16 Ways Enterprise AI is Changing

0:00Today on the AI Daily Brief, 16 trends in enterprise AI. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:33like this. This episode got a little long, so we just have a main episode. We will be back to our normal headlines plus main format tomorrow, but for now, let's dive in. Welcome back to the AI Daily Brief. Today, we are doing one of my favorite things on this show, which is when someone comes out with a set of predictions or observations, we get to then go through them and talk about if and how what we're seeing both at the AI Daily Brief and at Superintelligent resonate with those trends and observations, or if we're seeing something different. Today's source for our show comes from Andreessen Horowitz, who recently posted their 16 changes to AI in the Enterprise 2025 edition.

1:09Now, this is the second year in a row that they've done this piece. And back in 2024, there were a few big trends. First of all, they saw budgets for generative AI skyrocketing. Second, enterprises had a lot of concerns around enterprise data security. Leaders from enterprises they found weren't building models from scratch, but instead customizing models through fine-tuning. Purchasing decisions were being heavily influenced by the existing cloud providers the company was working with, and enterprises citing a dearth of existing high-quality vertical applications were building tools rather than buying.

1:40Now, what you might notice wasn't there at all was, of course, agents. In March of 2024, yes, there was conversation about agents, but we were still pre-reasoning models. The big coding agent companies didn't exist yet, or at least weren't doing this. And so just based on that, of course, that represented a very different time. So what have they found this year? Because this is the contemporary piece, we're actually going to go through bullet by bullet. Starting with their first, budgets are bigger than expected with no signs of slowing down. In fact, A16Z found that enterprise leaders expect an average of 75 % growth over the next year in their AI spend.

2:16Positively, the growth in spend seems driven by the discovery of and implementation of new use cases. As one company said, we've been mostly focused on internal use cases so far, but this year we're focused on customer-facing gen AI where spend will be significantly larger. I.e., it's not like they sat down and said, we're going to spend a bunch more money. It's just that the use cases demanded it. Now, the one thing that I would add to this is that it's very clear that agents are having an outsized impact on this as well. You might remember this slide from my presentation on why your company needs to move faster on AI.

2:45It's from PwC, and it's about the AI budget increases that are due to agentic AI. Basically, they found that like 88 % of organizations were increasing their AI budget due to agents, and about three-quarters of them were increasing their budget by 10 % or more. The second trend that A16Z points to is the fact that this spend is coming from different places. Here's a really important stat. Last year, innovation budgets represented around a quarter of Gen AI spending. That is now down to just 7%. Reallocated central IT budget jumped from 28 % last year to 39 % this year, so a big jump. Business unit budget went from 21 % last year to 27%.

3:23Net new central IT actually was down a little from 19 % to 17 % of spend. And central R &D was up from 8 % to 10%. Still, the point that they take away is that Gen AI spend is graduating to permanent budget lines, and I think that's a good way of putting it. This also sort of reflects something that I've been noticing a lot, which is that the intention around agent pilots is very different than it was around other Gen AI pilots that we saw maybe a year or a year and a half ago. A lot of these pilots still had the feeling and flair of if. Meanwhile, all of these agent pilots are really just about figuring out what's possible now and building infrastructure for what is presumed will be available in the future.

3:57I think that because organizations are in the mindset of redesigning themselves around agents, it's natural to see the budgets shift consequently to different areas. Number three is an interesting one because it sees enterprises behaving more like consumers. If you are a consumer of AI, especially if you're someone who's enfranchised enough to be listening to this show, you're probably the type of person who maybe uses different models for different use cases. You might use Anthropics Quad for coding, for example, but use GPT 4.5 for writing and O3 for business strategy tasks. A16Z finds that enterprises are also using multiple models and that model differentiation is the key driver.

