Subscribe | Advertise | PRO Member Hub

TOGETHER WITH HEIDI

Perspectives from the people building the future of health AI…

We sat down with Dr. Thomas Kelly, CEO and Co-Founder of Heidi, to unpack how the company is moving from ambient documentation toward a broader AI Care Partner platform.

Tom is a doctor by training and previously specialized in vascular surgery, where he saw the downstream consequences of delayed diagnosis, fragmented primary care, and a healthcare system stretched beyond capacity. That clinical background shows up clearly in how Heidi is built. The company is not just trying to generate better notes, it is trying to build AI that supports clinicians before, during, and after the visit.

Heidi is now used across more than 190 countries, supports more than 110 languages and handles millions of consults each week.

Tom shared why the scribe was only the starting point, how Heidi thinks about regulation and medical device standards, why clinical AI needs different model behavior than consumer AI, and what agent-driven healthcare could look like when clinicians remain firmly in the loop.

Let’s start at the beginning. What’s the founding story behind Heidi?
I’m a doctor by background. Before Heidi, I was specializing in vascular surgery, which meant I saw patients at two very different points in the healthcare system.

On one side, you have very chronically unwell patients who often had years of poorly managed primary care behind them. High blood pressure, high cholesterol, smoking, diabetes, all the things that compound over time. By the time they reached vascular surgery, they were often at the end of the road.

On the other side, you have emergencies where every hour matters. Someone loses blood supply to their leg, and if the diagnosis is delayed, the outcome can be devastating. I saw patients lose limbs because the right diagnosis wasn’t considered early enough in the emergency room.

That was the original insight. It felt strange that clinicians were still doing history taking and initial diagnostic reasoning without more support. Even in 2017, if you typed the right symptoms into an AI system, you could often surface the correct diagnosis. I had studied computer science and math before medicine, so I had always been interested in AI. When transformer models started emerging, it became obvious that this technology could eventually become a partner to clinicians.

We started Heidi in 2021 with that broad idea: build an AI partner for clinicians. We didn’t know exactly what the first product or go-to-market would be, but the mission was clear: support clinical reasoning, reduce administrative burden, and ultimately expand the healthcare system’s capacity to care for more patients.

How do you describe Heidi today, and how has the company evolved?
Heidi started with an AI medical scribe because documentation was the most immediate pain point for clinicians and the richest source of clinical context. We intentionally made the scribe widely accessible because we viewed it as the foundation for something much larger, and we still have dozens of templates clinicians can generate notes from for free.

The consultation captures much more than what is ultimately written into the note. It contains clinical reasoning, patient preferences, what the clinician is thinking about, and the decisions that can power a much broader set of workflows. The clinical note is important, but it is still a narrower dataset than the visit itself.

Today, Heidi is evolving into what we call an AI Care Partner. Documentation is still central, but we now support clinicians before, during, and after the visit through products like Evidence, which brings trusted medical literature into the consultation. We’re also expanding into coding, pre-charting, knowledge base workflows, and other downstream tasks that happen around the visit.

The scribe was the wedge, but the broader opportunity is to understand the clinical encounter and support the work that happens around it.

Where do you see Heidi heading over the next five years?
In the next couple of years, there is still a lot of room for AI scribing to expand globally. In many countries, scribing has not really happened yet, often because of regulation, language, legacy systems, or product quality.

In the US, we see ourselves becoming the leading AI Care Partner for mid-market and specialty organizations. That includes groups like ChenMed, The Urology Group, and other specialty practices where the quality of documentation and the ability to configure agents around specific workflows really matter.

Specialty care is a good example. If patients are paying thousands of dollars to see a specialist, the note can’t feel generic. It needs to reflect how that specialist thinks and practices. That’s where being excellent at transcription, templating, and note generation matters.

Longer term, we want Heidi to be the best AI system that works across many systems underneath. If all the information needed to complete a task lives entirely inside the medical record, that is probably a feature of the medical record. Heidi’s opportunity is in the work that crosses systems, websites, records, clinical context, and external knowledge.

That’s where we want to win: safe, high-quality AI that can work across clinical environments.

Ambient AI has become a crowded market. What differentiates Heidi?
The ambient documentation market has become crowded quickly, but we don’t think the long-term differentiator is just generating a good note.

For us, the question is: what is the irreducible value five years from now? My view is that the durable opportunity is real clinical practice. As AI gets closer to influencing investigation, diagnosis, management, triage, specialist advice, or repeat scripts, the expectations change. You can’t treat that like a lightweight software feature.

That’s why we approach Heidi more like a medical device company than a traditional software company. Different countries draw the regulatory line differently. In the UK, even a medical scribe is treated with that level of seriousness. But the broader point is the same: if AI is going to influence care, companies need to prove very high levels of quality and safety.

There is real risk in these tools. A system can infer a diagnosis a clinician never actually considered, insert it into a note, and subtly change what happens next. Evidence tools can also hallucinate or incorrectly source information. As AI moves closer to care delivery, the bar gets higher.

