Loading summary
Transcript59 lines
- [00:00]
A
Foreign.
- [00:08]
B
Welcome back to AI to roi, the Big Story edition. I'm Ray Reich, founder and CEO of Benchmarket, and joining me as always, is my co host, Peter Buchanan.
- [00:19]
A
Yep, I'm Peter Buchanan. I'm the founder of New Plan, and Ray, it is great to be here. This week we're diving into what I think is the most important AI success story that nobody talks about.
- [00:33]
B
Okay, a little bit of a tease here, but that's exactly right, because if you read and just pay attention to all the headlines, you think the AI market is all about OpenAI raising $110 billion, or anthropic in the federal government getting into a war, so to speak. A war of words and contracts and then using cloud and combat operations. Hours later, hyperscaler spending 700 billion funded by debt. The issue around private credit. It's just chaos, Peter.
- [01:04]
A
It is chaos. Yeah, totally.
- [01:06]
B
But spoiler alert, while everyone's focused on the infrastructure and foundation, there's a category of AI companies that are quietly delivering massive, measurable AI to roi. So maybe today we break down vertical AI. What it is, why the business model is fundamentally different from traditional SaaS, and how the funding environment has exploded for vertical AI companies. Then we'll profile for companies that really do define what vertical AI and why it matters.
- [01:42]
A
Sure. Let's start by defining vertical AI, because the distinction between different types of AI is critical. There are two types of companies in the AI application layer, which rides on top of all these famous models from OpenAI and anthropic and meta and deep seq. The first category is horizontal AI, and we all use those products. It's things like Copilot, Google Workspace, AI notion AI. These tools give you a capability and they leave the business application to how you're going to use them up to you. So they're useful, but they're general purpose and they're not vertical. They're not embedded deeply into functional vertical workflows. But vertical AI is different.
- [02:32]
B
Yeah, and I want to talk about vertical AI, but I want to ask you a question because we've talked about this before here in the podcast. Is that these coding assistants, Claude, Code Cursor, Lovable, Replit, the. They've really exploded. Would you consider them a horizontal AI application or vertical? Peter?
- [02:52]
A
I think they're a combination of the two. I think they're horizontal in that you can use them basically to do anything you want that involves software code. But of course, you're building something that's specific and vertical with them, and you're getting deeply embedded into a workflow where if the product is really successful and your pipeline is much faster and your code quality is higher than yeah, it's pretty much like a vertical AI product. So it looks kind of horizontal, but the top products replete Claude code, lovable cursor. They're incredibly sticky.
- [03:31]
B
They are. Well, but let's talk about the classic definition of vertical AI applications that primarily are built and purpose built for a specific vertical industry or functional workflow. And these companies don't make the AI models they typically use them. They're not raising a lot of capex to build data centers, they're renting space in data centers and then they're leasing the foundational models by paying for the inference and tokens on typically a consumption basis. But they're delivering real value, Peter, because they often don't just augment, but they can replace expensive labor. They can compress process execution time and even increase revenue. And I'm going to say this probably two or three times throughout today's podcast, an important part of these vertical AI applications is they're often not just assisting humans, they are actually replacing quite a bit of the labor that humans do. But here's a question I have for you. How is this different than what some incumbent, I'll say technology or software companies like Thomson Reuters are adding to Westlaw, they're using AI too. What's the difference?
- [04:47]
A
So it's a great question. So when Thomson Reuters adds AI features to Westlaw or Viva adds AI to CRM, it's a big improvement. Users appreciate it. But without major investment in product architecture, the way they price, the way they organize, organize their work, they operate in the pre AI world. So they are helping a human do the job rather than actually taking over the process of doing the job like a typical vertical AI product would do.
