Ertas + OpenClaw for Healthcare

Healthcare providers using OpenClaw for patient communication, clinical note processing and appointment management often send patient messages to cloud AI APIs. With Ertas, you fine-tune a model on de-identified examples of your own workflows, export it, and run it locally with Ollama, so inference on patient data stays on infrastructure you control.

The Challenge

Healthcare providers see immediate value in OpenClaw: automating appointment scheduling through messaging platforms, triaging patient inquiries, processing clinical notes, generating referral letters, and managing insurance pre-authorisation correspondence. These workflows take up a large share of front-desk and clinical admin time.

HIPAA's Privacy Rule limits how Protected Health Information (PHI) can be shared with third parties. When OpenClaw processes a patient inquiry through a cloud AI API, the patient's name, condition, treatment history and contact details go to that provider's servers as prompt input. The provider then has to be covered by the right contract and reviewed by the practice's privacy and security team, and many practices would rather keep patient messages off third-party AI servers altogether.

The Office for Civil Rights (OCR) has taken enforcement action against organisations that disclosed PHI through systems without adequate safeguards. Every message, note and appointment request an agent sends to a cloud API adds another data flow the practice has to account for.

Meanwhile, generic AI models frequently hallucinate on clinical terminology, misclassify triage urgency, and produce responses that do not align with a practice's specific protocols. The productivity gain from OpenClaw is undermined when staff spend time correcting AI-generated clinical content.

The Solution

With Ertas, healthcare providers run OpenClaw on a fine-tuned model that they host themselves. You train the model in Ertas Studio, export it from Hub as GGUF, and serve it on the practice's own hardware through Ollama. From then on, every inference request from OpenClaw goes to that local model, so patient messages are processed on the practice's own servers.

Training happens on Ertas's cloud GPUs, so the training set is uploaded to Ertas. Build it from de-identified examples of the practice's own workflows: its triage criteria, scheduling protocols, referral letter templates and patient communication style. A model fine-tuned on a dermatology practice's examples learns to tell urgent cases (suspicious lesion, rapidly changing mole) from routine ones (annual skin check, cosmetic consultation) the way that practice does.

Healthcare networks and agencies managing several practices can train a separate model per practice in Studio, each capturing that practice's communication style, clinical specialty and scheduling rules, and run each one on infrastructure the network controls.

Key Features

Hub

Local Inference with Ollama

Export the trained model from Hub as GGUF and run it with Ollama on the practice's own hardware. OpenClaw sends its inference requests to that local model, so PHI in patient messages is processed on your servers and stays out of cloud AI APIs.

Studio

Clinical Workflow Fine-Tuning

Fine-tune in Studio on de-identified examples of the practice's work: triage protocols, appointment scheduling rules, referral templates and patient communication patterns. The resulting model learns the practice's clinical context, which cuts down on hallucinations and improves accuracy on its own tasks.

Studio

Multi-Practice Deployment

Healthcare networks, management groups and agencies can train a separate model for each practice in Studio, run the training jobs in parallel and compare them. Each practice gets a model shaped by its own examples, deployed on infrastructure the network controls.

Studio

Staff Review in Data Craft

Practice staff build and correct training examples in Data Craft, Studio's guided dataset builder: they add USER and ASSISTANT turns, edit replies to match the practice's guidelines, version each example and approve it before it goes into training.

Example Workflow

A US dental practice group with five locations and 12 dentists deploys OpenClaw to automate patient communication over WhatsApp and SMS. Front-desk staff at each location currently spend hours every day answering appointment inquiries, sending reminders, handling rescheduling requests and answering pre-appointment questions about procedures and insurance coverage. The group exports 8,000 de-identified patient communication threads from its practice management system, covering scheduling, procedure inquiries, insurance questions and post-procedure follow-ups. The group's operations manager reviews a sample in Data Craft and approves the replies that match the group's communication guidelines. The dataset is uploaded to Ertas Studio, where a Qwen 2.5 7B base model is fine-tuned with LoRA. The team checks the model against a held-out set of threads, scoring intent classification (booking, rescheduling, inquiry, emergency) and reply quality against its own guidelines. Each location's scheduling availability, dentist specialties and insurance panels are passed to the agent with each request from the practice management system. The model is exported from Hub as GGUF and served with Ollama on a Mac Mini M4 Pro at the group's central office, and OpenClaw instances at each location connect to it over the group's private network. Staff review and approve AI-drafted replies for complex cases. Inference on patient messages runs on the group's own hardware, and the group reviews its WhatsApp and SMS providers as messaging services under its own HIPAA program.

Compliance & Security

Meeting HIPAA, and any state health privacy law that applies, is the practice's own work. Here is what this setup changes. With the exported model running locally through Ollama, OpenClaw's inference requests go to the practice's own servers, so PHI in patient messages is processed there and stays out of cloud AI APIs. Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas: de-identify it under Safe Harbor or Expert Determination before upload. Datasets are stored in Studio so runs can be repeated, and you can delete them at any time. Access controls, logging and the messaging channels themselves (WhatsApp, SMS) remain part of the practice's own security and privacy review.