# Building Agents for Healthcare

Lars Maaløe, Corti | Agents in Production 2025 | 30:34

Source: https://www.youtube.com/watch?v=f9QCWHKF7qc
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/building-agents-for-healthcare
Published: 2025-09-19
Tags: agents, governance, healthcare, synthetic-data

## TL;DR
- Healthcare clinicians can spend 25% to 50% of their time on documentation and administrative work, which Corti aims to reduce with healthcare-specific AI.
- Corti combines a healthcare foundation model, patient and guideline memory, workflows, and specialized experts instead of relying on a single general-purpose language model.
- Healthcare agents need governance, auditing, deployment controls, and domain-specific models because clinical errors and data handling have regulatory consequences.

## Summary
Lars Maaløe describes Corti's approach to building AI agents for healthcare. He starts with the pressure on clinicians, including growing patient volumes, more complex treatments, and administrative work that can take 25% to 50% of a clinician's time. Corti combines audio and text models with patient context, clinical guidelines, workflows, and specialized experts. One example is FAX R, which extracts clinical findings from long patient records and conversations for downstream tasks. An orchestrator interprets a user's request, retrieves information, calls the relevant expert, and applies a governance layer with auditing. Maaløe explains that experts are built from several model types, including classifiers and reasoning models, rather than from prompts alone. He also discusses data anonymization, synthetic data, regulatory status, tenant isolation, sovereign cloud deployment, and the need to support environments without internet access. The talk is specific about the engineering complexity behind healthcare agents and cautious about diagnostic use.

## Key ideas
### Healthcare administrative work is a direct target for AI assistance
[01:17](https://www.youtube.com/watch?v=f9QCWHKF7qc&t=77s)
Maaløe frames healthcare as a sector with too few clinicians for the number of patients, while ageing populations and more complex treatments increase demand. He says clinicians face burnout partly because of documentation and other administrative duties. Studies cited in the talk report that clinicians spend between 25% and 50% of their time on documentation and administration. Those tasks are separate from the patient-facing work for which clinicians are trained. Corti's stated goals are to protect patient safety, reduce administrative work through automation, and keep documentation as complete and accurate as possible. Maaløe also says the number of healthcare administrators has grown relative to physicians, while earlier digital systems have not solved the burden.

### Corti combines a healthcare foundation model with clinical memory and workflows
[03:59](https://www.youtube.com/watch?v=f9QCWHKF7qc&t=239s)
Corti's system starts with a foundation model trained on audio, text, and classification tasks. Maaløe describes a combination of text language models, audio models, and classifiers that communicate through an orchestration of model parameters. The model is connected to a large memory containing historical patient interactions, structured diagnoses, laboratory results, and other patient information. It also uses guideline data covering best practices and triage. On top of this foundation and memory, Corti builds workflows and tools for triage, physician dictation, ambient documentation, and research. The capabilities are exposed through an API with streaming and asynchronous interfaces for applications that use text or voice agents.

### FAX R extracts clinical facts before other healthcare agents use them
[08:31](https://www.youtube.com/watch?v=f9QCWHKF7qc&t=511s)
Maaløe introduces FAX R as a reasoning agent that can process healthcare dialogue and patient records, including records that run to hundreds of pages. Its job is to find and extract clinical findings for downstream tasks. He presents this as a response to the difficulty of getting the right clinical information out of a consultation. In the example workflow, a system can use extracted findings to summarize an interaction, assign a diagnosis code, and generate follow-up questions for the clinician to ask. FAX R can work across different sources of information, after which other agents can operate on the extracted facts. The talk presents clinical fact extraction as a separate capability rather than leaving every downstream agent to interpret raw dialogue or records.

### An orchestrator routes requests through governed specialist agents
[10:20](https://www.youtube.com/watch?v=f9QCWHKF7qc&t=620s)
Maaløe contrasts Corti's workflow with a user sending a prompt directly to a general-purpose language model. He says a general model may produce a convincing answer even when it should say that it does not know. Corti's orchestrator interprets the request, sits behind a governance layer, and audits information entering and leaving the system. For a request to find mentions of diabetes in a transcript, the orchestrator can retrieve the transcript and call an interaction expert. A later request to predict a diagnosis code can route to a revenue cycle management expert, which uses clinical guidelines and coding databases. The orchestrator can also ask the user for more information and combine several prompts in one workflow.

