AI Agents Winning Hearts and Wallets, A New Report from G2

Live from Schenectady, New York and the South of France, 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 Metrics Brothers, and we go together like figs and brie.

Ah, a little homage to France with a touch of fromage, huh, Kek?

Oh my gosh! Oh God! A little homage to fromage, my Ray Ray Ray Ray. I guess your never-ending trip to Europe has you now in the south of France, Perpignan, I believe. And that was the perfect opportunity for another food analogy.

Yeah, but you know, we stopped with SaaS talk. Now we need to stop with the food talk, both figuratively and literally, and get back to something that is even more appealing. And to me, that is how AI agents are winning hearts and wallets.

Sounds like you're talking about the recently published report from G2 entitled A Leap of Trust: AI Agents Are Winning Hearts and Wallets. But before we dive into that, which is the topic of this week's episode, Let's hear a word from our sponsor.

The Metrics 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 @kellogg and Ray on LinkedIn @rayreich. Thank you so much for listening to the Metrics Brothers podcast.

So, Ray, I know it's 10 o'clock there in France. I think you're probably half in the bag. I'm not sure. He won't admit to this, folks. But I know you're super excited about this report, Ray. What is so exciting about this report to you?

Well, you know, it comes upon our last episode, which was all about ROI and why it's so important in AI software investments. And this report from G2, one of the first pages, it talked about why the investor enthusiasm for AI agents is palpable, because AI agents are actually moving beyond efficiency and productivity gains to true ROI. And I'm like, that's exactly what I've been thinking about, talking about. And it was research that backed up that concept.

Okay, so this thing starts— kind of comes as a stark contrast to the recent MIT study, quote unquote, the GenAI Divide, where they said stuff like, despite $30 billion in investment, 95% of projects have no ROI, yada yada. This is more of a ray of sunshine, and it's talking about how to use agents to successfully deploy AI and how people are today successfully deploying AI agents. within enterprises. Is that roughly right, Ray?

Exactly right. Where that MIT report, which has been fairly widely debunked, said that 95% of enterprise AI projects were seeing no ROI, this report highlights— and this is over 1,000 respondents— that 25% of them said that they're seeing and being able to prove meaningful outcomes in 3 months or less with agentic AI.

Yeah, so I think we'll finish up with the MIT report now, but because it got so out there and was so visible, you know, it's important to remember this is a survey of 1,000 people here from G2, I think 1,050. MIT study was interviews with something like 53 people done by a group called NANDA, which stands for the Networked Agents of Decentralized AI, a group that actually had something of an agenda. And a lot of people feel like they cherry-picked their interviews to kind of Confirmation bias their way into saying what they wanted. So the thing has been pretty widely debunked. The thing that bothered me about it, it was just such obvious clickbait. If you read the headline in the first paragraph, it's just designed to be consumed by the media. It wasn't kind of any kind of serious academic study. But in contrast to this, we have G2 out with a big survey of a lot of people, and they're finding much better news about AI. Ray, tell us about this survey.

Yeah, well, it results in this 57-page report. And as you said, it was over 1,000 B2B enterprise software buyers, and it covered a wide array of topics from how AI agent adoption is expanding, that the results are actually rolling in, that these agents are gaining more trust and autonomy, that the orchestration of the agents across the entire process is critical. And then, as I mentioned earlier, that agent strategy is driving outcomes that have real tangible ROI, and that these vendors need to adapt and adopt these end-to-end strategies to win. So I really liked it, and 57 pages, and I picked out 7 or 8 key topics that I think the audience is really gonna love, Dave.

Awesome, Ray. Before we do that, let's just quickly get you to define AI agents, either using your or their definition. Preferably theirs. What is an AI agent, right?

Yeah. So they define it as AI agents are systems that can reason, plan, use other technology and tools, and most of all, act. So it doesn't require human beings to do the action. And unlike chat-based generative AI, which a lot of companies have put that chatbot using generative AI on their website, which that often stays in the background. Agents are the direct interface to the customer and they're talking to the real world. They coordinate with suppliers and they move work through a system versus just a point of interaction on a website.

Okay, that's awesome, Ray. Tell me a little bit about translating efficiency gains into outcomes.

