Building a High Growth Vertical AI Company in Healthcare: Ganesh Padmanabhan, CEO of Autonomize AI (Part 5)

Sramana Mitra: However, the first health insurance company is the real turning point for you in terms of product-market fit. So tell me a bit about the technology infrastructure that you had to plug into. This is obviously not happening in vacuum, there are all these systems that you have to work on top of.

Ganesh Padmanabhan: Enterprises are messy. You cannot expect a greenfield and a clean slate anywhere. Our first customer that was a turning point is a Fortune 100 health insurance company. So, you can imagine the complexity there. Another lesson I learned was that it is easier to go down market than to go up market in a B2B context.

They had their own medical management system, they had other vendor tools like TriZetto from Cognizant, ZeOmega, medical management systems, and Salesforce infrastructure. I would say 70-80% of the existing infrastructure were modern web-based tools and APIs. We are always not that lucky when we go to some of the other places. There are mainframes hosting the claims system and so forth. But there is a messy middle that you have to get your answer on.

What we really focused on is where we’re going to generate value. You have all this existing stuff that’s doing a lot of things. Let’s identify the lever that will completely unlock the value for you when changed.

We learned that in that prior authorization process, the initial intake of cases coming in from faxes was a huge bottleneck. There was no tool for it. There were just people doing it.

Sramana Mitra: You have OCRs. OCRs have been around forever.

Ganesh Padmanabhan: OCR are basically optical character recognition that will read from left to right. You can get OCR for other languages as well. The age-old OCR problem has been on structured layouts inside unstructured data.

For example, if you fill up a form above and beyond, the key value pair extraction gets jumbled up because OCR is not intelligent. Now, the language models are supposed to solve that problem, but most language models are generative in nature. The amount of hallucination that you’ll see is not acceptable for these kind of use cases.

Lastly, there’re also a lot of handwritten components and badly scanned non-high resolution images. You have to use special techniques like super resolution and other things to enhance the image and then reason against what you’ve extracted.

So that’s just the OCR and the extraction of information. These people upstream are not just typing up information; they have to check other databases and ensure there are patient matches in the database. They’re making sure of the benefit eligibility, which is probably a document that has got five steps. They’re checking a spreadsheet to make sure the ICD codes are covered in that particular plan.

So, it’s a little bit more involved than just reading OCR.

Sramana Mitra: There’s sufficient workflow there to automate if you want to automate it.

Ganesh Padmanabhan: Exactly. We did that as a first step. The second part of the bucket was the medical necessity review wherein you have these nurses who literally have hundreds of PDF documents that are each 100-200 pages. They have to do a keyword search to identify the right document, go to the right section, review the information, understand it, and then look at the medical history of this patient that’s in PDF form, review all of that, and then keep that in their head. They usually use spreadsheets or notebooks to compile that and then match the two and make a decision.

So we said, “We will not give you another new mousetrap for prior auth, but we can solve these two problems for you in a very simple pilot. If we solve that, what does this mean for you?” That’s the partnership engagement we got into, and we solved it for them.

Since we did the patient matching, we had really good understanding of medical charts review. So, we could do that for the second problem. For the OCR problem, we actually built a vision language model and a technique called compound AI backed by Berkeley’s paper. AI Research Institute at Berkeley had a paper called Compound AI, which says that the ensemble approach of using the right tool for the right problem in cohort will deliver much better results than a general model that just does one thing.

Sramana Mitra: Small language models are much better for this kind of use cases than large language models.

Ganesh Padmanabhan: Exactly. Then if you want to summarize it, you use a large language model. You can feed the context to it, and then summarize it. We use that technique to do that. So those two agentic workflows that we built got us the entry into it. This was circa early 2023. The quality of the language models wasn’t as much as it is today, right? We were just ahead of the curve and solving a real problem. That gave us the opening. Now, once we did that, we mapped out the process and found other parts like benefit checks.

Then, there’s letter generation of a decision, which is often sent to a different group. Beyond this particular workflow, three months later, the medical management team and care management team are requesting the same charts that go through prior auth to look for a different context care gap. Are they doing their BMI under control? Do they have the BP reading under control? So, we started contextualizing the same charts coming through this use case for another use case. Six months later, the claims team is looking to review the same charts for DRG codes or itemized bill reviews for payment integrity and payment accuracy.

We could help them with the same thing. We realized that it was very critical to lay this out. We realized we found the product market fit, when two things happened – one, we went from a $10,000 pilot to millions of dollars of ARR in that one customer in a span of 18 months. We also learned that it’s the same common problem that we are solving at other customers. We went from one customer in one quarter, to two customers in the next, to three, then four to five customers every quarter on quarter. We could see the velocity of the market picking up.

We also learned the entry point is prior authorization. There’re a lot of players in prior authorization. Everybody wants to be prior auth, but we won because we are not just a prior auth company. It’s very counterintuitive. We go in and say, “We are laying down for you the foundation to solve all the other problems that’re going to come up across the enterprise to give you a better AI sidekick to run your operations.”

This segment is part 5 in the series : Building a High Growth Vertical AI Company in Healthcare: Ganesh Padmanabhan, CEO of Autonomize AI
1 2 3 4 5