- [05:23]
B
Yeah, and I'll just harken back to a podcast episode we did about two weeks ago where we talked about three companies that really had made the progression from SaaS, a legacy SaaS company, to an AI first company. We talked about Notion, Canva and ServiceNow. So highly recommend people to go listen to that if they want to see how some companies are actually trying to make the transition to be more AI first or AI native. But to your point, vertical AI companies are typically built on top of an AI foundation from day one. Everything about how they operate is designed around that reality. Both a product architecture but also go to market. And the results, they're growing faster than any enterprise software company or market ever has. So that's why I think the vertical AI business model is fundamentally different. An example is Bessemer Venture Partners has invested over a billion dollars in AI native startups since 2023 and their vertical AI playbook that they recently published lays this out clearly. Traditional SaaS companies target IT budgets or IT and software budgets and vertical AI companies are targeting labor budgets. How big of a difference is this?
- [06:44]
A
Peter It's a huge difference. So in the US the total labor market is 10 times the size of the enterprise software market. So think about that. So the best, best vertical AI applications replace or radically reduce the cost of skilled professional labor, which is a huge profit driver and it's a huge efficiency driver.
- [07:09]
B
So Peter, that really does change the operating model for these vertical AI software companies. I'm going to call them software companies. It starts with something that you said earlier, that by focusing on the labor and the work being done, it opens the aperture of that target addressable market by almost 10x. So it's a much bigger market. The other thing that's really different is their pricing. A lot of the companies I've seen in the vertical AI marketplace, they're charging either on a usage consumption basis or an outcome basis rather than a per seat. So what's interesting here is the revenue scales not just based upon how many employees are using their software, but on the actual value that they're delivering. And there's one other aspect that I believe is going to be big for a lot of these vertical AI and agentic AI software companies that they may be servicing hundreds if not thousands of agents that are operating across a company and that could become another source of agent based revenue. But I'm kind of projecting out there a little bit, but I really see how operating models are very different and that even is go to market workflow. Because they deploy one workflow, they earn the trust of their target market, their prospects, and they expand from there. Now, I'm not ready to go to this is fundamentally going to shift how many employees are in go to market versus engineering yet, Peter But I do think we should watch how a vertical AI software company's R&D function versus their go to market sales and marketing, what percentage of revenue they're spending and how many people, because I do think that could change, but I'm not ready to make a prediction yet. I don't know if you want to say any more about that.
- [09:08]
A
I don't think the predictions are going to be as dire as you're reading in the paper, but I think a lot of jobs are really going to change. I think the bigger thing to really consider is once these products get in, you've attacked a few workflows. The first one is successful. You do workflows 2 through 5. These products become incredibly hard to displace. Why is that?
- [09:33]
B
Well, this is where words and the nuance of these words are really going to matter. Peter, we all talk about moats, right? We're vertical AI products. They don't sit on top of existing systems. I know, I saw the OpenAI, I think they called it their frontier model where they had agentic AI systems sitting on top of traditional systems of record. I think it's different. I think that these vertical AI and agentic AI systems are embedded within a business's workflow and they integrate into legacy systems of record. But what they're doing is they're ingesting this proprietary data. They're building domain specific models that don't even just improve, but they become more intelligent with every interaction. So that's why I think vertical AI is so, has so much potential because it becomes smarter and is able to actually implement continuous improvement with, with every workflow and every transaction that they make. So it's going to get hard to replace them, Peter, because it's like having an employee that's been there 5 years, 10 years, 20 years, get smarter because of their applied knowledge that 10,000 hours to be, you know, really good at something. And guess what? These vertical AI in their agents aren't going to need an annual raise and, and they're probably not going to be asking for a promotion.
- [11:06]
A
No. Until you read the science fiction books and then it's going to happen. Displacing it vertically. If you want to take that out, you're basically ripping out your institutional memory. So that's actually a bigger switching cost than switching from SAP to Oracle or PeopleSoft to Workday was back in the day. All right, so let's move to following the money, as in the funding explosion for vertical AI companies. So vertical AI companies raised $42 billion last year. That's a big fundraise, doesn't meet the level of anthropic or OpenAI, but it's spread across a lot of companies and that's up from 22 billion in 2024 and 8 billion in 20 in 2023. So investors have definitely gotten their religion on the potential of vertical AI companies. Ray, that's a massive acceleration. Where are these companies putting that money to use?
- [12:12]
B
Well, I'll talk about that in a minute, but I will tell you, I thought maybe you'd go down a limb here, Peter, and provide some estimates of how much more we'll see in 2026 and 2027. Because honestly, I think we're in the first one or two innings of vertical AI software investments right now.
