The State of Generative AI in the Enterprise 2025

Live from Schenectady, New York, it's the Metrics Brothers, Growth and CAC.

And I'm Growth, better known as Ray Reich, founder and CEO of Benchmarkit.

And I'm CAC, better known as Dave Kellogg, EIR at Balderton Capital, independent consultant, and the author of Kellogg.

And together we are the Metrix Brothers.

And we go together like AI and coding.

I'm sensing some vibes here, Dave. Did you spend some time using Replit or Lovable this weekend?

Well, Ray, I've actually not used Lovable before, but I have used Replit to build some prototypes. Um, but mostly because I'm intrigued by all the kind of buzz and growth flooding into the AI market, including tools like Lovable and Cursor, but also more generally in enterprise AI.

Well, both of those companies were part of the episode when we talk about the new AI metric months to $100 million ARR episode. So the impact that AI coding is having on software development productivity and honestly enterprise spend on AI tools is very intriguing.

Yeah. And that's— look, I think this is covered in part by a recent report from Menlo Ventures. Entitled, oddly, 2025: The State of Generative AI in the Enterprise. And I know it touches on those topics and more. And I think that's going to be the topic of today's episode. It is.

I read this report in late December. It just came out in December, and I really liked some of the first-time data that I saw, and I look forward to discussing it. But before we do that, let's hear a word from our sponsor. The Metrix Brothers is proudly sponsored by Benchmarkit. The B2B AI and software industry's largest, broadest, and most contextualized benchmarking index. Dave Kellogg and Ray Groth Wright have put the Metrics Brothers podcast on hiatus for the remainder of 2026. All episodes we have ever recorded are still available on your favorite podcasting app. You can reach Dave at Kellogg and Ray on LinkedIn at Ray Reich. Thank you so much for listening to the Metrics Brothers podcast. Okay, Dave, I'm going to give you just a quick overview of the report. Okay?

Sure, right.

And, and for the listeners out there, if you want to download that report, the overview is not in front, it's actually on page 18, and it was based upon a survey from 495 US-based enterprise AI decision makers. They conducted this research throughout November. It was a wide variety of participants that didn't give us specific ends for each category, but they said included C-suite executives, VPs of engineering and product, technical leaders responsible for AI purchasing and development. So more of an engineering and development than a finance or procurement report. And they also on each page throughout this, they do use some third-party research and proprietary data that they footnote on each page, Dave.

Got it. So Ray's talking about really kind of methodology and sources page because you said the overview is on page 18 and that definitely confused me. But, but yeah, the sources and methodologies are on the last page of the report as Ray covered. Bear in mind, this is a report produced by a venture capital firm, Menlo Ventures. And while they do a pretty good job of footnoting where they're getting stuff and how they're kind of deriving or imputing it. Um, bear in mind they do have skin in the game, among other things. Like, as we speak, Ray, there's kind of a Twitter controversy happening between people from Menlo and people from Khosla, uh, because I believe Menlo is an investor in Anthropic, and the Khosla guys who are investors in OpenAI are challenging some of the very claims in this report. But in any case, with that caveat, I think we're ready to jump in.

Okay, well, on the first page of the report And they, they call it AI Boom versus Bubble. And what they show here is how enterprise AI software spend has increased from $1.7 billion in 2023. Makes sense since ChatGPT was just released in November of '22, all the way up to $37 billion in spend in 2025. And that grew from $11.5 billion last year, Dave. So a 3.2x growth. And generative AI software spend. And I always like to compare things to the SaaS industry. And when you look at the SaaS industry, which in earnest really started in '99 to 2000, it took almost 16 years to surpass $37 billion. So my big takeaway from this is how fast the enterprise spend on generative AI software is growing.

