Healthcare AI & Agentic Systems

Agentic AI workflows deployed in production clinical environments. Claude and Gemini APIs routed through zero-retention endpoints. GraphRAG over clinical documents. HIPAA-compliant endpoints with a signed BAA before any PHI touches a model.

AI in healthcare is a compliance minefield

Most AI vendors train on your data. Read their terms of service. The big consumer LLM providers reserve the right to use your inputs to improve their models. In healthcare, that means Protected Health Information could end up in a general-purpose model's training set. That is a HIPAA violation waiting to happen, and the fine is not hypothetical. OCR has levied penalties in the millions for less.

The other problem is trust. A clinician will not act on an AI summary if they cannot verify where the answer came from. "The model said so" is not documentation. You need grounding: the ability to trace every AI-generated claim back to the source document, the specific section, the specific patient record. Without that, you have a chatbot, not a clinical tool.

Then there is the agentic problem. When an AI agent can write code, submit pull requests, modify data pipelines, or trigger workflows, you need guardrails that are mechanical, not aspirational. A CI gate does not care whether the PR came from a senior engineer or Claude. The same tests run either way. If your AI agent can bypass your CI gate, you do not have an AI strategy. You have a liability.

Most organizations I talk to are stuck at the "we know we need AI" stage. They have a few people experimenting with ChatGPT. They have no BAA with any vendor. They have no policy for what data can go to a model and what cannot. They have no monitoring for when an AI agent goes off-task. They need someone who has actually deployed agentic AI in a HIPAA-governed environment, not someone who has watched a demo.

My approach: zero-retention, BAA, human-in-the-loop

Every AI deployment I build starts with the same non-negotiables:

• Zero-retention API endpoints. I route Claude and Gemini through enterprise API configurations that contractually prohibit the vendor from retaining, training on, or accessing your data. No exceptions. If the endpoint does not support zero-retention, it does not get used.
• Business Associate Agreement. Before any PHI touches a model, a BAA is signed. The BAA specifies the permitted uses, the data handling requirements, and the breach notification obligations. No BAA, no PHI. This is not a recommendation. It is the law.
• Human-in-the-loop by design. AI agents do not merge to production. Agents write code, run tests, open pull requests. A human reviews every PR, checks the CI output, and approves the merge. The agent cannot bypass the gate. This is the same pattern I use for the dbt platform, and it works for the same reason: mechanical guardrails do not get tired, distracted, or overconfident.
• CI/CD gates for agent PRs. Every pull request, whether from a human or an autonomous agent, runs through the same CI pipeline. Compile checks, tests, contract validation, code review. The branch prefix tells you who opened it. The CI gate does not care.
• GraphRAG for grounded responses. Instead of asking an LLM to answer from its training data, I build a retrieval graph over your clinical documents, policies, and guidelines. The agent retrieves relevant passages, grounds its response in them, and cites the source. The clinician can verify the answer by clicking through to the original document.

What zero-retention actually means

Anthropic's Claude for Enterprise and Google's Vertex AI both offer zero-data-retention configurations. When enabled, the API processes your request, returns the response, and does not store the input or output beyond the transaction. No training. No logging beyond what you explicitly configure. No human reviewers reading your prompts.

This is table stakes for healthcare. If your AI vendor cannot offer zero-retention, find one that can. Anthropic and Google can. AWS Bedrock can. Azure OpenAI can. There are options. Picking a consumer chatbot with no enterprise controls is a choice, not a constraint.

GraphRAG over clinical documents

Retrieval-Augmented Generation is standard. GraphRAG adds structure. Instead of flat vector search over documents, GraphRAG builds a knowledge graph that captures relationships between entities: patient conditions, medications, providers, facilities, clinical guidelines, policies.

When a clinician asks "what is the current care gap for this member's diabetes management," the agent traverses the graph to find the member's diabetic panel status, their last HbA1c date, the HEDIS measure threshold, and any recent lab results. It grounds the answer in specific data points and cites each source. The clinician sees not just the answer but the evidence behind it.

This is the difference between a chatbot and a clinical decision support tool. The chatbot gives you an answer. The GraphRAG system gives you an answer with proof.

Deliverables

Every engagement produces concrete artifacts. Not a slide deck. Not a proof-of-concept that dies when the pilot ends. Production-deployed systems with monitoring.

• Deployed AI workflows. Production agentic systems running in your environment, routed through zero-retention endpoints, integrated with your data platform and clinical workflows. Not a demo. Not a prototype. Running code.
• Governance framework. A written policy for AI use in your organization: what data can go to which models, what requires BAA, what the human-in-the-loop workflow looks like, what the escalation path is when an agent produces incorrect output.
• CI/CD pipeline for agent PRs. The same guardrails I use on the dbt platform: agents open PRs on agent/* branches, CI validates them, humans review and approve. The agent cannot merge to main. The agent cannot bypass tests. This is mechanical, not aspirational.
• Monitoring and alerting. Production monitoring for AI workflow health: response latency, grounding rate (how often the agent cites sources vs. hallucinating), error rates, and anomalous behavior detection. If an agent starts generating output that deviates from its training patterns, you get an alert.
• Documentation. Every AI workflow documented with its purpose, data sources, model endpoints, BAA status, and runbook. If someone on your team needs to understand or modify the system six months from now, the documentation exists.

Proof points

These are not hypothetical capabilities. This is work I have shipped.

• Production agentic AI. Claude and Gemini workflows running in a HIPAA-governed clinical environment through zero-retention endpoints with a signed BAA. GraphRAG over clinical documents for grounded, cited responses. This is not a lab experiment. It handles real clinical queries in production.
• CI/CD for human and agent PRs. The same pipeline I describe on the data engineering page runs for AI agent code changes as well. Agents submit PRs on agent/* branches. The CI gate validates them identically to human PRs. Same tests. Same contract checks. Same human review before merge. I have written extensively about this architecture because it works.
• Zero-retention endpoint architecture. I have designed and deployed API routing that ensures no PHI reaches a model without zero-retention guarantees and a BAA in place. The architecture is documented, the contracts are signed, and the monitoring confirms it.

Related reading

I have written about the compliance, operational, and technical details of healthcare AI in depth: