Host Intro (Gene Kim)

The person speaking next actually shipped Amazon Prime Now. And when we met in 2017, he told the amazing story of how in 2015, as an engineering director at Amazon, he actually made that capability happen.

And so this was actually the third attempt to make it happen. It was so complex because it touched over 300 services across three of the most powerful pillars at Amazon — shopping, transport, and fulfillment. And so it takes a really unique type of leader to be able to make that happen.

So I'm so pleased that since then, he's been CEO and founder of Skypoint, which serves in the healthcare space. They serve over 800 healthcare and senior living organizations. And Q4 of last year, he had his aha moment of coding agents at scale.

And this is an amazing story when he told me about it — where they went from... they shipped more products in one quarter than they had for the entire eight-year history of the company. The team size has changed drastically. And so I'm so grateful that he's going to share his story. And this is actually the third presentation today told from the perspective of a CEO.

So with no further ado, Tisson.

Tisson Mathew

Since... probably started with Amazon L8, and so I was an L8 engineering director there, and I left after two years because I couldn't get to L10, you know what I'm saying? It is a hard path to get there.

So, what I want to talk about today is my own personal journey and how AI kind of helped unleash our own internal capabilities within the company — to make real applications and agents in production with our customers. So I kind of coined the term "from analytics to agentic apps."

The reason why I did that is Skypoint as a company started as a data integration and analytics company. And providing analytics to decision-makers in healthcare is powerful, but analytics doesn't do anything. You know what I'm saying? It just gives you more reports. There's more dashboards, and healthcare people look at it, and they go home. That's what happens, right? And agents and agentic apps actually do work. So the transition from analytics to agentic apps.

The second modality that we touched on is how to replace legacy SaaS applications with agentic apps. This could be built in-house. We have a set of commercial applications — I'm talking about the names that you all know within the industry, the companies that have been around in the SaaS space for 20, 30 years, and they have old code bases, a lot of legacy customers. But the majority of the customers that use our products and others within healthcare usually use like 5 or 10% of the features, and they don't need these legacy applications that millions of dollars are spent on. So they are looking at this and like, "Can we build these things in-house? Can we buy from AI-native companies like Skypoint?" So our product line replaces about 100 or so legacy SaaS vendors. And I'm not talking about SaaSocalypse or anything like that right now, but there is an element of truth here, so I'm going to get into that as well.

Currently we are in about 1,100 locations. So a location for Skypoint is a hospital, a provider group clinic. We serve the senior care market, which would be a senior living community at a specific location. So we are deployed in about 1,100 locations. Each location has a lot of users. Healthcare, in general, is very people-heavy — 60% of all expenses are for people. So there is a lot of TAM for agents to automate a lot of the drudgery that people are working in healthcare. For example, the administrative expenses in healthcare is about a trillion dollars a year that goes on just paperwork.

So we've got now about 20 different AI products. I would say over 90% of it is AI-coded. And we had integrations before, so now these integrations are bidirectional. We've got about 50 FTEs. We were about 120 — it came down two-thirds of the way — and then we have added a few business QA folks internally. So our goal is to get about a million dollars of revenue per employee. So essentially, how do you measure now velocity or output value? What is your moat? There's a lot of things in question right now for commercial software companies. So these are some of our pillars and where we are and where we are heading.

This is my journey personally. In 2017, I was with Amazon Prime Now — great organization. Most of the work that I did was creating services and scaling them in logistics. And if you look at those teams, there were a lot of engineers trying to figure out and optimize algorithms and things like that. And back then it was microservices. I don't know if anybody here is a fan of microservices anymore. Did it go away? Is it all gone? It's gone. So yeah, I'm completely irrelevant. It's good that I left and started this thing, right.

So 2020, I founded Skypoint, primarily around data integration analytics to solve problems within healthcare. So fast forward to 2026 — when we started, for the first four and a half years, we had one product, which was primarily an analytics and integration product. And a lot of our customers didn't even know what it actually did, because what they see is a dashboard in their Power BI or Tableau or somewhere else. We just never had the direct interactions with customers. And it's very difficult to scale data integration and analytics products at the application layer. You have the Snowflakes and Databricks of the world, which is at the infrastructure layer — but nobody sees that. Nobody knows these things within the business users, right? They don't know. They don't interact with it. What they interact with is a Power BI report or a Tableau thing or a custom dashboard they're acting on, right?

