Subscribe | Advertise | PRO Member Hub

TOGETHER WITH ARTERA

Perspectives from the people building the future of health AI…

We sat down with Guillaume de Zwirek, Co-Founder and CEO of Artera, to discuss how the company is continuing to evolve and support healthcare providers as an agentic AI partner.

Guillaume started his career at Google before his own experience navigating serious health problems as a competitive Ironman athlete exposed how difficult it was to access care. That experience inspired Artera's founding mission: make healthcare the best industry in the world for customer service.

Artera began with HIPAA-compliant text messaging and expanded into automated, deterministic workflows. Today, its Harmony platform and AI agents support communication across voice, text, and web, and manage scheduling, referrals, eligibility, billing, and other administrative tasks. The aim is to resolve routine requests automatically while connecting patients with staff when human help is needed.

Artera serves 1,000+ healthcare organizations, including specialty practices, health systems, FQHCs, and federal agencies. The company says its platform supports more than 2B patient-provider communications each year and has engaged 200M+ unique patients. In December 2025, it announced a $65M growth investment and said it had reached $100M in contracted annual recurring revenue (CARR) by year-end.

More recently, Artera introduced an AI Services model in which specialized teams of builders work directly with healthcare practices to deploy agents around their workflows. Guillaume discussed that operating shift, his views on the next generation of healthcare agents, why clinical and operational context is hard to replicate, and how the company tests reliability at scale.

Let's start from the top. What's your background, and what inspired you to build Artera?
I started my career at Google, working in commerce and search. Eventually, I realized I wanted to build something myself. My boss had also insisted I get a hobby. When I asked what his was, he said Ironman triathlons, so I decided that would be mine too. I took it pretty seriously and started competing with the goal of earning my professional card.

During races, I started losing my vision. In one race I lost it completely, and an EMT warned me that I might be at risk of sudden cardiac arrest. I ended up traveling around the country seeing cardiologists and specialists, trying to figure out what was happening.

My mother is a physician, so I'd grown up with a fairly privileged experience of healthcare. Suddenly, in my early 20s, I was managing my own care and couldn't believe how difficult it was, especially when I was already scared and vulnerable.

I remember thinking about walking into Tiffany's to spend $10K on a diamond ring. Someone greets you, gives you champagne, and makes sure the experience is wonderful. In healthcare, you can spend far more and have a terrible experience.

That became the founding idea behind Artera: how do we make healthcare the best industry in the world for customer service? I was naive about how difficult the industry would be. More than a decade later, the mission hasn't changed, but the technology has finally reached a point where I think we can deliver on it.

For readers hearing about Artera for the first time, what does the company do today?
We automate the routine, undifferentiated work between patients and healthcare staff that gets in the way of people receiving care. That includes phone calls, texts, scheduling, insurance eligibility, billing, referrals, and other tasks that happen outside the exam room.

Today, our agents can talk to patients, access information, and take action in EHRs and other systems, 24/7. Patients are able to get help through their preferred channel, in their language, at a time that works for them. If a request is complex, acute, or outside policy, we recognize that and connect them with a person immediately.

We started with HIPAA-compliant texting between patients and staff. Then came automated messages, deterministic workflows, and now AI agents capable of navigating conversations that don't follow a script.

The goal has always been the same. Take away the routine administrative work so patients get what they need and staff have more time for the interactions that genuinely require a human.

Despite years of investment in digital health, patient access remains difficult. What's fundamentally broken?
I think it's a business model problem. Providers are generally reimbursed by payers, and many operate in markets where patients have limited alternatives. That can make patient experience feel like a cost to control rather than something worth investing in.

I'll give you a personal example. I haven't seen a primary care physician in two years. My mother is a doctor, I run a healthcare company, and I have a history of heart-related issues. But I don't want to call between 8 and 5, wait on hold, or remember my MyChart password. That friction has been enough to put it off.