Yeah, SaaS was a bit of a slow burn, right? Salesforce kind of started the whole movement and then kind of category by category, people created SaaS companies, but it was kind of a slow replacement category, certainly by contrast to enterprise AI. And by the way, a lot of the AI numbers you see out there are not enterprise only. So I like this report because it is focused on enterprise AI in specific. But yeah, it surprised me a little bit to see the SaaS equivalent was 16 years to get to the same number. I might have guessed 5 or 10. But, but in any case, yeah, this, this growth, and this is why the VCs are all going insane, this growth is incredibly fast.

And then once again, I haven't seen this. The next page, they have this chart and it breaks it up into 2 slides. On the left-hand side, it shows infrastructure GenAI spend equals about $18 billion. On the right-hand side, it shows application GenAI spend, $19 billion, but it breaks it down between foundation models, model training, in AI infrastructure on the infrastructure side, and then horizontal AI versus departmental AI versus vertical AI on the right-hand side. And what it shows is, not surprisingly, that foundation models by far the biggest percent of infrastructure spend, $12.5 billion out of the $19 billion. But more interestingly, that foundation model spend in enterprise grew from $3.5 billion to $12.5 billion. So almost 3.7x increase year over year, Dave. So I don't know what jumped out at you that chart, but that one really jumped out at me.

Yeah, I mean, it's definitely the biggest line change on the chart there on bottom of page 2. But to me, yeah, I like the taxonomy. I think it's nice how they split it between infrastructure and apps and then break apps into horizontal, departmental, and vertical.

Yeah, and even on that horizontal AI versus departmental AI on the right side of the chart, Horizontal AI was $8.4 billion out of the $19 billion spent on applications, and $7.3 billion was on departmental. And what's going to be really interesting to see which department is consuming the majority of that $7.3 billion spend for departmental AI.

Okay, well, let's move in, let's move in that direction, right?

Okay, well, something we talked about, and this was over a year ago, and I was shocked, and it was from a Battery Ventures report, and they said about 33% of AI software is going to be bought from existing software vendors, their current SaaS vendors. About 33% was going to be bought from new native AI application companies, and about 33% was going to be built internally. So on this page, I believe it's page 5, what Menlo says is that last year, 2024, 53% of enterprise AI solutions were purchased. They didn't say who they were purchased from. So that means 47% were built internally. But in 2025, 76% of enterprise AI solutions were purchased and only 24% were built. And that kind of blew me away, Dave, as far as people are moving to a buy versus build.

Yeah, I'm a little skeptical of all the data that's doing the buy-build analysis because of the following. Jeffrey Moore has this notion of context and core, and core is what makes you special or different. And context is that, you know, you run the business, you need an office, you need a cafeteria, you need a coffee pot, you need an accounting department, you need a finance department, you need a legal department, right? All that's the context. It's a huge amount of work in a business is spent on context, but core is what makes you special. Like, so for Amazon, it might be distribution, warehousing, inventory, automated picking and packing, right? Uh, for Costco, it's going to be all the the various efficiencies they drive into the business model. So, and, and I think the general rule of thumb is you buy context, you build core, um, because you can't buy core, right? Whereas context is, well, everybody's got a legal department, everyone's got a finance department, everybody's got an HR department. We all have employees and performance reviews, context, context, context, context, but our inventory management system is what makes us special, right? Or our, if we're selling an HR package, our skills management AI detection is what makes us special. So I think you're mixing variables here, which is if you went to a company and said, what would be your first projects with AI? I would probably say core, cuz I see a new potential source of competitive advantage coming down the pike and I want to try and grab that now before my competitors do. And I'm gonna have to build it. It's got nothing whatsoever to do with my preference for buying or building AI applications. And it has everything to do with my AI strategy. which is what I'm trying to do. So I think we're kind of crossing the beams here, if you will. And it's, it's making the data confusing because I don't think business people think about, do I want to buy or want to build? I think they think about what they're trying to do and is it available? Does it make sense to buy it or no, this is something that only I can build because I'm trying to get core competitive advantage.

Yeah, it's also, I think, one of the underlying variables when people talk about Vertical AI is going to be a big winner. It's that subject matter expertise and that core knowledge of a business process and regulatory environments and compliance requirements that could only be in a vertical AI versus a horizontal platform. So I do think that core is a great way to look at it, Dave.

