Our latest episode was a special one. We sat down with Zach, the CTO of Xsolis, a healthcare company based in Nashville, Tennessee, that has spent more than ten years using machine learning and predictive AI to reduce the friction between hospitals and health insurance companies in the US. Here is what we learned.
The problem: hospitals and insurers don’t get along
In the US, healthcare providers treat patients and insurance companies (the “payers”) foot the bill. When the two sides can’t agree on the level of care a patient needs, the payer issues a denial and refuses to pay the claim. The hospital has already delivered the care and can’t recoup the cost, and at the end of the chain a patient can end up personally liable for a bill they thought was covered. Xsolis was founded on a simple hypothesis: if both sides work from the same data, much of that friction doesn’t have to exist.
The Care Level Score
At the heart of Xsolis is a proprietary predictive algorithm called the Care Level Score, or CLS. It quantifies how sick a patient is and therefore the level of care they need, and it updates in near real time as the patient’s condition changes during their stay. Xsolis gives the exact same score and data to both the provider and the payer, so everyone works from the same information. Payers use it to authorise care while the patient is still in hospital, and the highest score reached during a stay (the “max CLS”) is the marker that should determine fair reimbursement. Xsolis positions itself as a neutral third party: not favouring providers or payers, just letting the data speak.
The data challenge
Unlike the UK, where the NHS keeps electronic medical records nationally, the US has no central repository of healthcare data. Every provider has its own data store, in its own format. A common framework called HL7 (an XML-based standard) helps, but Xsolis still has to normalise and standardise every new client’s data, and it is now using agentic AI tooling to automate some of that mapping. Their platform, Dragonfly, processes almost 10 million HL7 messages a day, and their models are trained on clinical data from over 200 million patient encounters across the US. All predictive models go through quarterly retraining events where they are analysed for bias, hallucinations and model drift. On security, the basics (encryption, role-based access) are treated as building blocks, and the focus is shifting to proactive security: using AI to predict attack vectors before any alarm goes off.
Humans stay in the loop
This was one of our favourite parts of the conversation. Xsolis does not create care pathways or treatment plans. It makes recommendations, and a trained clinician always makes the final call. Two examples stuck with us. First, a model that predicts early in a hospital stay where a patient will need to be discharged to: home, a rehabilitation centre, or a skilled nursing facility. Second, length-of-stay management: if the system predicts a patient will likely be ready to go home on day three, the care team works towards that goal from day one, which gets the patient home sooner (safer and more comfortable) and frees up a scarce hospital bed. And if the patient isn’t ready on day three? The clinician overrides the prediction. Rinat shared the famous study about judges giving harsher sentences just before lunch: humans have weaknesses too. But in medicine, the final decision belongs to a person.
Does it work?
Zach shared a before-and-after study with Beacon Health, a US healthcare system and a client of almost six years. Before Xsolis, clinicians wrote every “initial review” (the first case summary sent to the insurance company) manually, combing through the medical record. After implementing Dragonfly, the same number of staff reviewed over 140% more patients. The next step is already rolling out: generative AI now drafts the initial review at the click of a button, and the clinician edits and approves it. The human is still in the loop, just with far less repetitive work.
The ethics are the hard part
We pushed Zach on prevention: with all that data, could insurers incentivise patients to stay healthy, the way Vitality does in the UK? Xsolis doesn’t do anything patient-facing today, and Zach was clear-eyed about why: there are serious ethical and legal considerations, from GDPR in Europe to California’s privacy law, and any positive-reinforcement scheme should be based on people opting in. He did point to one promising use of data, identifying “gaps in care”: comparing a patient’s journey against an evidence-based treatment protocol and flagging missed steps. Drawing the line between a helpful recommendation and a programmatic requirement is work for medical ethicists, and it is a growing field of research.
What we took away
- AI works best in high-stakes fields as a recommendation engine, with a knowledgeable human making the final decision.
- Data symmetry can turn an adversarial relationship into a cooperative one. Zach’s proudest moment: at their last two user conferences, a payer and a provider presented together about a better way to work. That almost never happens in US healthcare.
- The ethical questions, bias, privacy, consent, are just as important as the technical ones, and the industry is still writing the rules.
Thank you to Zach for a fascinating conversation. Listen to the full episode on our Captivate page or watch it on YouTube, and let us know what you think.
0 Comments