- Services
Product Management
Product Development
Product Capability
- Client Stories
- Insights
- About
- Contact Us
Medbill Medical Billing Platform
Terem designed and built “Clara”, an LLM-powered medical billing engine for Medbill, an Australian scale up. Leveraging AWS Bedrock and Anthropic foundation models, the full-stack platform automates the processing of Medicare, private health fund, and workers compensation claims for physicians, surgeons, anaesthetists, and medical billing staff, reducing processing time from weeks to seconds, materially reducing claim rejection rates and ensuring compliance across every submission.
Medical billing in Australia is a complex process spanning Medicare, private health funds, and workers compensation schemes. Clinicians and billing staff face a tangle of ever-changing claim codes, compliance rules, and payer-specific requirements. The consequences of errors are significant: rejected claims, delayed revenue, and compliance risk.
Medbill set out to solve this with an AI-driven engine, but the path from vision to product presented several formidable challenges:
Perhaps the most nuanced challenge was architectural, how do you build a system that relies on AI as a core component, yet keeps a human firmly in the loop? The team needed to wrap the inherent non-determinism of LLM outputs with a layer of predictability, giving billing staff confidence in the system’s recommendations while maintaining their ability to review, override, and approve every claim and stay across every rule change.
Medical billing codes demand precision. A single incorrect code can mean a rejected claim or an audit. Training an LLM to reliably map clinical notes to the correct billing codes across Medicare / Private Health funds required extensive prompt engineering, evaluation frameworks, and rigorous testing against real-world billing scenarios.
Medical billing data is among the most sensitive information in any system. The platform needed to meet strict privacy and security standards from day one, ensuring patient and billing data was handled with the appropriate safeguards throughout the AI processing pipeline.
Medical billing codes demand precision. A single incorrect code can mean a rejected claim or an audit. Training an LLM to reliably map clinical notes to the correct billing codes across Medicare / Private Health funds required extensive prompt engineering, evaluation frameworks, and rigorous testing against real-world billing scenarios.
Terem embedded a cross-functional AI native squad to take Medbill’s Clara from concept to a live production platform. Terem’s Claude Code Pipeline and Processes were used by our engineering team to deliver at pace.
The centrepiece is a purpose-built AI processing pipeline where Anthropic’s foundation models analyses government documentation about the Medicare Benefit Scheme. The pipeline combines this information with each private health fund’s specific requirements and any additional context (like known types of rules). The result is a deterministic set of rules that can be consumed by the rest of the application. Every AI-generated rule flows through a review interface where billing professionals inspect the reasoning and edit/override the rules. These edits form part of the context for future runs (the MBS updates quite regularly).
AWS Bedrock was chosen deliberately for its enterprise-grade security, data privacy controls, and managed scaling essential for a healthcare application. The decision to use managed foundation models (in this case Anthropic’s Sonnet and Haiku) was deliberate – allowing Medbill to benefit from model advances without sacrificing security or operational control. Unlike end-to-end “black box” AI systems, Clara separates reasoning, rule formalisation, and execution into controlled stages.
The platform has been in continuous use since July 2025, improving the speed and accuracy of Medbill’s service to doctors.
What previously took operations staff days or weeks of manual processing is now completed in seconds by the AI engine, with the billing professional reviewing and approving the output.
By combining Anthropic’s reasoning with structured validation rules, the system catches coding errors that manual processes could miss.
For a significant portion of claims, the platform cuts human effort by more than half, freeing operations staff to focus on complex edge cases and higher-value work.