Yeah. The only thing I'd say there, Ray, is even in vertical SaaS, like knowledge of pharma regulations, I would argue in pharma is context. We all have the same regulations, right? It is definitely a source of vertical application differentiation from horizontal apps. Hey, use our CRM because it understands pharma. Got it. But once you're selling to a bunch of pharma companies, then it's like, is that a core activity? Like, a core might be accelerating our clinical trials so we can get new drugs out faster than the other guys, but contacts might be regulatory compliance. So I hear what you're saying, but just It's all relative to where you sit.

Yeah, I understand. Well, let's move on to the next topic, which based upon the episode last week when we covered the famous, or maybe it's infamous, MIT Nanda Report, where 95% of AI projects fail or don't deliver ROI. I believe this was page—

Right. Don't repeat the myth. Every time you repeat it, you're reinforcing it.

I was just trying to pitch last week's episode, Dave.

I know, that was good. Okay. But anyway, yeah, we talked about the Nanda report. Keep going.

So on page 6 of this report, it basically highlights that AI software pilots convert almost 2 times more than traditional software pilots. And what they have here is they have 2 funnels. They have AI buyers funnel on the left and they have software buyers on the right. They look at, okay, so we're starting at do people have a use case for AI software? So 100% who have identified the use case. 73% of AI buyers say they have identified vendors. 58% then piloted from that vendor they selected. And 47% actually end up in a closed one. Doesn't say production, but it says closed one. below the pilot. So I'm making a poor assumption maybe that close won after pilot means it's being put into production. And in traditional software for companies who went into pilot, only 25% went into close won. So it basically says that AI software buyers are converting to production almost 2x.

What say you? And I, I, yes and no, and mostly no. But, but, but I think this chart says they're converting, they become deals at 2x the rate. Right? Like, this is a deal chart because the last step is closed won. It's not in production, right? And you made a massive leap in my mind to run a proper pilot. You actually have often have to buy the software first. You can do kind of like a POC or some sort of kind of deal-driven pilot. But to me, in any real enterprise project, that's step one. You buy the software, you run, you run in production, but for one division, for one product line, for one factory.

Right.

Right. For one teeny piece of the business, you run an experiment that is going to be in production, but in one little place. And if that works, then you roll it out. Right. So, so this chart very much is a sales funnel. So I think it's saying there are fewer tire kickers in AI than software. And I think that's true. I think— why? Because I think people are under a lot of top-down pressure to go try AI. So when the boss says, Ray, go try to use AI in our marketing department, You go out and evaluate 10 things, you're going to buy 5 of them because you want to go back to the boss and say, yeah, we're using AI. Whether those work is a completely different thing, right? Whether that was a closed one advanced directly to production across all of my marketing team, or whether that's, yeah, in the, you know, in Australia we're going to run an experiment with automated outbound, right? I don't know. But either way, I mean, it'll end up being a deal.

Yeah, the, the only thing I would, I would respond to that is there was some lack of clarity maybe in the report itself, because under their, the same page where it says AI buyers convert at a higher rate, the specific language says, we found that once an organization commits to exploring an AI solution, deals convert at nearly twice the rate of traditional software, semicolon, set 47% of AI deals go to production. Compared to 25%. So they use the word production, but you're right, Dave, the chart says closed won.

Yeah. So again, like, I haven't really tried to dissect this thing like we did in the Nanda report. Because fundamentally, I'm not suspicious on this one. It seems to check out. I don't see that. I just think it's a sloppiness in my mind. I think it's kind of assuming that every closed won deal goes to production. But I could be wrong. And there's a little footnote down there. Um, but I don't think it's relevant. I think it's for another point. So, so yeah, I, I'm, I'm, I'm going to take the chart at face value. And to me, it's at the bottom of that funnel is deal, not production app.

