Benefits sit on top of trillions of dollars, thousands of carriers, and millions of choices. This is the ground every model we build has to stand on. For every person, in every language, every year.
Strip each plan down to what actually matters for this person, so no one pays for coverage they will never use.
Recommend the plan that fits a real life, and end the 24% premium most employees overspend by defaulting to the wrong one.
Catch the overpayment, the unused HSA, the better option, every day. Not just in the two weeks of open enrollment.
Every carrier, plan, rider, employer, broker, and member is a node. Every commission, rule, and relationship is an edge. We build a living ontology of the whole network, resolve it to one source of truth, and reconcile it in real time as the data drifts.
The same plan shows up under five names across five systems. We resolve every record to one canonical entity.
Every carrier speaks a different language. We normalize each feed into typed nodes and labeled edges.
Benefits nest: carrier, plan, tier, benefit, rider, rule. We model the whole hierarchy, so answers come from the exact level they live on.
Every node is paid differently, and misalignment is where money leaks. We encode the incentive structure into the graph itself.
Fathom is our benefits-native model. It reads any carrier's plan documents, reasons over the living graph, and answers the questions people actually ask. Is this covered? What will it cost me? Where do I go? Every answer points back to the exact line in the document, so people can trust it.
It is safe with private health data, and it always shows its work. Fathom is live inside Keel today, helping real employees enroll.
Benefits AI lives inside one of the most regulated environments in the country. HIPAA, ERISA, the ACA, and a fast-moving wave of state AI laws all govern how a model may touch a person's coverage. We do not bolt compliance on at the end. We build the agent so that every action it takes passes through governance first.
We run on BAA-covered, zero-retention infrastructure. A model that keeps prompts is a breach waiting to happen.
Fiduciary duty means a black box cannot hold final authority. We keep the human in charge.
The 2024 rule covers decision-support algorithms directly. We test inputs and outputs for disparate impact, always.
Coverage decisions require individualized human review. Our agent routes to a person. It never denies care on its own.
About 25 states expect documented AI governance with vendor oversight. We ship the controls and the exam-ready records.
Govern, map, measure, manage, plus the security and healthcare certifications expected before anyone integrates.
It all runs on one spine: the living graph of benefits. The lab turns that graph into science. The model turns it into answers. Amanda turns it into a conversation a real person can have at 11pm on their phone.
We map the whole network into one living graph, publish the science behind it, and build the models no general-purpose system is allowed to.
Read the research →Our benefits-native model. It reads any plan, reasons over the live graph, and proves every answer back to the exact line it came from.
Meet Fathom →The voice people actually talk to. Fathom, given a face and a warm bedside manner, live inside Keel and answering real questions today.
Talk to Amanda →In most of the field, a wrong answer is a bad demo. Here it is someone's healthcare, money, and family. These are the problems we work on, and would love your help with.
Finding the right answer inside messy, contradictory plan documents, and proving it is right. A wrong answer here is not a bad demo. It is a real problem for a real person.
There is no benchmark for whether benefits advice is correct, so we are building one. You cannot ship "mostly right" when it is someone's coverage.
Understanding someone's real life without ever exposing their private data. Privacy is not a feature we add later. It is a constraint we design around from day one.
Software that enrolls, files, and follows up for a person, with a full record of everything it did and why it did it.
Today's note: Arbitration was built for 17,000 disputes. It got 1.2 million.
A small, remote team building the AI that reads benefits. Three roles open, and a screen worth taking even if you do not apply.