Recently, I asked a personal AI agent to book an appointment for me. It called a large health system and another AI agent answered. My agent asked to book the visit, but the health system's agent offered to text a link. Mine couldn't receive texts and asked to complete the booking directly.

Eventually, it was transferred to a human representative, who said corporate policy prohibited them from speaking to an AI. Can you believe that? My agent had my calendar and insurance information. I'd outsourced 25 minutes of work and still couldn't get an appointment.

The alternative for a patient often isn't another provider. It's doing nothing. That's a missed opportunity to catch health problems earlier, and the provider also loses the chance to deliver that care. We need to make access unbelievably easy for health systems to adopt and show them that reducing friction can bring in patients who otherwise wouldn't be seen. That's a revenue opportunity as much as a service improvement.

Artera has moved from texting and deterministic workflows to autonomous agents. What changed, and where do copilots and flows fit now?
For a long time, the technology wasn't reliable enough. The early GPT models hallucinated too much for healthcare workflows. You can't have an agent mistakenly refill a controlled substance. We needed better reasoning, stronger guardrails, and a secure way to access the systems where work happens.

The foundation models improved quickly. We used to maintain natural language understanding libraries for all the different ways patients express an intent. Now a model understands that "I need a refill" and "I'm taking a red-eye tonight and running out of blood pressure medication" are related requests. Interoperability improvements, FHIR, and newer approaches like MCP have also helped make actions inside EHRs more feasible.

Frankly, I think a lot of our older technology should eventually die. Hundreds of customers still use deterministic flows, but rigid decision trees are a relic of the past. We also built copilots that suggested messages for staff to approve. If an agent can safely resolve the request itself, why have a person approve the text?

Scheduling, standard refill requests, billing questions, and insurance updates can increasingly be handled autonomously. Humans should handle situations that are acute, complicated, or outside policy.

A big part of navigating this transition is having amnesia about sunk costs. You might have spent millions building software that won't be relevant in two years. You need to help customers move to what's better rather than protect the old product.

You've said domain expertise and distribution matter more than proprietary models. What's Artera's advantage, and why the focus on specialty practices?
I'm skeptical of a vertical AI vendor claiming it can out-innovate the frontier labs on fundamental model technology. I believe vertical AI will produce successful businesses, but the important advantages are domain knowledge and distribution, with data playing a supporting role.

Healthcare has endless exceptions. How do you handle a controlled substance or a patient proxy? Those rules can change by state. One physician won't see patients under 60; another has unusual scheduling requirements. You need to understand the EHR, the practice, the specialty, and its policies. Customers don't want to teach their vendor all of that.

We have years of interaction data, and it's useful. But I think data is sometimes overplayed, particularly as synthetic data improves. The main administrative tasks are reasonably well understood. The difficult part is completing them reliably in thousands of different circumstances.

Distribution matters enormously. We have relationships with more than 1,000 healthcare organizations, and when we ask customers what they value, the word we hear most often is "partner." We have to keep earning that trust.

Our first customers were large health systems, but today we're spending more time with specialty clinics and physician groups, from 20-doctor practices to much larger organizations. Large EHR companies can bundle competing tools into health systems where they're already deeply embedded. Specialty groups give us an opportunity to bring more tailored expertise and move quickly. Strong relationships only matter if we keep solving problems well.

You've introduced AI Service Squads that put builders alongside providers. How do they work, and what does implementation actually involve?
It's our version of forward-deployed engineering. Healthcare is too variable to assume that one piece of software will work the same way everywhere.

Take ten customers using eClinicalWorks. Their FHIR and HL7 feeds may all be configured differently. Some use proprietary APIs or RPA, and the scheduling policies aren't the same. Mary at the front desk knows how Dr. Jane likes her schedule, but those rules may never have been written down.