Yeah, well, this was really on page 3 of the report and it's where, you know, the first paragraphs talked about how Fully autonomous agents can remove a primary production constraint, and that is humans managing the loop. One of the things we talked about all the time is not only in training models you need that human feedback, human reinforcement, but you need humans to actually often intervene into the agentic process and actually take over and interact with the humans. And with true agent AI, Humans aren't needed in the loop. They can do kind of review and audits of the process the agents are actually executing, but humans don't need to be in the loop full-time. So that was one of the big things that came out. The other thing that came out was this topic of Parkinson's Law and the relevance to AI agents. Can I talk a little bit about that?

Yeah, well, let me do that, Ray. I'll tee up Parkinson's Law and I'll have you do the relevance to agents. So What was it? Last episode, we talked about Jevons' paradox, which came from an 1860s economist. Now, we've got Parkinson's law, which comes from a naval historian, C. Northcote Parkinson, and an article he published in 1955 in The Economist, which I didn't know The Economist was actually existing in 1955, where he coined Parkinson's law, which roughly says, work expands so as to fill the time available for its completion. So, and I think we've all seen this, Ray, that if I took something that took 4 hours and then all of a sudden I can make it take 3, what's going to happen to that other hour? Is it going to get used doing something else unproductive, or will it translate directly into an efficiency? So, how does this relate to AI agents?

Well, so when you're using AI to make a human more efficient, to your point, let's say I can do things 25% faster. So it frees up those 2 hours. And I think in the previous episode we talked about, are people eating bonbons and playing video games? Are they doing more work? Right? And Parkinson's Law says that most humans will fill that time with nonproductive activity. And with agents, they're going to fill that 2 or 3 hours that are freed up with more work. So over time, as the agents drive more and more productivity, they can do more and more work, unlike humans. So that's what's going to drive continuously growing productivity gains over time versus just giving the person a higher quality of life and maybe a less stressful role.

Yeah, I'm not sure I'm as pessimistic on humanity as you are there. But certainly, at least some of the potential productivity gains do get lost. I would agree with that. And certainly in a world of agents, they don't. It just gets cracking on the next task. So interesting law, yet another economics law here on the Metrics Brothers. Let's talk now about which industries are moving to AI per this survey, Ray.

Now, this actually surprised me because when it looks at the industries leading the way in AI agent investments, so it measures by how much money they're spending per year. Healthcare. Healthcare was leading the way with 23% of healthcare participants reporting that they were investing more than $1 million in agent investments, and another 24% were investing more than $5 million. So if you add that together, 47% of the healthcare participants said they were spending more than $1 million per year in And on AI agents. And then retail, retail was the second with 33% investing more than $1 million and another 11% investing more than $5 million. So 44% of retail participants were spending more than $1 million a year with AI agents. And honestly, Dave, I didn't know that it was that mature already.

Yeah, I mean, the issue here I have, Ray, is first, I get what G2 is doing. They want to show where people have big absolute dollar budgets because they work with vendors and vendors want to find where the money is. So that all makes sense. Just analytically, I don't love this methodology because the healthcare companies might be bigger than the retail companies, right? Well, so how much is $5 million to a $5 billion company versus $5 million to a $200 million company? And the answer is, if you wanted to do this more correctly, I would do it in percent of sales basis, right, or even a percent of IT expenditures, but they're just doing it in absolute dollar size on the budget, and that is useful if you're a vendor and you want to know what verticals to target. There's a lot of $5 billion budgets kicking around in healthcare.

Yeah, and that was on page 9, and then they went a little bit further, Dave, on page 10, and, you know, we've heard that there's a lot of experimental budgets, proof of concepts, and pilots going on out there, but they're not converting into full production. But in this particular research, it said that 57% of these 1,000+ respondents actually have systemic AI agents in production, not in pilot, actually in production. And another 22% have AI agents working in a pilot. So basically saying almost 80% are using agentic AI either in production 57% or in pilot. And this is the other thing, Dave, that is unlike some other research I've seen. 50% of agentic AI pilots actually are converting into production, which means to me, even though it didn't say in the report, the customers are seeing enough benefit and value to move from a proof of concept or pilot into production.

Yeah, I think one of the more interesting observations in the report is on page 11. Where it says, quote, agent adoption looks very different from the GenAI spending wave of 2023, 2024. During that period, companies rushed into enterprise-wide seat licenses, carved out budgets for API calls, even spun up their own models. They went searching for proofs of concept that might show value only after the purchase. And I think they're saying that that era is now behind us. We're kind of in the era of early production and POCs and pilots that we want to put into production.

Yep, totally agree.

Okay, so how about use cases? So we talked a little bit about industries. Let's now kind of intersect industries and use cases.

