The AI scribe helping 2.7 million patient visits a week, and why its database matters as much as its model

Key points

• Heidi's AI scribe supports roughly 2.7 million patient interactions per week across more than 190 countries, as of 2025.
• A pilot at Beth Israel Lahey Health, one of New England's largest hospital networks, found 74% of clinicians reported less after-hours paperwork after using Heidi.
• Heidi stores all patient data inside the same region where care was delivered, enforced by infrastructure rather than policy promises.
• Migrating to MongoDB Atlas cut response times on key parts of the app by nearly 33%, according to the company.
• Heidi co-founder and CTO Yu Liu says the AI model accounts for roughly 20% of what makes the system work; data architecture carries the rest.

A doctor in London finishes a consultation. Within seconds, a structured clinical note appears. The doctor reviews it, edits a line, and moves on. That's what Heidi Scribe does, and it's doing it roughly 2.7 million times a week.

Heidi is an Australian company that builds AI tools for clinicians. Its flagship product listens to a consultation and automatically drafts the documentation that follows: the notes and structured records that can consume hours of a doctor's day. The company now operates in more than 190 countries.

What makes healthcare AI different from every other kind?

A two-percent error rate in most software is a minor nuisance. In a clinical setting, it's a patient safety problem.

"The architecture has to be built around the assumption that every output may be scrutinised and relied upon in a patient's care," says Yu Liu, Heidi's co-founder and chief technology officer.

That shapes every engineering decision Heidi makes. Data residency, the rule that a patient's records must stay physically within the country where care happened, isn't optional. A clinician in Sydney works under Australia's Privacy Principles. One in London works under GDPR, the European Union's data protection law. Tokyo brings Japan's APPI; Denver brings HIPAA, the US health data law. Heidi runs a fully isolated deployment in each region to satisfy all of them.

Auditability is equally non-negotiable. If a regulator asks what the AI saw during a session months later, what it produced, and what changed, Heidi must answer. That capability has to be designed in from the start.

Why the database choice was a founding decision

A single Heidi Scribe session isn't one tidy file. It's a collection of transcripts, structured notes, patient context, records from electronic health systems, and other pieces that all belong together and change shape as the product evolves.

Traditional databases store data in rigid rows and columns. That structure breaks whenever the data model needs to change, which in a fast-moving AI product is constantly. Heidi chose MongoDB Atlas, a cloud database service that stores data as flexible documents rather than fixed rows, letting the team reshape records without freezing the whole system.

"The model is maybe 20% of the system, and the data architecture is what determines whether the other 80% holds up under real clinical load," Liu says. Switching to Atlas cut latency on key APIs by nearly 33%, the company says.

It's worth comparing this to the broader pattern we've seen across regulated industries. Our 17 July story on NVIDIA's Nemotron models found hospitals shaping general-purpose AI into specialised tools precisely because off-the-shelf infrastructure doesn't survive contact with compliance requirements. Heidi's database decision is the same instinct applied one layer deeper.

How the AI actually retrieves clinical knowledge

Heidi uses a technique called RAG, short for retrieval-augmented generation. Instead of asking a model to recall medical facts from training data (where it might produce something wrong with complete confidence), RAG pulls verified information from trusted sources and hands it to the model as context.

Heidi's sources include licensed clinical knowledge bases from organisations such as BMJ Best Practice and NICE CKS, a UK clinical guidance body. The system is jurisdiction-aware: a UK clinician gets UK treatment guidelines, while an Australian clinician gets Australian drug formularies. Citations are bound to their source records before the model ever sees them. The AI can't invent a reference.

This design matters more than it might appear. We've covered fake AI doctors racking up millions of TikTok views with fabricated health claims. Heidi's hard-coded citations are an architectural answer to the same failure mode, applied inside a professional clinical workflow.

Should you worry about what happens to your data?

Each region runs its own fully isolated database cluster with its own compute and its own encryption keys. Patient data physically cannot cross a regional boundary. That's an infrastructure guarantee, not a contractual one, and it's the distinction that lets Heidi walk into a US health system or an NHS trust and give a clean answer on residency.

For clinicians, less paperwork time means more time with patients, or simply finishing at a reasonable hour. The Beth Israel Lahey Health pilot put a number on it: 74% of participating clinicians said after-hours documentation fell after using Heidi. For patients, the promise is a doctor present in the room. The risk, as with any AI clinical tool, is over-reliance on outputs that haven't been carefully reviewed. Every note goes through the clinician before it enters the record.

Common questions

Is Heidi approved as a medical device?

Heidi markets its product as a documentation assistant, not a diagnostic tool. Regulatory classifications vary by country; clinicians remain responsible for reviewing every note the system produces.

Can a hospital trust that patient data stays in their country?

Heidi says yes, enforced by infrastructure: each region runs its own fully isolated database cluster. The data physically cannot cross a regional boundary.

Does using an AI scribe mean the AI is making clinical decisions?

No. Heidi Scribe records and drafts documentation. The clinician reads it, edits it, and signs off. The AI doesn't diagnose or prescribe.