ZERO Health: From Redshift Latency to AI-Powered Healthcare Analytics

A new model for healthcare billing

ZERO started when a surgeon in Oklahoma City got tired of dealing with insurance companies. He posted his cash prices online and said, "I just want to send an invoice and get paid."

"We became a direct contracting company that could make employers working through us look just like they were coming to his place with a bundle of cash," explains Stan Schwartz, MD, ZERO Health Co-Founder. "We reduce all the friction, take all the sludge out of the system."

The result is a system where everyone wins: patients pay nothing out of pocket, employers save 30-40% on procedures, and providers get paid immediately without haggling over deductibles. ZERO's goal is to extend this model to everything in healthcare that can be scheduled.

"The unspoken truth is that the more we are unlike insurance, the more they love us."

The Billing Chaos Problem

Healthcare billing in the US is notoriously complex. There are 65,000 encoded procedures, each with multiple components. A single knee surgery might generate 50 different line items across multiple bills. ZERO's job is to bundle all of that into a single price that employers and patients can actually understand, passing along dramatic savings to employers and members.

"We really tried to codify that into an algorithm that identifies those bills and bundles them up into that single procedure," explains Greg Inman, CTO. "So what does a knee surgery cost? If it's 50,000 different permutations of whatever, we're going to bundle that up into a single line item that says knee surgery, $16,000."

The data problem compounds at every step. Prospects send massive, messy files—sometimes 15GB CSVs—in unpredictable formats. Engineers had to spend hours just getting a look at the data to see if it was usable. And their previous business intelligence platform, built on Amazon Redshift, couldn't keep up.

"We called it the blue bar of tyranny," says Inman. "It was taking a minute to load a list of 100 employers just to filter on. We hired an outside consultant who came highly recommended. They told us, 'It looks pretty good to me.' And I said, 'Why is it taking a minute to load a filter list?'"

ZERO's engineers had been experimenting with DuckDB and saw an opportunity. They rebuilt the pipeline and ran the same analysis.

Six minutes to load a messy, massive CSV. All local.

"It is incredible to think back—before this type of technology, how do you even look at a 15 gigabyte CSV file? We had this reality where we would have to involve engineers and spend a bunch of time and hack things together just to get a look at this data. That friction is gone."

Enter MotherDuck

The new architecture streams data directly from S3 into MotherDuck, where a dynamic parser normalizes the chaotic billing formats into a consistent structure.

• Ingestion: Raw CSV files upload to S3, then stream directly into MotherDuck
• Parsing: AI-generated parser definitions (stored in MotherDuck) normalize inconsistent billing formats
• Processing: CTEs bundle 50+ line items into single procedures; bundling algorithm runs opportunity analysis and claims adjudication
• Analytics: MotherDuck MCP server exposes sanitized data to Claude for self-service querying

"We upload to S3, point MotherDuck at it directly, and stream it in," Inman explains. "We developed our own parser format that AI can write to pretty easily. It writes the parser definition, which is stored in MotherDuck, and then MotherDuck uses that to normalize the data. Then we have CTEs that bundle 50 different line items into a single procedure."

The MCP Breakthrough

The real transformation came when ZERO built an internal AI toolkit using MotherDuck's MCP server. The goal: let subject matter experts—like co-founder Stan Schwartz, a retired infectious disease physician—query data directly without engineering involvement.

"The prospect of having true self-service BI was right there in front of you," says Inman. "LLMs are exceptionally skilled at writing SQL—better than I am in a lot of cases."

The team built internal tooling to streamline MCP configuration, manage tokens, and push sanitized data into a MotherDuck database. Then they opened it up to Stan.

"I've been doing this for five months now," Stan says. "And I can now, with fair confidence, go in and get questions answered. The remarkable thing is that I don't have to know anymore where things are. Claude can find things."

What They Built

The team has developed several Claude skills that combine MCP access to MotherDuck with domain expertise:

Drug Infusion Pricing A task that used to take 40 minutes of manual research now takes about 4 minutes. "There was a lady here in Oklahoma that needed an expensive drug infusion," Stan recalls. "It was literally a six-minute process—I could get coffee while Claude was thinking about it."

Contract Proposal Generation A Claude skill pulls prospect data from HubSpot via MCP, matches it against provider data in MotherDuck, and generates dynamic contract proposals with up-to-date pricing. "Claude goes and generates a prospective contract with dynamic pricing at given target intervals for savings—all just dynamically generated based upon our core data sets."

Automated Rule Generation Stan used to manually write YAML files defining how to extract components from each of the 65,000 medical procedure codes. It took 45 minutes per code. Now he has an interactive Claude artifact that generates the YAML in 2 minutes. The rules go directly into a MotherDuck table and are immediately available to the algorithm—no engineering review required.

"The beauty of it is that we can use the same system we build them with to turn around and test them," Stan notes.

What Changed

The performance gains were dramatic. Opportunity analysis that took six hours now runs in six minutes. Drug pricing research dropped from 40 minutes to 4. Rule generation went from 45 minutes per procedure code to 2 minutes.

But the deeper shift was cultural. Going all-in on MCP-powered analytics fundamentally changed how technical and non-technical team members collaborate.

Before, Stan and other subject matter experts would come to engineering with questions. They'd go back and forth for weeks. Now Stan explores the data directly with Claude, develops his understanding, and comes to engineering with much more refined requests.

"Stan is a very curious person and would come to us all the time with questions," Inman reflects. "A lot of times we would say, 'I don't know about that' or 'I think maybe this is off base.' Now Stan has those conversations with Claude and we're in much more of a support role than a gatekeeping role," says Greg.

"It's turned laborious and boring tasks into fun. I mean, that's as clear as I can state it." — Stan Schwartz, Co-Founder