Our builders work directly with the practice to collect that knowledge, build the tools and rules the agents need, and integrate with the EHR. The details matter: different physicians have different scheduling preferences, exception policies, and even technical configurations within the same system. We then test against a simulated version of that customer's actual environment before deployment.

The traditional SaaS process is too slow for this. A request goes from implementation to product management to an engineering queue, through several rounds of telephone, and two months later the result may not be what the customer wanted. We put builders close enough to the problem to solve it directly.

That doesn't mean every deployment is custom code. We use common infrastructure and a library of tools, and turn recurring solutions into reusable components. Each squad benefits from what we've learned serving more than 1,000 organizations.

I also don't think healthcare operators should be handed the keys to customize complex agents on their own yet. Adding lines to a prompt can make performance worse; sometimes you need a better tool or integration. Our teams understand those tradeoffs and the security standards involved.

What does the technology stack look like under the hood, particularly around models, voice, and security?
We're largely an AWS shop, running model workloads through Bedrock. That gives us model flexibility within an environment built around our security requirements. For voice, we still use cascading pipelines: speech-to-text, an LLM, and text-to-speech. I'm interested in direct speech-to-speech models, although their guardrails are still tricky to tune.

One of our more interesting investments is secure log analysis. We handle billions of patient communications annually, and a conversation can go anywhere. Someone could call about an appointment and disclose suicidal ideation. We need to understand how the agent performed without letting staff pull sensitive transcripts onto their computers.

We built an internal tool where a planner agent generates a query against conversation logs inside our AWS account boundary. The analysis runs there, and services including Amazon Comprehend remove protected health information before the report is returned to our team. We can investigate problems and improve the software without exposing the underlying conversations.

We've also done extensive work on the security requirements of serving federal customers, and we apply similarly rigorous controls in our commercial environments. I think healthcare buyers need to scrutinize this. A small mistake involving patient information can be catastrophic.

How do you test patient-facing agents and decide when a new workflow is ready for production?
Every agent starts from a common foundation with our security policies and operating requirements built in. We then configure the tools and rules for that particular customer.

We also build synthetic replicas of their EHR environment. If a clinic uses NextGen, for example, we'll create a simulated version reflecting its actual scheduling patterns, provider rules, API responses, and particular FHIR or HL7 configurations. It's their version of NextGen, not a generic test system. That can include an unusual JSON response, a customized interface, or a third-party hosted installation that behaves differently from the standard configuration.

Then we use de-identified information from real patient interactions to develop synthetic personas and realistic scenarios. The agent runs through those test suites, and we score how reliably it completes the work.

If a change introduces hallucinations, causes an agent to call the wrong tools, or otherwise degrades performance, we can block the release. Once it meets our thresholds, we introduce it to a small portion of live traffic, compare the results, and expand gradually.

We take this seriously because the agent has to complete the task according to the clinic's rules, protect patient information, and recognize when a human needs to take over. A convincing conversation isn't enough if the work doesn't get done correctly.

Can you share a concrete example where AI agents improved patient access or staff workload?
We worked with a pediatric organization operating 12 clinics in underserved communities. They needed to reach families whose children were overdue for well-child visits, which traditionally required a lot of manual outreach.

We deployed agents that texted families in multiple languages, held conversations, and booked appointments directly into the EHR. In the first four months, those agents booked more than 10,000 well-child visits for families they reached.

That freed staff from thousands of manual outreach calls, giving them more time for patients in the clinic. I like the example because it shows what the technology should accomplish: make it easier for people to access care while freeing staff to provide it.

At more than 2B annual patient communications, how do you measure whether agents are actually working?
We have two North Star metrics: controllable non-resolution rate, or CNRR, and patient experience quality, or PXQ.

CNRR measures the work our agents should be able to handle but fail to resolve. It tells us whether the problem is an unsupported request, failed verification, patient abandonment, or something else. LLMs review conversation transcripts, extract those patterns, and return findings to our team after protected health information is removed. Sometimes the system automatically creates tickets for fixes.