That is where we want to sit. Our focus is clinical quality, safety, and breadth. We support clinicians in more than 190 countries and over 110 languages, which forces us to build flexible systems that work across healthcare settings rather than optimizing for one market or one EHR.

We are not trying to be the most deeply embedded EHR feature or a white-labeled tool inside someone else’s record. Our unique value is owning the models, safety, regulatory work, clinical quality, scale, and clinician experience needed to become a true AI Care Partner.

There’s been a lot of debate around specialist healthcare models versus general-purpose models. How do you think about that tradeoff?
General models are very impressive, but we need to be careful about how we evaluate them in healthcare.

A lot of benchmarks use public question sets, which means the models may have seen similar data before. That doesn’t tell you how well they perform when the answer depends on proprietary clinical context, new information, or a specific clinical workflow.

The bigger issue is objective alignment. Consumer AI models are trained to produce answers humans like. They are optimized through reinforcement learning from human feedback, and that often makes them agreeable, conversational, and good at keeping the interaction going.

In medicine, that can be dangerous. Sometimes the AI needs to give the clinician an answer they don’t want to hear. If a clinician frames a case in a way that confirms their existing bias, a general model may be more likely to agree with the framing instead of pushing back on the clinical facts.

But the job in medicine is not to keep the chat going. The job is to give the correct clinical answer, even if it is inconvenient.

That is why healthcare AI needs its own quality and safety bar. You need models and evaluations that optimize for clinical correctness, not user satisfaction alone.

What does Heidi’s AI stack look like under the hood?
We build on top of leading foundation models, but we also build our own models where we think we can outperform general-purpose systems on specific clinical tasks.

Our stack includes open-source models as starting points, internally fine-tuned models, proprietary reward models, and orchestration layers that route to the right model for the job. We are constantly balancing quality, latency, and cost rather than assuming one model solves everything.

For example, after six weeks of fine-tuning, we built a model that performed as well as Sonnet 4.6 on our Evidence task. The key is having a clear reward function. We look at clinician preference, but also train and benchmark around safety and quality because we need to meet medical device expectations.

There are a few reasons to build models ourselves: consistent behavior, product quality, speed, and cost, especially at the scale of millions of visits per week.

More broadly, I think application companies that generate large amounts of high-quality proprietary clinical data will increasingly build more of their own intelligence. In our case, we see more visits in our context than any general model does. If the goal is to write the best clinical notes in the world, it makes sense for Heidi to become better at that task than a general model with many competing priorities.

How do you approach quality, safety, and evaluation?
We think about quality and safety before, during, and after deployment.

Before anything reaches real users, we run internal evaluations. For Evidence, we use public datasets like HealthBench as one input, but we also use our own internal evals, including synthetic visits. We evaluate transcription, word error rate, output quality, and whether the model behaves safely across different clinical scenarios.

We also try to be more precise about what people call hallucinations. That word gets used for everything from a simple clerical error to a dangerous fabrication. Those are not the same risk. A grammatical issue is low risk. A strange, obviously wrong hallucination can be damaging to trust, but clinicians usually notice it. A subtle mishearing of a medication or medical term can be more dangerous because it is easier to miss during review.

So the goal is to measure what actually matters clinically, not just what is easy to count.

After deployment, we have a product called Verify that checks transcripts and notes for meaningful deviations from the facts and flags possible issues to users. We also do post-market surveillance, document safety incidents, and maintain a large internal medical team. Because we operate in environments where these tools are treated as medical devices, this is not optional. It is part of how we build.

What does implementation look like, and how do customers adopt Heidi?
One advantage of documentation is that clinicians can begin using Heidi almost immediately without requiring large implementation projects.

Many organizations start with individual clinicians or small teams before expanding across departments and health systems. Once clinicians begin experiencing the benefits directly, adoption often spreads organically through peer recommendation rather than top-down mandates.

For larger systems, we configure Heidi around the organization’s security, compliance, and EHR requirements. The goal is to fit into the existing environment rather than forcing clinicians to change how they practice.

Adoption patterns differ by segment. Solo practitioners and smaller groups can move very quickly because they control their own software stack. Health systems move more slowly because of committees, security reviews, and procurement, but adoption often accelerates when clinicians inside the organization are already using the tool and asking leadership to support it.

That is our core belief on go-to-market: clinicians are the ultimate arbiters of healthcare technology. If you solve a painful daily workflow, they will adopt it, rely on it, and push to keep it.

What kind of impact are customers seeing?
One example is Beth Israel Lahey Health, where we recently announced a system-wide rollout covering more than 6,000 providers across 14 hospitals and 175 primary care practices.

In the six-month pilot with 1,000 providers, 90% of clinicians said they felt more present during patient visits and could maintain eye contact because they were no longer taking manual notes. We also saw 89% user satisfaction with the quality of generated documentation, 88% physician-confirmed accuracy in capturing complex medical terminology, 82% of clinicians reporting reduced cognitive load, and 74% reporting less after-hours charting time.