So what transpired for us is we now have real applications that are used by end users. Now we have direct exposure to those end users. It's completely transformed our company and what we did. And I was talking to Gene about this — everything that we did in four, four and a half years, we completely redid in a matter of a few weeks, the same products, and then scaled it to other products in the last six months or so. It's a complete transformation, but there was a lot of hard work. There were several weeks and months. I started like 6:00 a.m. in the morning. I don't go to bed until like 2:00 a.m., right? Trying to code these things, the version ones of these products, figuring out what to do, things break, and I don't know what happened.

It's very hard to deal with these things. It is not like not everything is smooth. So then getting this to engineers and saying, "Okay, now you take what I did," and it was like, "Oh, this is all slop. I've got to throw it away, and I have to redo this whole thing again." Right.

So that's kind of my journey. And then what was legacy at the so-called SDLC within Skypoint was two-week sprints, and there was a bunch of product managers taking so-called requirements, writing PRDs, and then getting it to engineers. They misunderstand those PRDs. They code, then go through QA every two weeks, something comes out, and then it takes sales several more weeks to get it to the customers. It was just slow. A lot of money being spent, but the business value we couldn't even measure.

Then the second one was the team size ceiling. There are a lot of things our customers want. We are constrained by people. And then we have a forward-deployed engineering function within the company, and those FDEs started ignoring the product entirely. Like, "Hey, you guys are really slow. I'm just going to code this myself." So essentially, we have a product line which our own FDEs are not excited about because it's slow and it doesn't move fast. So they are just coding things themselves.

So essentially we have a whole bunch of problems, right? You have customers who don't really know our product, and then our FDEs are kind of creating their own thing, and our product guys think they are creating something magical. But none of them connect.

Integration — that is another thing. We create integrations to systems and get some parts of the data over there. It takes more time. Things break. Pipelines, data pipelines break all the time, and we don't know what the hell happened. Then we have to retool it, rerun it. Most of the time, our L1 support guys are like, "Oh, we just reran the pipeline and it worked." Who knows what the hell happened, right? There is no way to kind of go through the telemetry and figure out, okay, what exactly happened.

Now things have completely changed. Compliance was another thing. So we are a HiTrust R2 certified product, which is more like FedRAMP — mid-level of the FedRAMP regulation. So it is a highly regulated product. You need to have all the checkpoints, things like that. So compliance was a problem when shipping.

So I've kind of thought through: what is the true value of software now? Is it code? Is it something else? Is it the culture of the company? Your velocity to ship, proprietary data, your sales and distribution? What is it actually, right? So I'm kind of thinking of this harness — I know there is an agent harness, but it's more of like an organizational harness. So if you can actually build an organizational harness around context — you have data and your telemetry and all of these things together — then you have a moat. You can really go to a customer and say, "Hey, I've run this product through thousands of users. It understands context and workflows very well. This product integrates to your workflows. We have tested this." That is a real moat. And code — because code is now, you can keep shipping code, but it doesn't really mean that the output business value is there or not.

So we are primarily a Claude Code shop. We started about a year ago on it. And we use Replit for some of the early prototyping and V1s. Before that was GitHub Copilot. I personally use Replit for some of my early prototyping, Claude Code, and Codex — three of those. I've used Codex quite a lot for validating certain things. And Cursor — I try to kind of put it in plan mode, get the output, give it back to Claude Code, say, "Hey, can you go figure this out?" for some complex scenarios.

So right now we are working on a few things around multi-channel integration, like SMS, text, phone calls, things like that. So how do you actually coordinate all of this stuff? So that's kind of our turning point.

So now, how agentic code is in production. We have customers now who replace a legacy SaaS app that existed for 25 years with our products and they don't know the difference — because they are in Teams interacting with an agent and the agent is giving them answers, or they're in Slack or any of these media.