- [12:29]
A
Oh, that's totally right. You're seeing fundraisers four, six, eight months apart where a company raised $100 million and then raises three or 500 million after that. You're seeing companies that were at 100 million in ARR in the middle of 2025, ending the year at $200 million in ARR. So there's hit records across multiple vertical industries and VCs are going to notice and the valuations look high now. But the revenues of a lot of these vertical AI companies are justifying the valuations. They're in essence growing into them.
- [13:10]
B
Well, in fact, if we double click on the data, in 2025 we've saw multiple mega rounds of 100 million or more. In fact, those kind of mega rounds accounted for almost 79% of all AI funding. According to CB Insights, we saw multiple vertical AI companies closing rounds of 300 million or more in a single year. Harvey, I know you're going to talk about them as a legal vertical AI. They raised two such rounds in 2025 alone. And once again, this is only the beginning of vertical and agent AI software. So you know what I thought about and I thought about this from the COVID era. It's like, oh, we were talking about all these unicorns that were being minted. I think there was like, boy, if I'm not mistaken, it started with 100 and almost went up to 1000 unicorns. Are we seeing the same phenomena here in vertical AI?
- [14:07]
A
Unicorns are no longer rare. Little girls all over the world are going to be really upset that their unicorns are all over the place and they're not special anymore. Because healthcare AI went from 2 unicorns in 2023 to over 15 at the end of last year. Legal tech went from one unicorn in 2023 to eight plus now financial services grew from three to 12. Industrial and manufacturing went from one to five. So you know, we're talking about this continuing. There doesn't seem to be any blockers.
- [14:43]
B
Yeah, one of the things I want to start tracking here with you at ROI, I want to see how many of these vertical AI companies become decacorns. $10 billion valuations. And even you may have heard it here first to listening audience, how many center corns that we create companies with 100 billion above valuations, right. And here's a pattern that, you know, I don't know. It's a pattern, but definitely there's some common variables of why vertical AI is so attractive, because they're targeting sectors with high labor cost. I mean, come on, health care, legal, Right. Low prior technology adoption, an example. Healthcare industries have always been one of the laggards in their IT investments. We're not talking capital equipment like X ray and MRI machines, but software infrastructure. It was typically only about 1.5%, so real laggards. So that's why I'm so interested in healthcare, because not only have they traditionally invested less in it, but Menlo Ventures report that we talked about in one of our AI to RI newsletters. Healthcare is the number one industry as measured by, I think it's been by healthcare companies. Now, they primarily are using IT for clinical documentation transcriptions. But legal has also been an early winner in vertical AI, and there's some fast followers. There's things in financial, back office, industrial operations, and dare I say, maybe we'll see government services becoming at least a little bit more efficient right around the corner. So it seems like these waves are coming faster and faster. PETER.
- [16:24]
A
Oh, absolutely. So, so let's dive into these four companies that we previewed before. Harvey is the first one. Harvey is basically a legal workflow application was founded by a former litigator named Winston Weinberg and a researcher who left Google mind. Their Last valuation was $8 billion at the end of last year. And they're apparently in talks to raise money at an $11 billion valuation only a few months later. Their ARR went from 100 million in August to 190 million by the end of the year. So that's 90% growth in four months. They have 1,000 customers in 60 countries. Over half of the Amlod 100 use them, and they have over 100 lawyers on the platform. And so those are gobsmacking numbers. And so what makes Harvey, from a product architecture and delivery perspective, so compelling? Ray?
- [17:35]
B
Well, before I answer that question, first of all, it sounds like we need to welcome Harvey into the Decacorn category. If they're raising at an $11 billion valuation.
- [17:45]
A
Yeah, they're not there yet. So, you know, they have to work for it. Ray.
- [17:49]
B
Exactly. At least another a month or so, right?
- [17:52]
A
Yeah, yeah, exactly. Work really hard, and then they can take a vacation after April 15th like all the accountants do.
- [17:58]
B
So to answer your question, and this is becoming more and more common today, that their product architecture is deliberately modular, and what that really means is the right model for the right job or the right task. So instead of having every query go through a single model, they orchestrate the workflows using different AI models depending on the task, the document type, whether it's research, whether it's agreement or contract drafting. They do this for two reasons. The right model for the right job, to have better performance. But also we're seeing a lot of CFOs and the head of AI software development having discussions about, well, what if millisecond latency is not needed for that particular task? What if we could use this small language model or cheaper language model to try to get those gross margins up from 40, 50 to 60% plus. So that's one is the orchestration across multiple models.
- [19:01]
A
Right? Plus they have this no code workflow builder. And so that lets law firms construct their own task specific agents without Harvey basically being a shoulder angel on top of their lawyers. So they can configure things to deeply integrate into the document management systems. They can attach themselves to firm specific data. That again, makes displacement really difficult. There's a quote, admittedly from the Harvey website that sort of captures the stickiness, which is, if we took Harvey away from our staff, there would be a riot. So kind of like that.
- [19:44]
B
You know what this reminds me of? I managed a lot of sales organizations and I can promise you, not once did I Hear a professional B2B sales representative say, if you take salesforce.com away, I'm going to riot.
- [19:57]
A
No, no. They'd be offering to buy you drinks.
- [20:01]
B
Exactly. Okay, let me move on to my favorite vertical AI story. And that's Healthcare. Can you share an example of a vendor there?
- [20:08]
A
Sure. So there's a company called Abridge. It was founded in 2018 by a cardiologist named Shiv Rao. And Abridge builds what they call ambient AI that converts physician patient conversations into clinical documentation in real time. So the stats are $5.3 billion valuation from their June 2025 Series E. So they're getting up there, they're catching up to Harvey. They have contracted. They don't release revenues quite as frequently, but as of the first quarter of last year, they were $117 million in revenue. So they were the leader. They work with over 150 healthcare systems, including 40 different locations of Kaiser Permanente, University of Pennsylvania Medical Center, Johns Hopkins, Mayo Clinic. They work with physicians and clinicians in 55 specialties in 28 languages. So what's different about their transcription tool, Ray? Because this seems like it's basically a Godsend.
- [21:21]
B
You know, I feel like I've done this to you a couple times during today's podcast, but you just. Something just struck me that I have to share with the audience. You know, I was in Silicon Valley and around software companies for almost 30 years, and almost every founder was more of a product centric technologist. The two that you've mentioned here so far, Harvey and Abridge, founded by deep domain experts, a lawyer and a doctor. So it's just something I think we should keep track of because that subject matter expertise is so critical to vertical AI.
- [21:58]
A
Yeah, well, you know, the lawyer has a Google DeepMind person sitting right next to him. And I think it's not in the script here, but I think there was a very similar person next to Mr. Rao. So the smart professionals are going and finding the technical geniuses to turn them into, turn their brains into software, you know.
- [22:19]
B
But you asked me a question. How's Abridge, or Vertical AI for health care different than a general transcription tool? And the key is, you know, a general purpose transcription tools. There's great ones out there. Heck, we're recording this on Riverside FM and they're going to transcribe our podcast. So we have the written transcript. But Abridge produces a very structured kind of soap note, it's formated to the relevant specialty. You know, pediatrics versus ob, GYN versus ga. It integrates with the electronic health care record. And that's so critical because one of the big advantages of Abridge is their deep integration into epic, which is the leader in electronic medical records. And all this is done without the physician having to touch a keyboard. So to me, that is very different. They also have a partnership with Highmark Health, which reduces the prior authorization timeline from weeks to minutes. Think about those of you in the audience, if you have to wait to get the approval to go see that specialist. And even though they're not doing it today, I can see them continuing to expand their footprint where they may even be able to over time start to do things like insurance coding, which is a downstream process. But if you're capturing all the information from the doctor, it's a natural extension to start doing coding. So it's just amazing how they could capture larger and larger part of the pie.
