When Ada Health was founded nine years ago, hardly anyone was talking about combining artificial intelligence and physician care — outside of a handful of futurists.
But the chatbot boom gave way to a powerful combination of AI-augmented health care which others, like Babylon Health in 2013 and KRY in 2015, also capitalized on. The journey Ada was about to take was not an obvious one, so I spoke to Dr. Claire Novorol, Ada’s co-founder and chief medical officer, at the Slush conference last year to unpack their process and strategy.
Co-founded with Daniel Nathrath and Dr. Martin Hirsch, the startup initially set out to be an assistant to doctors rather than something that would have a consumer interface. At the beginning, Novorol said they did not talk about what they were building as an AI so much as it was pure machine learning.
Years later, Ada is a free app, and just like the average chatbot, it asks a series of questions and employs an algorithm to make an initial health assessment. It then proposes next steps, such as making an appointment with a doctor or going to an emergency room. But Ada’s business model is not to supplant doctors but to create partnerships with healthcare providers and encourage patients to use it as an early screening system.
It was Novorol who convinced the company to pivot from creating tools for doctors into a patient-facing app that could save physicians time by providing patients with an initial diagnosis. Since the app launched in 2016, Ada has gone on to raise $69.3 million. In contrast, Babylon Health has raised $635.3 million, while KRY has raised $243.6 million. Ada claims to be the top medical app in 130 countries to date and has completed more than 15 million assessments to date.
Startups like Ada are pushing at an open door. Health professionals are warming to apps that have built-in collaboration with trained clinicians because the alternative is allowing patients to rely on a great deal of non-scientific health information found online that often verges into outright disinformation.
“The number of possible combinations of symptoms that could be entered into Ada is more than atoms in the universe,” says Novorol. “So building something that works instantly, that you know is fast and smooth and works at that scale, with that number of possible combinations… it hadn’t been done before.”
The question is: how different is it to create an AI-driven startup versus a standard enterprise software company?
“I don’t have experience of an enterprise company,” says Novorol, “but I suspect it is different from the experience of building other types of tech companies in that we spent many years on it [prior to launch], almost like a research team would do. The team was very focused on building out the core capabilities of the technology.
So for the first few years, we were a team of doctors, scientists, engineers, and mathematicians building that core medical reasoning capability: the medical knowledge base, the questioning algorithms, honing that constantly, testing it internally and then externally with doctors; stress-testing with patients and so forth before we could launch anything,” she explains.
Novorol notes that Ada undertook several pilots and multiple iterations of its engine as it strove to solve tough, technical and mathematical problems that weren’t addressed in medical literature.
“At that point in time, in those early years, the only business person was Nathrath,” she says. “Everybody else was working on the technology platform. There were no customers, apart from doctors who were using and testing and helping us improve the product, but through the pilots and co-development rather than through a commercial relationship.”
Once the consumer app launched, complete with patient-friendly language, it was then that the startup needed marketing and business development. Furthermore, Ada is operating in a space now where regulators are starting to catch up.
“There’s a lot more public discourse about digital health in general,” says Novorol. “All health care organizations are now looking to add digital, whether that’s through a partnership or trying to develop it themselves. So there’s a level of responsibility to service now that we have partnerships.”
She says Ada is pushing the boundaries of medical devices (how the app is classified) because regulations require them to create internal compliance and clinical evaluation teams and bring on a Patient Safety Officer. Ada’s experience attests the to the fact that AI-driven healthcare will come under a great deal more scrutiny in years to come. Arguably, more scrutiny than AI-driven startups in other sectors receive.
But the AI aspect of this journey does make an AI startup a different beast, says Novorol. “It’s not like you just build a product to see if it works. We had to spend years building the tech before we could actually launch.”
Does she feel that the weight of expectation by the press and consumers was something to deal with at the time? As AI-driven healthcare has risen up the agenda, it has met with controversy.
“We’ve been talking about this sort of product for many years but it’s only now coming to fruition,” she says. “In a way you are quite clearly pioneering this sort of thing. So comes down to the difference between the hype and all the predictions made and the reality of actually launching an AI onto an unsuspecting public.”
Of course, launching an AI startup is often not that at all.
“When we started this eight years ago there was no talk about AI,” says Novorol. “People were starting to talk about big data, but there was no talk about AI and we never called it AI. It’s a form of AI but it’s not the sort of deep learning which everyone’s been excited about recently.”
She says the so-called AI systems that we know now won’t be the technology that solves everything. “We need a more nuanced approach. It’s more of a probabilistic system. It’s medical reasoning. So we always talked about medical reasoning or medical knowledge — we never really even talked about AI, because we were thinking from the medical perspective such as how doctors think.”
“I remember in the very early days having a conversation with a very smart person at Cambridge University and they said, ‘Oh, but people have tried this before and it didn’t work.’ So there was a lot of skepticism, understandably.”
Another key to building this startup was not hyping it in the press.
“We didn’t talk to the press for years,” says Novorol. “Nobody knew whether or not it was going to work. We were totally focused on building technology. I think it gave us a real head start in this field, allowing us to basically focus on quality and doing it right.
That meant that by the time we launched, we’d already achieved a lot behind the scenes. So we weren’t officially in stealth but effectively, to some extent, we were. And so we launched knowing that it already worked and knowing that it performed better than anything else that was out there, that was capable of supporting specialists in their particular field. It’s not the same as starting with something very basic that can cover the common cold.”
Ada clearly put a great deal of effort into shooting for the most complex scenarios as well as the simplest because their app had to perform at the highest level.
“We started with building something that would be good enough to support highly trained professionals in their field and then we expanded that out for consumers. And we got it right before we launched,” recalls Novorol. “We continue to expand the knowledge and improve the performance. But it was already far better than anything else has ever been in this space before. So that made a big difference.”
Novorol says she believes the Minimum Viable Product approach used in many tech sectors simply won’t be sufficient for the AI-driven future.
“I guess the advice would still be to make sure that you do the groundwork first. You know that you don’t go out and launch something promising more than you can actually deliver and overstating the capabilities of your technology.”
“In healthcare if you launch a symptom assessment app and you say it can do X Y Z and it doesn’t, that wouldn’t be a good approach, to put it mildly. You’ve got to do the work basically, and it takes time.”
In the final analysis, Novorol says she’s aware that there are some things Ada could have done more quickly. “With the benefit of hindsight, for certain pivots or iterations, some things could have been stopped sooner and others doubled down on, But I don’t think I would change anything about the ultimate journey and the direction, although in an ideal world you’d learn things sooner and move faster.”