a full brief already prepared. Our adjusters are doing adjuster work for the first time in years.
NK
N. Kaur
Chief Operations Officer · Clariva Group
Work With Us
Ready to build an agent that handles the work your team shouldn't be doing?
Amroar has built and deployed agentic AI systems across insurance, logistics, healthcare, real estate, and professional services. Book a discovery call and we'll show you what's realistic for your operation.
Founders on every engagement. Not a junior consultant.
50+
Agentic AI implementations across industries — OpenAI, Claude, and custom model stacks
200+
Enterprise clients across US, UK, Australia and the Middle East
25+
Years combined experience — senior architects and AI engineers on every build
0
Failed builds across every agentic AI engagement we've delivered