- [23:54]
A
And in fact, that's where the prior auth and revenue cycle management things that they're doing are coming from, because all the data to facilitate that comes up comes out of the interactions with the patients. I have a client right now that's building a product for those specific use cases. Right now. And it's actually pretty much the hottest use case in healthcare right now is prior auth and getting revenue sooner. So getting back to abridge here, the Wall Street Journal just basically said they're the widely considered leader in the medical scribe market because once the healthcare system has trained clinicians, integrated the ehr, built the quality benchmarks around Abridges output, the cost of switching is very high and doctors and nurses don't want to go through the process of doing it because it takes time away from patients and they're comfortable with it.
- [25:02]
B
Okay. Hey Peter.
- [25:03]
A
Yeah.
- [25:04]
B
Now that we're all healthy, let's talk about another category. And it's customer experience. And quite frankly, customer experience could be viewed as one of those a vertical AI tool, but with a very deep horizontal process. And Sierra is a fascinating company because it was founded in 2024 by Brett Taylor, who was a former Cosio at Salesforce. He's the OpenAI board chair and sitting beside him is Clay Baver from Google Labs. They officially hit decacorn status that $10 billion valuation in September 2025, and they actually hit $100 million in AR in 21 months. So here we have an enterprise application acting more like an individual kind of lovable type revenue trends. But I think Brett Taylor also was recently quoted on potential growth with vertical AI companies, right?
- [25:56]
A
Yeah, he did. He said reaching 100 million in AR that quickly makes Sierra one of the fastest growing enterprise software companies in history. And he was wondering aloud whether the agent era could produce several trillion dollar enterprise software companies, which would be a first.
- [26:16]
B
In fact, I don't know if you heard, but Jensen Wang from Nvidia was at the GTC conference and yesterday said, hey, we're going to do a trillion dollars in revenue by the end of 2027. And there was a little bit of a misunderstanding because that was combining 25, 26 and 27. But what Brett's talking about is one year trillion dollars. Amazing. But you know, one of the things I want to do, because we're already coming to the end of this podcast, but I want to go into one other sector that's really near and dear to my heart and it's the industrial space manufacturing. There's even a vertical AI there, right, Peter?
- [26:53]
A
There is. It's called Maintain X. It covers a category that gets way less attention than legal or health care, but it's a massively underserved market and that's maintenance management for physically physical asset intensive industries. And so, so all those manufacturing plants with huge machines that have to keep working. And it's really expensive when they break. It's that industry. And so they are at a $2.5 billion valuation as of last July. They have 11,000 customers. They're managing 11 million assets. They processed more than 27 million work orders. So, Ray, why is this so important?
- [27:41]
B
Well, beyond the fact that equipment failures alone is. It cost about $1.4 trillion now, that's annually. And most companies are still managing those in a very manual way. Spreadsheets, paper orders, etc. But I am so bullish because I think manufacturing or industrial AI could also be a catalyst to onshore more manufacturing back to the US where we offshored because of labor arbitrage. I'm thinking we're going to see AI agent arbitrage be a reason for onshoring. But that's really a Ray Reich perspective.
- [28:19]
A
Right, Based on your career, spending so much time with manufacturers. Right, Right. So, so maintain X. Actually, they have a really powerful data network effect. So here's how it works. The platform sits on years of proprietary operational data sets, work order histories, asset failure records, maintenance logs for millions of assets. And when they recommend a preventive maintenance interval or flag an anomaly, they're drawing on patterns from across the entire platform, not just one company. So the customer outcomes are absolutely fantastic. So first, 34% reduction in unplanned downtime, 15% increase in production capacity. So costs are going down, revenues are going up, and 32% savings in monthly maintenance costs. So that's what you would call a trifecta.
- [29:24]
B
I agree. And boy, that institutional knowledge and intelligence they're building by collecting data across companies. Instead of 100 units, it's a thousand tens of thousands or even hundreds of thousands. Much more intelligent. In fact, that kind of brings me to wrapping up. And can we kind of summarize a little bit about why vertical AI has such big structural advantages?
- [29:47]
A
All right, so we've profiled four companies, so let's just talk about why, why these durable advantages actually occur. So first, Ray, you talked a lot about labor budgets earlier, so that's got to be one of them, right?