That is the outcome we care about. It is not just generating a note quickly. It is giving clinicians time, attention, and mental energy back.

At a system level, that matters because reducing after-hours work and cognitive load directly addresses one of the biggest sources of clinician burnout.

You’re building from Australia but clearly have global ambitions. Has that shaped the product?
Australia turned out to be a useful place to build from.

Many Australian doctors operate more independently than physicians inside large US health systems. If they see a tool that helps them, they can try it quickly. That forced us to build something clinicians would choose on product quality rather than something sold purely through enterprise relationships.

It also forced us to build globally from the start. Heidi is now used across more than 190 countries and supports more than 110 languages. That means we had to build a flexible platform that can handle different clinical templates, workflows, languages, and regulatory environments without custom-building everything from scratch.

We also built around high regulatory baselines rather than local compliance patchwork. Our platform supports domestic data residency, enterprise-grade security requirements, and international compliance frameworks across markets.

The result is a product that can adapt across settings. In the US, the focus is often on EHR integration and complex clinical workflows. In the UK, Australia, and New Zealand, it can be more about reliability, low latency, and safety inside public-sector or legacy environments.

How does medicine change as healthcare becomes more agent-driven?
The big shift is from passive tools to active systems.

Today, a lot of software in healthcare behaves like a filing cabinet. Clinicians enter data, click through screens, and manually coordinate the next steps. In an agent-driven system, the platform can understand what happened in the patient interaction and start moving the work forward.

After a consultation, an AI Care Partner should be able to analyze the encounter, cross-reference trusted clinical literature, draft patient instructions in the right language, coordinate with pharmacies, help with booking logistics, and flag relevant changes for the care team.

The clinician still stays at the center as the final decision-maker and verifier. That is important. But more of the operational execution can move into the background.

Our ultimate vision is to double the world’s healthcare capacity without dehumanizing care delivery. For clinicians, that means shifting their work away from data entry and back toward presence, judgment, and direct patient care. For patients, it means more eye contact, clearer communication, and more of the doctor’s actual attention.

That is what we mean by AI Care Partner. It is not just a scribe or a chatbot that sits outside the workflow. It is a system that understands the clinical encounter, supports the work around it, and helps care teams move patients forward without adding another layer of friction.

What stood out, what’s tricky, and why it matters…

Heidi sits in one of the noisiest corners of healthcare AI: ambient documentation. But the more we dug into the company it was obvious that the scribe was just the beginning of the story.

The bigger bet is that the clinical encounter is the context layer. The note captures part of what happened, but the visit itself captures much more: the patient’s story, the clinician’s reasoning, the questions asked, the diagnoses considered, and the follow-up work that needs to happen next. If Heidi can understand that layer well, documentation becomes only one output from a much larger system.

That is what makes the company interesting. Heidi is trying to build a flexible AI care partner that works across systems, not just inside one EHR or one narrow workflow. Tom’s view is that if all the relevant information for a task lives inside the EHR, that probably becomes an EHR feature. Heidi wants the work that cuts across the EHR, websites, external evidence, patient communication, specialty workflows, and downstream admin.

Heidi is growing quickly with its approach, reportedly going from around $1M to $50M in ARR in roughly two years, has raised ~$100M, supports more than 2.5M consults each week, and operates across 110 languages and 190 countries. That combination of growth, usage, and global breadth matters because this is not just a distribution story. It is a data story. The more visits Heidi sees across specialties, countries, workflows, and templates, the more useful its models and workflow intelligence can become.

Tom also framed the company in a way that felt more clinically serious than a typical AI productivity tool. Heidi is not just trying to write a cleaner note. It is betting that as AI gets closer to clinical practice, the long-term moat becomes clinical quality, safety, regulatory readiness, model ownership, and clinician trust. That is a harder path than building a narrow scribe, but it is also the path that could make Heidi much more durable.

What stood out

What’s tricky

Final thoughts

Heidi is clearly trying to turn ambient AI into a real care partner layer.

The near-term value is straightforward: better notes, less after-hours charting, lower cognitive load, and more clinician presence during visits. The Beth Israel Lahey Health pilot numbers speak to that. But the more important question is what happens once Heidi understands millions of clinical encounters and can start helping with the work around the visit.

That is the most interesting part. Heidi is not betting that the future of healthcare AI is one chatbot replacing the doctor. It is betting that clinicians need a flexible AI layer that sits across the workflow, understands the encounter, retrieves the right evidence, drafts the right communication, supports the right follow-up, and quietly moves care forward while the clinician stays in control.

If that works, the company gets more valuable as it scales. More visits create more context. More context improves models and workflows. Better models and workflows drive adoption. More adoption creates more scale. That is the flywheel Heidi is trying to build.

Ambient documentation may be chapter one, but the real question is whether Heidi can become the global AI care partner sitting across the messy, fragmented systems where care actually happens. If they can pull that off, the company could become one of the more important examples of what healthcare AI is actually supposed to do: expand capacity while making care feel more human, not less.