So what we are trying to understand is, when we are actually shipping an agentic product into the hands of these customers, how they're going to interact with it. So I'll give you one example. Are you guys familiar with prior authorizations in healthcare? Where you go see a doctor, they need to get a prior auth from your insurance plan for, like, an MRI. So this was entirely done by human beings. They basically take all of your medical history, take it into a payer's portal — like UnitedHealthcare or others — fill all of this out, and then they push a button. And you know what happens in the insurance company? They deny it, and then they clawback, and then it goes back in circles, right? And this is full-time employees doing this.

So we have a product called SkyAuth, which is doing this all entirely in an agentic workflow. And when we first deployed it into a dermatology group — they have 40 different locations in the Pacific Northwest — these employees were watching this thing happen, and they were like, "Okay, should I go?" So when this thing is moving, their hands were like trying to get under the keyboard, like, "Okay, is it going to go right or not?" So now I've seen, like, in the last few weeks, 200 prior auths went successful with no denial. Because most of the time the denial happens because you miss something or mistype something. So now AI is so good at this, and it's getting better and better.

So it is using computer use, it is using data integration, the application. So now all those prior authorizations — the benefit to a patient is you can get your procedure done the same day. So you walk into a dermatology clinic, you need a scan. You don't have to go home because it became a rescheduling event, right? So the true impact of it is not only labor replacement, which is great and amazing, but it actually is the value that you're delivering to your end customers. They can get their procedure done faster. The scheduling becomes a much easier problem to solve for.

Another one is referrals — referrals for a patient that is coming in. So it is all faxes. So we are reading faxes and then routing for certain things — prior auth versus not. So these things were traditionally done by UiPath-like companies, and it was brittle. Most of them broke and most people didn't use it. They spent a lot of money on it but really didn't see the traction.

So what we are doing now is the second part, which is Claude Code. We use the Claude agent SDK a lot in our products. So essentially it's Claude Code as a library. So we have skills and agents and the harness we kind of put around it. I actually looked at NeMo and OpenAI and adopted some of the things inside our agent system — like good memory, ability to text it, message it, and you can actually control it by messaging. But think of Claude Code as a library within your product, and your Claude Code creating your product, right? So you're kind of Claude-built quite a lot.

And we also use Gemini for a bunch of other things like vision modeling. Our voice works on Gemini quite a lot.

So the last one is on HiTrust guardrails. That is where we use Codex to do a lot of the checkpointing, validation of things, stuff like that. And then eval loops in production — that is something that is somewhat misunderstood. When we started releasing products into production, we started to see that the business QA of the products requires more human hands. You can do automated testing a lot, but how do you validate certain things — how a true business user flows through, right? We tried computer use; it didn't really work. Playwright, MCP, all of that — it kind of works, but it doesn't really truly flow through the business case. So we used a bunch of techniques and then added more humans, mostly interns, going through these cases and figuring out if it's actually working or not, and then working those through.

So this whole thing — we are releasing every day. We used to release every two weeks. And actually I look at it like, okay, there is so much software coming at our customers. How are they even going to absorb this, right? What is our adoption? How are we going to message it to our customers?

So every day we have a release report. We get a QA report, bug trends, how the output value is — lines of code and all of that kind of stuff — which is very interesting. Our lines of code in February... I don't know the numbers in March. February was about a million lines of code. But of the million lines, probably we deleted four million, right? So essentially there is a lot of rework. There is a lot of refactoring going on. There is three to four times more code deleted than added.

So essentially I don't know what this path is going to take, but depending on what features are getting built, you need a continuous refactoring technique. So I don't know what the SDLC is called. You're releasing daily. Your product managers are not writing PRDs anymore. You have GitHub and GitHub Actions completely running wild. And your products are getting out. And your users are using your products. Tickets are coming in. I'm looking at this thing, amazed — I don't even understand what is really going on, but things are happening, right? So it is just a lack of understanding of what is going on under the hood. But it more feels like physics to me. Like, okay, things are working, so if you can put these into the right type of frameworks...