4:35I think the thing that comes through here is that organizations are just getting a lot more sophisticated as they put more reps into these actual tools. A16Z writes, it's well known, for instance, that Anthropics models excel in coding related tasks, but there's more nuance to this claim. Within coding, some users report that Claude performs better for fine-grained code completion, while Gemini is stronger in higher level systems design and architecture. A reason for that, by the way, might be just the difference in context window and Gemini being able to interact with a bigger code base all at once.

5:03Still, this is a really big and important observation. I think a lot of people might have assumed that enterprises would just be locked into whatever one tool they had based on central IT. And instead, A16Z found the number of respondents who are using five or more models is all the way up to 37%. Number four in this trend list, while there are lots of models, there is definitely some clear consolidation and leadership around a handful. Specifically, A16Z points out that OpenAI continues to be in the lead with overall market share, but both Google and Anthropic made big strides over the last year.

5:34Interestingly, one of the things that they found around OpenAI's models is that in addition to companies using GPT-40 and O3, 67 % of OpenAI users have also deployed non-frontier models in production, which is not the case for Google and Anthropic. That number is 41 % for Google and just 27 % for Anthropic. So basically, the companies that are using Google or Anthropic are much more likely to be concentrated just at their highest end offering. When it comes to growth, and this one maybe won't surprise you, Google has gained more of a share within bigger enterprises. Interestingly, A16Z points to not just the fact that big enterprises are going to be inherently more trustful if you think about compliance and legal departments of a big company like Google, but also that Google has made a major play around its performance to cost ratio.

6:16And given how big the use cases are for large enterprises, that cost consideration really seems to matter. Large enterprises are also more likely to adopt open source models like Llama and Mistral, which I think has a lot to do with the trend towards building, which we'll talk about in a little bit. Overall, price is really interesting. One of the things that's happened is that closed-source models, especially non-frontier models, have come down so precipitously in cost that customers are more frequently opting for those closed-source models, given that they still get other ecosystem benefits as opposed to moving outside to open-source.

6:47One customer said, the pricing has gotten appealing and we're already embedded with Google. We use everything from G Suite to databases and their enterprise expertise is attractive. Whereas another company put it more simply, Gemini is cheap. For those who are interested in the long-term trajectory of open source versus closed source, the fact that closed source is seeing costs come down so precipitously is an interesting factor in how that battle might shake out. One big change from last year comes in trend number six, which will surprise no one who's actually paying close attention to enterprises, but is a big difference from last year, which is that fine-tuning is viewed as less necessary as model capabilities improve.

7:22And this is really simple. Basically, instead of taking a whole bunch of training data and taking the time to fine-tune models on that data, now that context windows are really long and the models are smarter, you can just use what's off the shelf. One enterprise said, instead of taking the training data and parameter-efficient fine-tuning, you just dump it into a long context and get almost equivalent results. Given how many startups were positioning themselves as helping enterprises fine-tune models, this is obviously a shift with a lot of financial implications. Trend number seven is all about reasoning models.

7:53And the TLDR is that while enterprises are still early in their testing, they're pretty excited about their potential. Basically, enterprises are looking at these reasoning models, and frankly, I think the agents that they represent, as opening up a new set of use cases. 23 % of enterprises right now are using OpenAI's O3 model in production. 57 % of A16Z respondents said that reasoning models are slightly accelerating their adoption. The thing that we see all the time at Superintelligent is that every single model improvement opens up in some way some new use case. In other words, model improvements are not just about doing the things that you're already doing a little bit better, a little bit faster, or a little bit cheaper.

8:30It's about doing things that weren't possible before. As one executive said, reasoning models allow us to solve newer, more complex use cases, so I anticipate a big jump in our usage. What about why enterprises are choosing different models? A16Z observes that the buying process is increasingly similar to traditional enterprise software procurement, complete with quote checklists and price sensitivity. The breakout chart here is really interesting where we see changes in how important a particular consideration was in the buying decision. The areas that went down the most as a major consideration were reasoning inaccuracy and reliability, whereas the area that went up the most is cost of ownership.