PXQ measures the quality of the interaction itself. An agent can technically finish a task while making a patient miserable. It might misrecognize speech, pursue the wrong goal, or make it unnecessarily difficult to reach a human. We want to catch those moments, including a transfer request that's being blocked for no good reason. Those failures create friction and undermine trust.

We're always balancing cost, quality, and speed, but we have minimum thresholds for CNRR and PXQ before an agent goes live. Security is non-negotiable, and certain findings simply block a release. Automating more interactions isn't useful if we're making the patient experience worse.

Three to five years from now, what does patient access look like if agentic AI works? Where do humans fit?
I think the agentic revolution is widely misunderstood. It's about removing barriers to care and freeing staff for conversations that matter, rather than eliminating human connection.

Given the labor shortage and increasing complexity of healthcare, my rough estimate is that we'll need to automate around 80% of today's undifferentiated administrative work to create enough capacity for the more complex, high-acuity interactions.

Patients should be able to get help when they want, through their preferred channel and language, without navigating phone trees or waiting on hold. And when they need a person, someone should answer immediately, know who they are, and have time to help.

That's what we set out to build at Artera. Back then, the tool was text messaging. Today, agents make it possible to pursue the same mission on a much larger scale.

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

Artera has spent more than a decade working on a fairly unglamorous but important healthcare problem: making it easier for patients to get the care they need. The company started with text messaging, expanded into automated patient communications, and is now building agents that can handle much of the administrative work behind those interactions.

Guillaume's frustration with patient access helps explain the opportunity. Despite years of investment in digital health, patients still deal with phone trees, forgotten portal passwords, long hold times, and complicated scheduling processes. His story about a personal AI agent being unable to book an appointment because a health system prohibited its staff from speaking with AI is a particularly good example. The technology existed on both sides of the conversation, yet the appointment still couldn't be booked.

Artera has a substantial foundation to build on. The company works with 1,000+ healthcare organizations, handles over 2B patient communications annually, and has engaged 200M+ unique patients. It also announced $100M in contracted annual recurring revenue and secured a $65M growth investment in 2025. That scale comes with established distribution, EHR integrations, and years of experience navigating healthcare's complicated administrative workflows.

The company's latest shift is operational as much as technological. Through AI Service Squads, Artera is putting specialized builders directly alongside providers to develop agents around their specific needs. Even two clinics running the same EHR can have completely different integrations, scheduling rules, and internal policies. Artera is trying to combine that customization with reusable infrastructure and lessons learned across its customer base.

The early results are encouraging. At a 12-clinic pediatric organization serving underserved communities, Artera's agents booked more than 10K well-child visits in four months, while freeing staff from thousands of manual outreach calls. It's a useful example of how automation can help patients get appointments and give healthcare staff more time for the people in front of them.

What is most interesting about Artera is that it's approaching the AI wave from almost the opposite direction of many newer startups. Most new healthcare AI companies start with the technology, then spend years building trust, establishing relationships, securing integrations, and learning how healthcare organizations actually operate. Artera has spent more than a decade doing that work. It already has the customers, distribution, integrations, and institutional knowledge that are often the hardest parts of building a successful healthcare technology business.

And Artera is very willing to adapt to this new era. Guillaume isn't just adding AI features to an existing platform. He's openly questioning products Artera spent years building, reorganizing teams around AI Service Squads, and leaning heavily into agents as the next evolution of the business. That's a pretty compelling combo: an established, trusted healthcare technology company operating with the ambition and willingness to change that one would typically associate with a much younger startup.

There are still challenges around scaling customized deployments and maintaining reliability as agents take on more responsibility. But Artera is already making that transition, bringing new AI capabilities into established customer relationships and workflows rather than starting from scratch. It's a compelling position in a market full of new entrants, and a good example of how established healthcare technology companies can help usher in the next era of AI-enabled care while continuing to serve the customers that got them here.