So we shipped about 20-plus AI-coded products. We put them into different categories — everything is "Sky something." And our customers can buy one of our products individually. They don't have to buy the whole suite. So we have one in growth and engagement: SkyAdvisor, Reputation, Voice, Sign, Engagement. Revenue Cycle is a bunch of them. Workforce Management. We have analytics products, and the last one we have is what we call SkyBuild, which is — you can actually build an AI-coded product within our internal IT teams, and they can deploy it in our O2 environment.

So all of these products, if you think about it, are displacing about 100 legacy vendors. Take SkySign, right — it replaces DocuSign, Adobe Sign, with the integration to operational workflows. So our customers are like, "Okay, so we replace this thing with your thing, you get a cost benefit, but also it integrates to our systems. It has a whole bunch of agentic workflows. What is the output value? Can we actually get consent forms signed from patients faster? Can we actually get our contract lifecycle management done faster?" So essentially we are measuring business value output from all of these products, and we can continue to churn out more Sky products.

Now we are in — I don't know, call it a software factory. Essentially, we have a factory that can create these product lines for solving specific use cases. And then we have a common agentic infrastructure underneath all of this, which at the moment is Claude agent SDK-based with Gemini wrapped around it. It could be something else in the future, I don't know. But that's kind of the current state.

And so these are the numbers that we got to. So again, I interviewed developers pre-AI agentic coding and then post. The numbers I got are from 5 to 12x overall output from engineers themselves. One of the things that I heard the most is problems that they never thought they could solve — they solved. And their biggest frustration is stupid mistakes by AI just swapping things happening, and they had to go clean it up. And it was like, "Oh, why did it happen?" Is it a skill or is it a validation issue, or what is it? And then we were like, "Okay, you can't have this kind of thing run wild. We need common libraries." So we have a SkyPoint UI library now, a SkyPoint logger library. We started to create a lot of frameworks, and those frameworks are common. So we can limit the blast area of these things producing bad code.

Then shipping in weeks, not in quarters — essentially shipping to customers. 20-plus products. Cost of R&D actually went down by 50%. So we are producing more — probably 20x more, or 50x more, I don't know, a very large number. But in terms of the R&D costs, it went down. But we are seeing a cost increase on customer success, awareness, education. We are asking our customers to hire business champions within their organizations to really deal with what these things are going to do for their business.

So — what broke, what worked?

Some of the things that broke is context collapse. It is a messy, hard problem. I don't know how many of you guys have solved this problem or are dealing with it. Inside our organization and with our customers, this is a problem. Systems not understanding what is really going on. Yes, things are improving in that space, but agents still make mistakes because they just don't have enough understanding of what is going on, right? This could be memory. We have a bunch of things that we put in to solve some of these problems, but it is not a solved problem.

The second is testing and business QA that I touched on, and then sensitive data exposure is another one. Because there is so much software, it can access so much information. Are we exposing sensitive data anywhere? What is the telemetry, how can we actually observe it?

Then compliance as code has worked. We are able to test things and run automated tests quite a lot for compliance. Quite a lot. So that's been helping us — hey, are privacy checks there or not.

The last one is small teams. We started to experiment with increasing the team size of some of our products — take SkySign — it's actually two guys. One is really good, and the other guy is kind of ramping up. But if I increase that team size to five, their throughput, their velocity and output value, actually goes down. It's a very interesting equation. I don't know if there is enough things to go in parallel with these products, or the way it is. So we are seeing there is a certain two to three number of engineers or PMs — they're full stack — that actually produce magical results compared to adding more folks into it. So essentially we shrunk the teams more to make sure that these things go in the right way.

So these are the three takeaways.

First is build a harness. I think it's a variety of different things within an organization. It could be data, it could be context, it could be an agent harness, whatever.

And the second is shift compliance to the left into code, just like security. Compliance means many things to many people, right? User experience to policy, or whatever.

And then redesign the team. So if you are trying to scale, you need a new way of looking at this whole thing. It's not necessarily headcount. What it means is a smaller company can have a bigger impact relative to others.

All right. That's it.