Yeah, now there's some things here that didn't surprise me, and I have one kind of, I guess, pet peeve about the report. And I'll go with what didn't surprise me was that customer support— customer support, customer service— was the number one use case across all 1,000+ respondents. 28% of respondents said they're using agentic AI for customer support. Then 26% of that population said they're using agentic AI for process automation. But to me, process automation or automating workflows, that's kind of horizontal and a little nebulous to me 'cause I'm not understanding or sure what processes are being automated. Are they supply chain processes? Are they sales, customer acquisition processes? So that did kind of leave me a little hanging out there to say, what processes are being automated? I don't know what you thought about that, Dave.

So I think I'm going to back up a slide on you. Ray was looking at slide 13. I'm going to back up to slide 12, which kind of crosses use cases by industry. And I find 2 interesting things here. One, as Ray pointed out, in this one, they've got customer support listed second. Not sure why. But they have software development listed first. But, but I guess because they exclude software development from the subsequent chart. But the thing I find most interesting is it doesn't vary much by industry. They're drawing these big circles around industries. But if you look at the first row, software development, the leaders are like 23%. Okay, one's at 27%, one's at 22%. The laggards are at 18%. That's not in terms of percent. That's not a very wide range. And in customer support, the range is even tighter. The lowest is 17% and the highest is 21%. It's an exceptionally tight range across industries, which to me says kind of everybody in an industry, kind of across-the-board industry way, is using software development and customer service or customer support as their top use cases.

Yeah, you know, that did jump out at me too, Dave, because on that page 12, where software development, you know, I think it ranged, what did you say, from like 18% to 22%?

27 is the high, but 20, yeah, 20, yeah, roughly.

But then when they did kind of the summary on page 13, it said that software development was only a top use case for 13% of the companies. So it didn't really jive with me on software development being so high when they looked at it by industry, but lower when they looked at it over the general population.

Yeah, I agree with you. It's not obvious what's actually going on there. Let's skip ahead here, Ray. I know you wanted to talk about time to value. This will be on slide 16, I believe.

Yeah. And one of the things they did was they showed the time to value with AI agents broken down by size of companies from SMB to mid-market to enterprise, large enterprise. And what was interesting was SMBs, you know, 24% were seeing time to value within 3 months and another 33% were seeing it within 6 months. And then if you went down to the large enterprise, 20% were seeing it in 3 months or less, and another 30% were seeing it within 6 months. So to me, it's pretty positive that almost 50% of even the largest enterprises are actually able to see real tangible, measurable value within 6 months. So that jumped out at me.

Yeah, for enterprise software projects, you usually can't get them deployed in 6 months, right? And here we are, we're having roughly 5% to 8% on average saying they're getting time to their first outcome, in effect, in 1 month. And roughly 20% on top saying within 3 months. And actually, you got to add the other 6 on this. So, probably 25% within 3 months and then 55% kind of within 6. So, I hear you, Ray. These are fast and this is good news that these systems deliver value quickly. Like faster than— this is, I mean, let's just think SAP took 2 to 3 years to deploy, right? Like, and now we're delivering value in 6 months or less. So, this is great news in my mind for software.

And then, you know, the last thing we didn't— they didn't show this on a graph, but across all these 1,000+ participants, that the median time to value was 6 months or less. Median time. So to me, that's a real positive. Another good sign.

Yeah. And just to be clear, that's not payback. That's just first outcome, kind of first value, just to be clear on the definition. But still, for enterprise software, it's really good. Now, Ray, I know this report introduced something we all love, which is a new metric. This one was new to me. Even though I worked in customer service at Salesforce, I'd never heard the term containment rate. I guess I was there before we had agents. But talk to me about containment rate, please, on slide 17.

Containment rate seems to be a term that really started to be used more in more common parlance with AI agents. And containment rate equals the— take the number of AI-handled interactions for a process. Let's say it's customer support. Divide that by the amount of total interactions with customers and then multiply that by 100. And that's the containment rate. What jumped out here was like customer service, which we saw was one of the top use cases, that 80% of interactions with an AI agent result in an outcome. 80%. That's the containment rate. Call centers, and they didn't differentiate customer service through call centers, Dave, but it's at a 70% containment rate. If you compare that to these GenAI chatbots that we're as humans interacting with on websites, They only have a 50% containment rate. So AI agents have a much higher outcome rate than any other type of interactive tool on websites.