Okay. Well, here's the thing that I really liked about this report is number one, it actually showed AI coding software, which they call part of the departmental spend and AI coding software. Was $4 billion out of the $7.3 billion that they're calling departmental AI. So it's 55% of departmental AI spend, which I thought it would be sales and marketing, but it's coding. And it also said that 50% of developers in this population was using AI coding, and at the top quartile of companies, 65%. And Menlo took a little liberty here with a term. They say that AI coding software is the first AI killer app.

Yeah. So, so look, I think there was no surprise on that, that, that, that clearly just looking at the growth rates and look at the hype around people like Replit and Lovable, that AI vibe coding software, quote unquote, is super popular category and it's growing like a weed and blah, blah, blah. That, that was no big surprise. Call it the first killer app. Sure. Uh, agreed. Stipulated. The thing I found more interesting than that even, Ray, was on page 6. And this one surprised me if I understood it correctly, but where they're showing kind of new dollars captured startups versus incumbents. And that's very much not vibe coding, right? It's market research. 100% is going to native AI startups. Finance and operations, 91 cents on the dollar going to startups versus incumbents, right? Sales, 78%. Product and engineering, 71%, which is probably where I would put the vibe coding. HR, 60%. Marketing, 46%. So this section just above that one, I think it is in the report where they're talking about who's getting the disproportionate share of new dollars. This was very optimistic for me because sometimes the first killer app is the only killer app. You know, if Vibe Coding is the killer app, will there necessarily be another one? Uh, and, and this is suggesting, which is what I believe for what it's worth, that no, there's going to be lots and lots of killer apps across the enterprise.

Yeah, it's interesting. On page 9, they show the departmental AI market map, and in that product and engineering, the only incumbent they have is GitHub Copilot, and then Cursor, Claude Code, Replit, Lovable, And about 12 others are in the startup. So it's interesting that they consider GitHub Copilot as the incumbent.

Yeah, yeah, yeah. I can't find that chart. Where was it, Ray? On page 9? Oh, there it is up there, top left, top left of the bottom chart. Yeah, yeah. I don't know how, because to me they're kind of, uh, in the middle there. They're not, they're not grandma and grandpa like Salesforce or, you know, maybe Gong or Workday, right? They're, they're kind of mid-generation, not, not old generation. But I mean, it's a Copilot too, so it's actually even newer generation. Yeah, that's surprising.

Boy, even though there's, as you mentioned, a little Twitter buzz around some of the numbers here, it'd have been really interesting to see market share between GitHub Copilot and all those startups, you know, vendor by vendor. But I digress a little bit because the next thing I wanted to talk about was, I believe it's page 10, and it shows the vertical AI market, which is totaling $3.5 billion. So The smallest of the application segments. But healthcare, Dave, healthcare was $1.5 billion of that $3.5 billion spend. So I believe that's around 43%. The next closest industry was legal at $680 million. And as you know, my wife works in the healthcare industry and they have always been a laggard in how much they spend on IT. Not MRI machines and X-ray, the capital equipment, but pure IT infrastructure and software. So it's really interesting to see healthcare is by far the number one adopter of vertical AI. And over 50% of that is basically transcriptions where the doctor is taking notes while he's with the patient. Everything gets transcribed and ingested into the medical records.

Yeah, yeah. Not surprising, to be honest, given the importance of that field. And how everything is taken down and noted and how good AI is at that and how the speech itself needs optimization. By the way, I've worked in Europe with a company that does multilingual healthcare AI transcription. So you need to know in all these different languages exactly how doctors say things and the synonyms they use. It's actually a difficult problem. So that makes sense to me. The thing I'd say here on the vertical spend, the second vertical they have is legal. They cite a company called Eve, which really is for law firms. But sometimes I worry that in legal people confuse departmental with vertical, i.e., if it's, you know, there are lots of good departmental legal apps that help regular companies, not law firms. In this case, I think they are separating it because Eve very much is for law firms. But I think it's another pretty obvious area for both departmental and vertical AI, and it's the number 2 in their chart.