9:08And I think this reflects back to what we just saw before around the idea that this is moving into more long-term budgets. This is no longer just about innovation. The usage of these tools is getting more ubiquitous across the organization, and so cost comes back as a major consideration in a way that it just didn't have to be in the innovation stage. Trend number nine is a little in the weeds, but I still think reflects a fairly significant change. A16Z characterizes it as hosting preferences still vary widely, though enterprises have quickly built trust for model providers over the last year.

9:37Basically, the TLDR on this one is that companies have gotten a lot more comfortable hosting directly with a model provider, even if it's a newer company like OpenAI and Anthropic, as opposed to last year when most organizations were opting to access models through a cloud provider that they already trusted and already had a relationship wherever was possible. The biggest reason for this shift is that leaders, quote, want direct access to the latest model with the best performance as soon as it's available. Early access previews are important too. One of the things that we see as a flagged problem from enterprises all the time is the disparity between the quality of the models that they have access to in their personal lives and in their consumer accounts as opposed to what they have access to at work.

10:15This is increasingly becoming an issue. It looks, for example, like Amazon might actually ditch its internally produced AI code companion and instead just let people use Cursor because that's what employees are demanding. We hammer on enterprises all the time that this is potentially an area where they have the most to gain because they have so much control. The closer that they can get to just getting state-of-the-art models rather than being stuck with whatever's available through their standard cloud provider, the more effective the use cases that the deploy are going to be. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context.

10:53Blitzy is used alongside your favorite coding copilot as your batch software development platform for the enterprise seeking dramatic development acceleration on large-scale codebases. While traditional copilots help with line-by-line completions, Blitzy works ahead of the IDE by first documenting your entire codebase, then deploying over 3 ,000 coordinated AI agents in parallel to batch build millions of lines of high-quality code. The scale difference is staggering. Copilots might give you a few hundred lines of code in seconds, but Blitzy can generate up to 3 million lines of thoroughly vetted code.

11:24If your enterprise is looking to accelerate software development, contact us at Blitzy.com to book a custom demo or press Get Started to begin using the product right away. Today's episode is brought to you by Plum. If you build agentic workflows for clients or colleagues, you need to check out Plum. Plum is the only AI-native workflow builder on the market designed specifically for automation consultants, with all the features you need to create, deploy, manage, and monetize complex automations. Features like one-click updates that reach all your subscribers, user-level variables for personalization, and the ability to protect your prompts and workflow IP.

11:59Make your life easier, your clients happier, and your business thrive with Plum. Sign up today at useplum.com. That's Plum with a B forward slash NLW.

12:33Vanta makes it easy and faster by automating compliance across 35-plus frameworks. It gets you audit-ready in weeks instead of months and saves you up to 85 % of associated costs. In fact, a recent IDC white paper found that Vanta customers achieve$535 ,000 per year in benefits, and the platform pays for itself in just three months. The proof is in the numbers. More than 10 ,000 global companies trust Vanta. For a limited time, listeners get$1 ,000 off at vanta.com slash nlw. That's V-A-N-T-A dot com slash N-L-W for$1 ,000 off. Today's episode is brought to you by Agency, an open source collective for interagent collaboration.

13:12Agents are, of course, the most important theme of the moment right now, not only on this show, but I think for businesses everywhere. And part of that is the expanded scope of what agents are starting to be able to do. While single agents can handle specific tasks, the real power comes when specialized agents collaborate to solve complex problems. However, right now there is no standardized infrastructure for these agents to discover, communicate with, and work alongside one another. That's where Agency, spelled A-G-N-T-C-Y, comes in. Agency is an open source collective building the Internet of Agents, a global collaboration layer where AI agents can work together.

