MSFT Copilot Health goes FHIR 🔥, Lovable hits $400M ARR 🚀, Vision scribes see more 👓
🔬 The Big Thing
Microsoft just shipped the largest FHIR consumer aggregation product ever built — and it wasn’t aimed at clinicians.
Yesterday Microsoft launched Copilot Health, a new module inside their Copilot assistant that connects to a user’s health records, wearables, and lab results. The data access story is what caught my attention: it pulls from HealthEx, which has TEFCA-compliant FHIR connections spanning 52,000 healthcare organizations. Oura, Fitbit, Function Health labs — it’s all in there. Microsoft is positioning it as “prepare for your doctor’s appointment, not replace your doctor.” The tool serves up Harvard Health answer cards, finds in-network providers, and gives your patients an AI to interpret their lipid panel before their cardiology follow-up.
Here’s what I keep turning over: this is FHIR consumer access at a scale nobody has done before, and the product it’s powering is entirely patient-facing. Not a clinical tool. Not a builder tool. A thing your patients will use before they walk into your exam room.
What happens to the clinical encounter when the patient has already asked Copilot Health why their ferritin is low and it told them to “eat more red meat and consider supplementation” — before a workup that might reveal celiac disease, or occult blood loss, or a hematologic problem nobody’s looked for yet? The clinical context problem runs in both directions now.
The builder angle: HealthEx’s architecture — TEFCA individual access services plus direct FHIR endpoints across 52,000+ organizations — is the best public proof that FHIR consumer access at scale is a solved problem. Not a regulatory ambition. A production system Microsoft is staking a consumer product on. If you’ve been building patient-facing tools and punting on the “how do we get the patient’s real records” problem, that problem has gotten materially easier. The standards exist. The infrastructure exists. What still doesn’t exist is someone with clinical judgment deciding what to do with the data once you have it. That’s the unfair advantage nobody at Redmond has.
One thing I’m genuinely uncertain about: what happens to HIPAA liability when an LLM interprets PHI in this context? Microsoft is explicit that this doesn’t provide diagnoses or treatment. But when Copilot Health tells a patient their potassium of 5.4 is “slightly above normal, which can sometimes indicate kidney issues” — is that medical advice? I think the legal answer right now is “we’ll find out in litigation.”
Microsoft announcement · Fortune · The Verge · HealthEx partnership
📡 Builder’s Radar
Lovable hits $400M ARR — vibe coding is no longer a curiosity, it’s a market.
Lovable, the platform that turns natural language descriptions into production-ready apps, just hit $400M annual recurring revenue — up 33% in a single month. Meanwhile, Replit launched Agent 4 with parallel AI agents that build backends and frontends simultaneously, and Cursor is reportedly in talks for a $50B valuation. These aren’t research projects. This is a multi-billion-dollar ecosystem built on the premise that the engineering barrier has collapsed. Every one of these numbers is evidence that the clinicians.build thesis is playing out: if you have domain expertise and a problem worth solving, the “but I can’t code” objection is increasingly irrelevant.
Vision-enabled AI scribes catch what audio-only scribes miss — and it’s a lot.
A new paper in npj Digital Medicine built a scribe using Google’s Gemini model and Ray-Ban Meta smart glasses to document medication histories — because med rec requires seeing the bottle, the label, the tablet count, the inhaler the patient holds up. Audio-only transcription of “yeah I take the little blue one twice a day” is exactly as useful as it sounds. If you’re building in the documentation space and your architecture only processes audio, you may be systematically missing the things that will bite you clinically. The medication reconciliation problem — the one that causes a real and measurable percentage of adverse events — is fundamentally a multimodal problem.
Oracle reportedly cutting up to 30,000 jobs to fund AI data centers — and Oracle Health is in the blast radius.
Oracle is reportedly planning to cut 18% of its workforce to free up $8-10B for AI data centers. Stock is down 54% from its September high. This matters for clinician-builders because Oracle Health (née Cerner) is Oracle’s healthcare business. If you’re at an Oracle Health site and wondering about your EHR vendor’s AI roadmap, the honest read is that Oracle is in financial triage mode — pouring cash into AI infrastructure while the health tech side competes for attention internally. Not the time to bet on Oracle Health delivering your agentic workflow tools.
OpenAI published a framework for defending agents against prompt injection — read it if you’re building clinical AI.
OpenAI outlined how prompt injection attacks against agents increasingly resemble social engineering, and argued that defenses should focus on limiting the impact of successful manipulation, not just detecting malicious inputs. If you’re building any agent that touches patient data or clinical workflows, this is required reading. The attack surface for clinical AI agents is real: imagine a malicious referral note that instructs your triage agent to escalate everything to the ED. The defense architecture matters.
🛠️ From the Workbench
HealthEx as reference architecture for patient record aggregation — If you’re building any patient-facing tool that requires pulling a user’s actual health records, HealthEx’s approach is worth understanding: TEFCA individual access services plus direct FHIR endpoints across 52,000+ organizations, with a consumer-facing consent model. They’re not an open API, but their architecture is a useful template for solving the “get the patient’s records” problem in production. HealthEx
What are you building this week? Reply and tell me — I read every one.
— Kevin