### Healthcare experts use several model types and domain resources
[24:02](https://www.youtube.com/watch?v=f9QCWHKF7qc&t=1442s)
When asked how Corti creates its experts, Maaløe rejects the idea that each expert is just a different prompt over one model. He describes a revenue cycle management flow with classification models built on an LLM body, a large corpus of around 140,000 codes, and a reasoning model that checks classification results against indexes and updated coding guidelines. The reasoning model can explain why codes should change, including when new guidelines prohibit combinations of codes. Other expert flows can retrieve research publications, guidelines, forms, or electronic health record information. Corti benchmarks different models and keeps testing whether a new out-of-the-box model performs better for a language or task. Maaløe calls this layered accuracy part of Corti's intellectual property.

### Healthcare data work depends on access, anonymization, and synthetic examples
[18:03](https://www.youtube.com/watch?v=f9QCWHKF7qc&t=1083s)
Maaløe says Corti first entered virtual healthcare communication because those interactions created a practical way to access many data points. The company worked on anonymization and de-identification to remove sensitive information. It is also building synthetic data from patterns learned from real-world examples. For coding, the large code distribution helps Corti identify where more examples are needed, then generate data resembling real cases in those areas. This gives the company another way to expand training data after learning from actual healthcare interactions. The approach is tied to specific healthcare tasks, such as coding, rather than being described as a general source of unrestricted training data.

### Healthcare deployment needs audit trails, isolation, and local control
[21:18](https://www.youtube.com/watch?v=f9QCWHKF7qc&t=1278s)
Maaløe says Corti encrypts data in transit and at rest and keeps audit logs for interactions. The logs can show which answer was produced for a prompt and which experts contributed to it, creating a trace of the interaction. Corti can deploy its technology in separate environments and isolate tenants. He gives a Swiss example where healthcare data must remain inside Switzerland, so the full stack can run there. The company can also deploy inside a customer's own firewalls, including environments that do not connect to the public internet. These controls address healthcare providers that need more than a shared online service and want to control where patient data and AI components run.

### Corti treats diagnostic use as a separate regulatory decision
[19:26](https://www.youtube.com/watch?v=f9QCWHKF7qc&t=1166s)
Maaløe says Corti's product is not registered as a medical device in the United States and says it should not be one for its current US use. He adds that the company is considering a medical-device application for capabilities related to diagnosis. Corti has research showing that its models can assist with diagnosis, but he is careful about how that claim is described because diagnostic use would require regulation. He says the UK and Europe have different requirements, including a class-one medical-device requirement for some ambient documentation use cases in the UK. Corti has a quality-management system in place, while the regulatory environment continues to change.

## Notable quotes
- "Healthcare is definitely a place that could need and benefit from some insightful and intelligent technologies." (00:59)
- "The big promise for this safe healthcare AI is first and foremost to focus on patient safety and secondly to ensure that the administrative burden is lowered." (03:40)
- "You need these very tailored models that are fine-tuned and structured towards the healthcare domain." (16:15)
- "I would love it if it was just one LLM that we could set another prompt for and then we could solve the use case." (24:24)

## Tools & references mentioned
- Corti
- FAX R
- MLOps Community
- Databricks
- FDA
- Switzerland
- UK
- Europe
- United States
- MCP

## Who should watch
- You are building clinical documentation, triage, coding, or research software and need to understand why healthcare agents require specialist components.
- Your team is evaluating healthcare AI and needs practical discussion of audit logs, tenant isolation, sovereign deployment, and regulatory boundaries.
- You work on model or data architecture and want an example of combining classifiers, reasoning models, retrieval, guidelines, and workflow orchestration.

## Related talks

- [Designing AI Agents for the Complex Realities of Healthcare](https://mlopstalks.com/talks/designing-ai-agents-for-the-complex-realities-of-healthcare) (Dr. Sarah Gebauer, Validara Health, 16:03)
- [AI in Healthcare](https://mlopstalks.com/talks/ai-in-healthcare) (Eric Landry, Zeteo Health, 51:06)
- [The Future of Healthcare: AI is Here](https://mlopstalks.com/talks/the-future-of-healthcare-ai-is-here) (, 24:56)
- [Governance for AI Agent Deployment](https://mlopstalks.com/talks/governance-for-ai-agent-deployment) (Spencer Reagan, Airia, 54:18)
- [Architecting Modern AI Systems: Platforms, Agents, and Integration](https://mlopstalks.com/talks/architecting-modern-ai-systems-platforms-agents-and-integration) (Allen Roush, BuzzHPC & Frédéric Bénard, Mila & Shuo Wang, Bell Canada, 57:00)