13:50It will connect systems across vendors and frameworks, solving the biggest problems of discovery, interoperability, and scalability for enterprises. With contributors like Cisco, CrewAI, LangChain, and MongoDB, Agency is breaking down silos and building the future of interoperable AI. Shape the future of enterprise innovation. Visit agency.org to explore use cases now. That's A-G-N-T-C-Y dot org. Now, one really interesting conversation that comes up pretty frequently around AI is where moats are going to exist in the future. If technology gets commoditized, and even state-of-the-art models everyone catches up to in a matter of weeks, where is lock-in going to come from?

14:31And one area that appears to be on the rise is switching costs. Turns out that switching costs are getting higher for AI as agents turn on more complex use cases. Last year, we found that most enterprises were designing their applications to minimize switching costs and make models as interchangeable as possible. As a result, many enterprises treated models as easy come, easy go. That might have worked well for simple one-shot use cases, but the rise of agentic workflows has started making it more difficult to switch between models. Now, this makes a ton of sense. Every once in a while, we'll try a different model inside our process, either for the voice agent in our agent readiness audits or at some part of the LLM sequence by which we process the results of those voice agent interviews.

15:14The changes when we do that can be dramatic and in not always positive ways. Andresen points out that agentic workflows often require multiple steps to complete a task, so changing one part of a model's workflow could impact all downstream dependencies. You can see this show up in competition with the agent platforms. Every agent platform is pushing some idea of wanting to get a bigger set of the agent use cases for their platform because companies are going to want agent interoperability. Now, I tend to think that even with high switching costs, the market is going to totally force interoperability on the agent platforms.

15:47But you can definitely see that this is a consideration that enterprises have as they invest in a particular agentic ecosystem. Number 11 is a really interesting one. This is about how enterprises evaluate different models. And basically, A16Z found a big rise in using external benchmarks as a proxy. Internal benchmarks actually went down a little bit and project-by-project benchmarks went down a little bit as well. I understand why they found this, but I think that this might be a temporary thing. The problem is I often talk about on this show is that most of these external benchmarks are really washed.

16:22We are crowding in around the very top ends of performance. And frankly, we just need a different, more sophisticated set of evaluations. If you take the AI Engineer Summit tracks as a leading indicator of where the market is heading, I think you're gonna see a ton of work over the next year around evals and model evaluation approaches that are going to find their way to the enterprise as well. So one prediction that I have is between May 25 and May 26, if they do this survey again, I would anticipate a big jump up and a return to internal benchmarks as a major consideration in how they evaluate model performance.

16:55Next up, we get into the build versus buy paradigm, and this is a fascinating question. So the short of what A16Z found is that as compared to last year, where you remember from the beginning of this show, a lot of companies were building their solutions by default because there weren't good AI applications for their particular use case. That's shifting now. A16Z writes, We've seen a marked shift towards buying third-party applications over the last 12 months as the ecosystem of AI apps has started to mature. It is absolutely the case that vertical agents and functional agents have been the biggest startup trend of the last six months at least.

17:32And because of that, more and more of the use cases that are common across lots of different types of enterprises are starting to have high-quality off-the-shelf-ish options. I say off-the-shelf-ish because when it comes to agents, nothing is as off-the-shelf as previous sort of enterprise software that we've seen. There's still going to be some amount of customization and a whole bunch of work to wire together the data that makes these agents work. So even the build versus buy paradigm as a conceit is a little bit challenged by how AI works. And to some extent, this is just a natural evolution of software.

18:02Enterprises did a bunch of building because they didn't see good options in the market for the applications that they wanted. But then when those applications come online, external companies who just focus on that thing are going to be in general in a much better position than the enterprises to maintain and update and improve these specific functional and vertical tools as opposed to what internal developers can do. And yet, I think that actually we're going to see a bit of a bifurcation. I become increasingly convinced, based on the huge array of conversations that we have at Superintelligent, that we're going to see organizations and maybe even different parts of organizations fall into two distinct camps.

