How Cara Built Industry-Specific AI for Insurance Brokers with AWS

Cara is a specialized AI platform for commercial insurance brokers, developed together with Amazon Web Services. In a technical write-up on the AWS Machine Learning Blog, the team described its architectural decisions and real-world deployment results, in what is one of the first detailed case studies of domain-specific AI in enterprise insurance.

Why brokers need their own AI

Commercial insurance brokerages handle thousands of complex documents: policies from different insurers, coverage requests, loss histories, regulatory forms and underwriting rules. A specialist has to keep a huge number of nuances in mind to quickly find the right coverage and correctly compare competing offers from different providers.

General-purpose language models handle this task poorly: they lack a deep understanding of industry terminology, standard exclusions and the specifics of individual insurers' products.

Cara was not built as a "fine-tuned" GPT for insurance, but as a system with domain expertise baked into the architecture from day one. That is a fundamental difference. Brokers at large firms spend a significant share of their working time searching for and analyzing coverage information, a task an AI assistant with the right knowledge base can perform many times faster and with fewer errors.

An AWS-based architecture

The key technical decision is a RAG (Retrieval-Augmented Generation) architecture: instead of storing knowledge in the model's weights, the system retrieves data from the corporate knowledge base in real time and builds its answer from it. This lets it work with brokerages' proprietary documents without sending confidential data to external APIs.

Choosing RAG over fine-tuning was a pragmatic call: insurance products change constantly, and retraining a model every time policies change is expensive and slow. With RAG, you simply update the knowledge base without touching the model itself.

The stack is built on several AWS services:

• Amazon Bedrock — managed access to language models without in-house ML infrastructure
• Amazon OpenSearch — semantic search across large collections of insurance documents
• AWS Step Functions — orchestration of multi-step workflows
• Amazon SageMaker — fine-tuning models on industry data
• Amazon S3 — storage and indexing of corporate documents

All processing stays inside a protected AWS environment, a critical requirement for an insurance industry with strict data protection standards.

What changed for brokers

Cara acts as an AI assistant, not a replacement for specialists. The system takes over routine search and first-pass analysis: it helps find the relevant policy terms faster, compare insurers' offers and prepare quotes. Final decisions remain with people. This approach reduces resistance to adoption in a corporate environment.

According to the company, the results were measurable: request processing time fell and the error rate in policy comparisons dropped. An additional effect is that expertise scales: new hires supported by Cara reach working efficiency faster, without months of immersion in product details.

What this means

The Cara case is an example of a durable trend: in regulated industries (insurance, healthcare, finance, law), vertical AI solutions with deep domain specialization win. Horizontal models, however powerful, do not replace industry expertise built into the architecture from day one.

AWS positions Bedrock as infrastructure for precisely these niche solutions, and Cara has become one of the first public proof points that this approach delivers concrete business results.