Healthcare Innovation has written features about health system implementations of ambient clinical intelligence from vendors such as Abridge, Ambience, Microsoft, Suki, and Nabla. One startup that has flown under the radar a bit is Knowtex, which recently became one of two prime awardees on the Veteran Administration’s $776 million ambient scribe enterprise contract.

The company’s CEO, Caroline Zhang, recently sat down for an interview about the VA contract and how Knowtex differentiates itself in the market. 

Healthcare Innovation: Could we start by talking about the origin story of the company? I read that you and co-founder Jocelyn Kang spent six months embedded in clinical teams working as medical scribes to learn the business. 

Zhang: That’s correct. We spent six months each in scrubs. I spent time in oncology and my co-founder in orthopedics. We're not originally from a medical background. We were AI scientists from Stanford. So it was a very intentional decision for us to really walk a mile in our users' shoes and be shoulder to shoulder with the doctors, and also see what the end-to-end workflow is like inside the clinic. That involves everything from knowing what makes a note great, why we need to build specialty-specific AI models of our own, and build end-to-end workflows to see the coding and the order placement for chemotherapy. That informed a lot of the initial platform that we built. 

HCI: Could you talk about your work with the VA?

Zhang: Healthcare is always a blend of the provider experience and regulation. I’ve worked in the U.S. Senate on AI policy, and saw that healthcare is going to have this massive transformation with AI technology. The VA is really an opportunity to lead the way here. They're the largest U.S. healthcare system. So when the VA said it was going to issue a 2024 AI tech sprint, we knew that was the perfect opportunity to demonstrate a product that we'd really painstakingly built out over the last two years. In a nutshell, that product experience led to us being a winner of the tech sprint. We were competing against a field of 150 competitors. The VA is very rigorous in their evaluation. They looked at product capabilities and integration. 

In July 2025 we were awarded a $15 million pilot award. Knowtex and Abridge each received a pilot award, and then we were both able to go deploy in separate regions of the country. We've had extraordinary success with our VA product adoption. We’re already in 310 clinics in the VA and 79 VA hospitals. In six months, we brought all that live, so it was a historic speed of roll-out.

HCI: With the VA, is this mostly in the ambulatory setting or is it also for hospitalists and nursing workflow in the acute care setting?

Zhang: The product we've built commercially does service 200-plus specialties today, but in the VA it's a controlled roll-out. They started in primary care, and in the pilot phase they did 13 specialties. But in the enterprise contract the scope covers all VA clinicians. So ED is an area that's on the roadmap for the VA that we've supported commercially before. I will say there have been demos from our team about nursing workflows, and we know that's also within the VA’s realm of interest.

HCI: I saw a case study on your website about work with the University of Rochester Medical Center. Were they an early customer, and were there things you you learned working with them?

Zhang: What we really liked was that they're an academic medical center with an Epic system. When we integrated, we did an initial pilot and the results were really strong. I credit that to working with a good chief medical informatics officer. With the initial cohort we had a focus on specialties as well as primary care workflows. The team in orthopedic surgery really felt the difference between our tool and more of a generic, standard AI scribe tool. What we have is not a wrapper; it's actually a specialty-specific model for each clinical specialty workflow, and that's the difference between getting 70% of the way there versus 99%. For us, 100% of the way is the goal, and we've actually released a benchmark tool to measure how much clinician effort we can reduce.

HCI: I was going to ask you about that. I saw a paper about your creation of KnowBench to measure effort reduction.

Zhang: KnowBench is about defining a benchmark criteria around real clinical adoption and impact. Looking at our deployment experience nationally, it’s aggregated and it's really measuring across notes, codes, orders, and decision recommendations. What is it that a clinician will actually accept and put their signature on, as it goes into the record and downstream? We're seeing today we hit 97.99%. It is a very high bar that we're looking to continue to advance.

HCI: But is one of the ideas that a CMIO would use this tool to assess not just your product, but a bunch of different clinical AI tools to measure the impact they're having?

Zhang: Yes. These are evaluation standards that work in terms of deployment and actual realized return when you invest in a clinical AI solution. There's been so much healthcare AI that's come out that it is hard to tell what's marketing and what works in production. We can have chief AI officers and chief medical informatics officers be part of this conversation and sharing what they're seeing, too. We’d like to hold general-purpose models to this standard, too. Because so much of what comes out in an OpenAI or Anthropic model is quite general, and they're not thinking about healthcare-specific needs in a way that we have been for the last four years. We have been very deeply embedded with our enterprise healthcare partners. It's about bringing awareness —  these are the guardrails and the expectations for healthcare adoption of AI for the long-term.