- [30:02]
B
Yeah, I keep going back to that. Now we've increased the target addressable market by almost 10x because we're going after labor, so it's a much bigger market. But then we also talked about a second advantage, Peter.
- [30:15]
A
Yeah, so proprietary data and intelligence compound. So the promise of AI people talk about it, they talked about it in science fiction since the 50s is the AI gets smarter and smarter and smarter in vertical AI that actually really happens. That creates a switching cost that's beyond just the cost of paying for new software and implementation. Institutional knowledge is hard to migrate from one application to another. And AI just builds and builds it.
- [30:51]
B
Hey, Peter, let me jump in there. Because what happens when the foundational models get better? Could that threaten these vertical AI companies?
- [31:00]
A
No, I think it gives them more power. It gives them the ability to do more knowledge. It gives them the ability to create better data. It gives them the ability to actually create modes. Ironically, the models are increasing at this incredibly rapid pace and you have so many choices now that, that they're getting commoditized. At the same time, they're creating moats for the companies that are running on top of their software.
- [31:27]
B
In fact, I think Sequoia Capital had a recent piece called Services the Network Software, the New Software, and it really nailed it because it says if you sell the tool, you're a race against the model. But if you sell the work and the outcomes, every improvement in the underlying model makes your service faster, cheaper, better, and harder to compete with. Hey, Peter, I hate to do this. We're going to wrap up. So can we just jump to some key takeaways regarding this vertical AI transformation? And I'm going to start. Okay, I'm going to start with that. I think vertical AI companies may be building the most successful AI businesses we're going to see. And they raised 42 billion last year and it's going to increase dramatically. The unicorn count went up from 9 to 57 to 56. So what is that like in 6x increase? And multiple companies in this category have already hit 100 billion ARR in under 2 years.
- [32:20]
A
Right. Plus the business model is different than SAS. Are they competing for later labor budgets? 10 times as much money to go after when you're selling software. Aligning with customers. So a seat model really didn't align with customers. It just made sure people had software to use. But when you're both betting on a particular outcome, both the vendor and the enterprise, that's really good. And you've got the stickiness of the embedd embedded workflows. And we have a lot of examples.
- [32:51]
B
Yeah, and those four examples we gave, there were some common themes I saw there. Founders and teams with deep domain expertise products that were built on an AI native architecture from day one. They're built into a process, not layered on top of a foundational model. And that their business models actually compound the benefits of, from revenue growth to customer relationships. And that proprietary intelligence that they're making every transaction, every workflow, every deployment. So I guess with that we should wrap up and kind of do our outro, if that's okay.
- [33:30]
A
Sure. So I think the big AI headlines are going to continue to focus on OpenAI, Google, Microsoft, the hyperscalers, all this, all the circular funding, all the government drama, all that sort of stuff. You know, data centers. But that's not where the real action is, I think, and it's not where
- [33:52]
B
the real ROI is. Right. We're already seeing tangible return on investment in vertical AI. Part of it is because they are, I'll say, supplacing supplement and replacing expensive labor. They're compressing that process execution time and increasing revenue. So another perfect trifecta.
- [34:13]
A
Right. And so if you're an enterprise buyer, there's things you should do. Like if you think about this. So evaluate vertical AI vendors. Whenever you're doing a category review for software, insist on outcome based pricing. And if they won't give it to you, that's a red flag. Ask hard questions about proprietary data switching cost, verified customer results. Start with one workflow and expand. Don't go all in on the first day. So the companies that figure this out in 2026, this is the year they're going to start to really build a durable competitive advantage where it's hard for their competitors to catch up.
- [34:56]
B
Yeah. And even though there's going to be a lot of focus on big IPOs this year, with both Anthropic and OpenAI possibly coming out, the real story, and I think it's a big story for the Future, is this. 2026 will be the year that vertical AI companies are building durable competitive advantage in the next generation of software. So I want to thank the audience for turning in here to AI roi. If you haven't already, if you want to dive deeper into some of these topics, go ahead and check out our newsletter at ai2roi. That's ai2roi.substack.com Leave us a review and. Hey Peter, we'll see you next week.
- [35:37]
A
Yep, see you next week.