Yeah. And I just want to tighten up your definition of containment rate because I think I have a slightly tighter one because I think you missed— well, anyway, here's total interactions minus escalated interactions divided by total interactions. And that is handled by the bot at all. Right. So if it goes straight to a call center and a person takes it, it's not counted. But if you start to engage, with an AI agent, then you're going to be counted in total interactions. If the AI agent is unable to handle you and says, I need to put you in touch with a human, or you ask to be put in touch with a human, either way, you would end up being an escalated interaction. So, it's in effect, the containment rate is the percent of interactions completely handled kind of autonomously by the AI agent. So, higher is better. And to Ray's point, it looks like, I think you said roughly, AI agents are doing 2x chatbots, or was it even higher in terms of the uplift in containment rate?

It's roughly 2x. It's 80% containment rate in customer service versus chatbots between 40% and 50%.

Okay. Yeah. So, it's 2x again. So, pretty impressive. All right. I know you wanted to talk about who's paying for all this stuff in terms of whose budget it comes from. I think that's going to fire us all the way to slide 29 to talk about Who's paying, Ray? Functional executives or central IT?

Yeah. And once again, we talked about this in the latest episode, and we talked about how what's old is new again, and that's managing the buy of software versus the build and maintain, or, or build, manage, and maintain. And this report said for agentic AI programs, 55% of the budget is coming from central IT budgets. And what that tells me is they're looking at this as traditional software development projects that the IT organization controls from the build to the deployment to the ongoing maintenance and management. And only 28% of the budget was coming from R&D. So it's not your research and development team running these AI agents. or building them and running them. It's your IT organization, which to me makes sense in large enterprise that has very large IT organizations who aren't developers.

Yeah. So Ray's looking at the chart on slide 29. And I think it's important to point out, Ray, here that this chart adds up to well more than 100. So it's actually looking at funding sources, i.e., one or more sources of funding will be contributing to these projects. And the number one source, as Ray pointed out, is central IT budget, followed by operating budget at 43. Followed by discretionary at 35%, which is interesting. People basically making money out of their budgets, right, presumably by not doing other things or asking for an overage. And the last one, 31% still coming from a dedicated AI or innovation fund, which I think is interesting because one way that big enterprises drive change is they don't try and put the full cost on the managers who are trying a new thing. They'll say, hey, we'll pay for half of it, or we'll pay for a third of it. and will chip in, you can get partial budget relief, uh, by tapping into the dedicated AI or innovation fund. So I thought that was interesting. It's down from earlier, I would almost certainly believe, but it's still there.

Yeah, so I— we're coming up to time already, so I guess what I'd love to do, Dave, is ask you kind of overview and summary of what you took away from the report.

Well, I don't think, Ray, I have anything really breakthrough here. I, I just think it's nice to see A report that I think is better grounded in actual reality than the MIT clickbait. So, I tend to believe everything these folks are saying. You know, AI is— I think for me, the biggest takeaway, Wayne, was we're no longer in the kind of FOMO phase of go buy AI even though we're not sure what to do with it. And they're arguing that that era ended in 2023, 2024. And in 2025, we're now in kind of the pilot-to-production era. And I do think it's important for vendors and their messaging. Like, as a vendor, as somebody who's spent my career at vendors, it means to me I need to be messaging a couple of things. One, outcomes, right? Two, probably not messaging so much like, oh, what were you going to do with AI? If you're not sure, call us. But three, instead messaging like, you can get tangible results by doing this. One of the things we didn't really talk about here that did come up as well, Ray, which was just trust, that somewhere in this chart they said like people only trust the vendor messaging about halfway. I think it was 48% of buyers trust what the vendors are saying. So you got to be very careful on the claims you make because people are now viewing you with skepticism. So that was it for me, Ray.

Yeah. And then the last thing I'll say is there's so many other great things we couldn't cover, like with agentic AI, that integration points to legacy systems is critical. And that these hybrid approaches are also emerging where, hey, I've got a robotic process automation tool like a UiPath, and I'm combining that with AI agents. And then the last thing, and I think it's really important for everyone to understand, is let's not throw together generative AI software and agentic AI systems. They're discrete and different solutions, and the use case and the deployment are going to be associatedly differentiated. So understand where each type of AI product fits best into your business, your process, and that's going to determine probably who buys and who maintains it.

Okay, well, go to bed, Ray. You've earned it after a long day and a 10 PM local time recording time. Thanks a lot, and we'll see you all next week.

Okay, good night everyone.

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 at Ray Reich. Dave Kellogg is based in Silicon Valley and available on X at Kelblog. Schenectady, which is French for unspellable, is not our actual production location. Thanks for listening.