18:37One camp is going to be where the off-the-shelf-ish solutions are good enough. So think CPG companies with customer service agents. A lot of the use cases of random CPG X and random CPG Y are going to be so similar that you really can have a high-quality off-the-shelf-ish option that just needs a little bit of customization and wiring in with the data of the particular company. However, I also think that there are going to be certain areas and industries, particularly heavily regulated, high-value industries, think finance, healthcare, where the default is going to be forever to roll your own agentic solutions.

19:14I think this is going to be made more viable by the rise of agents that can build other agents, which is the type of capability we're seeing from companies like Emergence. And those companies are going to be focused on building the infrastructure to allow their teams to build and share and provision agents dynamically across their organizations. There are going to be two almost totally different types of ecosystems and buying behavior around these different approaches, with lots of opportunities for companies on all sides. You can definitely see a little bit of the split in how regulated versus non-regulated industries are interacting with different use cases.

19:49In non-regulated industries, for example, it's fairly evenly split between organizations that are using a third-party app versus a custom solution for data analysis, whereas in regulated industries, the vast preponderants are using a customized solution as opposed to just like 20 % who are using a third-party app. What about pricing? This is another thing that we frequently keep an eye on here at this show. And basically what I've said before is that I think that we are in the experimental phase, where it's pretty clear that the traditional per-seat model isn't going to hold, but that also, a lot of these first attempts at outcome-based pricing are not necessarily going to get it right out of the gate either.

20:25And that's basically what Andreessen Horowitz found as well. Their trend 13 is that buyers are struggling with outcome-based pricing. CIOs are concerned around lack of clear outcomes, unpredictable costs, as well as attribution. Now, companies are still interested in usage-based or hybrid-style models for AI applications, and basically it feels like these outcome-based experiments are still just really nascent. Only 15 % of CIOs surveyed said that they preferred outcome-based models as compared to, for example, 21 % who said that they preferred seat-based and 39 % who said that they preferred usage-based models.

21:01When asked what their biggest challenges with outcome-based pricing, lack of clear measurable outcomes, was cited by 47 % of respondents, and unpredictable and unscalable costs were cited by 36%. Another trend is that we are starting to see certain use cases become absolutely ubiquitous and default. Specifically, A16Z calls out software development. They write, software development has seen a step change in adoption, driven by a perfect storm of extremely high quality off-the-shelf apps, a significant increase in model capabilities, relevance for a broad set of companies and industries, and a no-brainer ROI use case.

21:34The percentage of enterprises that had software development as an in-production use case went from less than 40 % to over 70 % in just one year. Some of the other use cases that saw big gains were enterprise search, data analysis, data labeling, and customer service actually saw a slight decline in in-production use. Trend 15 has to do with how different solutions and use cases are coming into the enterprise, and the short answer is that they're being driven by their employees' external usage. A16Z writes, much of the early growth across leading enterprise AI apps has been driven by the prosumer market.

22:08Many CIOs noted that their decision to purchase enterprise chat GPT was driven by, quote, employees loving ChatGPT. It's the brand name they knew. I talked before about how Amazon might be using Cursor in the future, so I think that this is just going to increase. Lastly, and this is a big one, one of the things that was extremely notable about AI when it started is how much incumbents had an advantage that they had not had with previous tech paradigm shifts. The pattern that we had gotten used to was that each new tech paradigm shift creates a new set of incumbents. who start as startups and move more quickly and nimbly and eventually outcompete and become the new dominant players.

22:48However, there were some really big factors that incumbents had when it came to AI that put them at the center of the industry much faster than they might otherwise have been. One of those was established trust, obviously a huge issue when it comes to sensitive data as a key part of the success rate of this new technology. Existing distribution was another advantage for incumbents. And then frankly, incumbents were some of the only funders who could actually play at the capital level needed by companies like OpenAI. The big, multi-billion dollar investments into companies like Anthropic and OpenAI by companies like Microsoft and Google was not driven necessarily by OpenAI and Anthropic's desire to work with their big tech competitors right out of the gate, but instead by the fact that they needed to raise so much money that those were basically the only games in town.