tariq remington

amsterdam

AI-assistant for Health Insurance Policy Q&A

Ask your health insurance policy a question. Get a real answer in under a minute.

5 min read

Challenge

Health insurance policies are long and full of legal language. Important rules sit across many sections. Answering a single question can take 15 to 30 minutes of careful reading. The aim was to build an AI-assistant that could give quick and accurate answers.

Goal

Be able to ask questions about a health insurance policy and get accurate answers with references to the exact relevant policy sections fast.

Proposed Solution

An AI-assistant reads the policy documents and answers questions only from the policy. Each answer shows the exact sections it used. This implementation runs locally so no data leaves the system.

Because dealing with own personal and medical data, the decision was made to opt for designing a system completely running on a local system infrastructure. Although perfectly possible to implement local infrastucture, the common practise for organizations is using either a hybrid (local & cloud) or cloud solution like Amazon Web Services, Microsoft Azure of Google Cloud Platform for working with data and calling models. Each approach comes with its own trade-offs.

Real-world relevance

Concept validation

Shows that AI can answer only from a specific document, making it suitable for regulated or sensitive domains.

Applicable to

Organizations

Health insurersHealthcare providersLegal practicesHR consultancies

Departments

Customer serviceComplianceLegalHR

Roles

Policy officersCompliance advisorsHR advisorsCustomer support agents

Skill transfer

Could work for any organisation with complex documents people must search, like terms and conditions, compliance manuals, medical guidelines, HR policies, internal knowledge bases.

Industry pattern

Any team that answers repetitive document questions can gain speed, consistency, and auditability with this approach.

Plan

  1. 1Prepare the policy documents so the right sections can be found fast. Without clean structure, answers would be vague or inaccurate.
  2. 2Limit the assistant to only answer from the policy text available. This keeps answers safe and avoids guessing or hallucinations.
  3. 3Show every answer with the exact policy sections it came from. This makes it easy to verify and builds trust.
Health Insurance Policy RAG — user interaction flow from question input to cited answer

User flow: a question enters the system, the assistant finds the most relevant parts of the policy, and returns a grounded answer with citations.

Impact

Time to answer a policy question

<1 minute

compared to 15 to 30 minutes of manual searching and reading

What I Learned

  • –Insurance language is precise. Small phrasing changes can change the coverage, so answers must accurately represent the conditional logic in the policy.
  • –Every answer needs a source. Confidence without proof is undesired in healthcare or legal contexts.
  • –Running locally removes data privacy risk, which is important for healthcare related systems that could contain sensitive information.

How documents are structured affects answer quality

Rules often span multiple paragraphs with conditions elsewhere. Understanding the structure helps appropiately designing the search functionality of the assistant.

Challenges & Solutions

Policy documents are not written to be searched

Coverage rules and exceptions sit in different sections and use different phrasing.

How I tackled it

Each piece of text was stored with closely surrounding context so the assistant could find the rule and its conditions together.

Preventing the assistant from making things up

A general Large Language Model (LLM) will try to answer even when there is no proof based on the policy.

How I tackled it

System design and strict instructions force answers to stay relevant to the policy.

Keeping language clear for policy holders

Policy wording can be vague. Answers should be accurate but easy to read.

How I tackled it

Returned text was simplified without changing meaning, and every claim links back to the original section for quick checking.

Required resources

Data

Existing policy, contract, or compliance documents

Infrastructure

Local file storage (no cloud required)A standard on-premise computer or server

Expertise

A domain expert to review and validate outputs

Technical Approach

Programming Languages

Python

Frameworks & Libraries

LangChain
ChromaDB
Ollama
Health Insurance Policy RAG — system architecture showing PDF ingestion, ChromaDB embedding, retrieval pipeline, and Ollama LLM

System architecture: policy documents are processed and stored locally. When a question comes in, the system retrieves the most relevant sections and passes them to the AI model with strict instructions to stay within the document.

Want to get in touch?

If this project is relevant to something you’re building, you have any quesions, feedback or interested in having a chat, feel free to reach out.