All right. Okay. We're going to wrap up with one last chart and you know what, it's page number 13 and maybe that's a little bit unlucky, but Menlo wanted to share the enterprise large language model API market share. And market share is defined by usage. And basically what it shows is that 3 years ago, OpenAI, of course, was the market leader with 50%. They've decreased to 27% in 2025. And Anthropic, which has clearly said they want to be a B2B enterprise AI vendor, they went from 12% up to 40% of market share. So much higher than, um, OpenAI. And Google And they didn't say whether this was due to Gemini 3, but they went from 7% market share of API usage to 21%. And I know that there's a little bit of buzz going on out there about this.

Yeah. So I think this is where we risk stepping into the middle of a controversy because obviously, I think we said it already, but Menlo is an investor in Anthropic and presumably very much behind the kind of come-from-behind victory tale that Anthropic is telling in the enterprise, which I personally tend to believe. Whereas Khosla is an investor in OpenAI and probably wanting to emphasize current market share and say it's not all enterprise, yada yada. But in any case, there's an active controversy going around about this very set of charts here at the top of page 13. That is the clip that's being shared on Twitter. So I think it's interesting data. I would take it with a grain of salt because people producing it have skin in the game. But, but I don't know the story I'm seeing. I know, Ray, I can't remember if you mentioned it here or the pre-call, but anecdotally, you seem to see Anthropic more in the enterprise too.

Yeah, it was during our Next Generation Solutions Showcase. We had 16 office finance platforms that had AI core to their product, and 14 of the 16, when they showed their generative AI functionality, it specifically said we're using Anthropic and we're leveraging the MCP protocol.

Cool. Okay.

Okay. Well, let's wrap up here. In summary, you know, some Captain Obvious moments, but I always like to see data behind my intuition and all the LinkedIn chatter. AI software, fastest growing software category of all times, hitting $37 billion last year. That enterprises appear to be moving more towards the buy versus build. And we talked a lot about that. That AI coding is definitely that first killer app. And that AI software pilots appear to be converting almost 2x more than traditional software pilots, which is good news from my perspective. And we got a little bit about that market share dynamic going on. Overall, I think it was a very good report from the type of data that I haven't seen before.

Yeah. And it's a, it's a 3rd year of them running this. And in the prep for the episode, Ray, I didn't tell you this, but I accidentally opened the 2024 version and was reviewing that. Um, and it took me about 15 minutes to realize I was in the wrong year. So it is fun to watch how it's evolved over time. Uh, so yeah, overall I like this report. I think it's worth reading. And, uh, you know, I think the build versus buy one, Ray, is probably the one where I think there's the most confusion. Because to me, I got to know what kind of system it is to know what build versus buy means. Because if it's a core system, I'm going to want to build it almost definitionally. If it's a contextual system, I should want to buy it. And maybe I get most interested with like, okay, this is just context, but you want to buy it? Talk to me more. I need to understand why you would do that.

Okay, well, that's the end of today's episode. This is the State of Generative AI in Enterprise 2025 from Menlo Ventures. Worth the read. Thank you, Dave.

Thanks, Ray. The Metrics Brothers AI, SaaS, and More is a production of the Metrics Brothers Growth and CAC and a member of the Benchmarkit Podcast Network. By accessing this podcast, you acknowledge that the Metrics Brothers make no warranty, guarantee, or representation as to the accuracy or sufficiency of the information provided or the humor content of the jokes told. Ray? Information, opinions, and recommendations presented are, according to our spouses, probably wrong and provided for information purposes only. This podcast should not be considered professional or, for that matter, unprofessional advice. We disclaim any and all liability for any direct, indirect, indirect, misdirect, incidental, special, ordinary, consequential, inconsequential, or other damages arising out of any use of, or God help you, reliance upon the information presented here. Ray Grothreich is based in New York City and available on X @RayReich. Dave Kellogg is based in Silicon Valley and available on X @Kelblog. Schedectady, which is French for unspellable, is not our actual production location. Thanks for listening.