AI Slowdown Calls Mount as Enterprises Push Ahead With Agents

0:00Hi all, Brian Felt, host of the AI Proving Ground podcast here. We're kicking off something a little bit different today. Starting today and every Saturday moving forward, we'll be talking with Worldwide Technology Chief Technology Officer Mike Taylor about the week's news and what it means for enterprise AI leaders. Mike spent nearly 20 years in the IT industry, all of it here at WWT. He's in constant communication with AI leaders around the world, spanning Fortune 100 enterprises, frontier AI labs, and leading technology companies, giving him a firsthand view of where AI is headed and how it's actually being put to work.

0:34The goal is to talk about and distill news that actually matters to those driving AI strategies within their companies, and to give them the context they need to hit the ground running Monday morning. It's Saturday, September 19th, and this is the AI Proving Ground podcast from Worldwide Technology. Let's jump in.

0:59All right, well, Mike, welcome. We certainly picked a heck of a week to start this. We're going to throw you into the deep end right away here. Talks of an AI doomsday have really hit a global fever pitch and after an anthropic researcher quit last week over safety concerns. On X, he said Frontier Labs were, quote, racing straight to self-improving superintelligence and gambling with our lives and that AI will soon be superhuman systems that can hack anything. And adding fuel to that fire, a separate anthropic safety researcher, excuse me, put a greater than 10 % chance on AI killing all humans within a decade.

1:36Now, that's heady stuff. And the New York Times actually called the last week, quote, one long freakout over the existential risk of AI set off not by critics, but by the people actually building it. And that freakout in some capacity has led frontier labs, such as Anthropic or OpenAI, amongst others, to call for government safeguards and a potential slowdown in new research and development. We'll get to the slowdown part here in a moment, but I do want to get your reaction to that. What do we make of these new alarm bells going off? Because AI doom warnings are nothing new, but this is snowballing like nothing we've seen already, garnering global attention.

2:11So how should the enterprise leaders listening to this grapple with all of this? And is there anything they can do to steady that ship? yeah i mean let's start with the softball topic right i mean why not let's just jump into that you know i think some of this from just a pattern of technology or even i think socioeconomic change we should expect this right this should be something we all expect from a standpoint of maturing capabilities endless you know imagination also validated by real circumstances that have come out here that raise alarms and bring new thinking to sort of old problems or old circumstances.

2:53The parts that I, at least from my perspective, that I think about here are, number one, I think everybody's got an agenda in all this, which is tough. You know, it's been politicized now, you know, as we think about regulations, we think about speeding things up, slowing things down. But if I just look at the core of what we're looking at here, the circumstances range from this achievement of AGI in some respects, whether you say, look, it can drive a car better than you can, or it's going and solving mathematical equations that we've been struggling with for 200 plus years. In the midst of that, these models still are not great at simple tasks.

3:37We've been using Fable and some of the advanced model capabilities to do deep research and insights, and it still just misses things. And so, you know, I'm somewhere between, certainly there is a version of this that, you know, ends tragically. I think the percentages on that are extremely low, given where the technology is today. And I think even projecting out to where it is in the future. I was joking with a colleague the other day. I don't think the human extinction thing happens without a bunch of robots, just my personal opinion, you know, to carry out some physical things that would need to be accomplished there.

4:14But, you know, all kidding aside on it, I do think there is a level of confidence that the frontier model providers, and by effect, the open source model providers, including recent transactions and other things that have brought these open models to the forefront. There's going to be responsibility and in an enterprise sense, liability. They're going to have to contemplate and do a better job of providing transparency to people, not just on what the models can use, but also how they're inferring data from prompts that go into some of these tools to a couple of examples and stories out there There were, you know, even the recent solving of the mathematical equation piece of it, all the scientists are arguing, well, they derived a lot of that theory based on, you know, how we were using the tools to get the research done.

5:13So they're just not at a point, I think, where they're truly doing these outside of some human supervision, as well as maybe even some human influence to how they want those things conducted. And in the same sense that some of these security circumstances are concerning, it also, I think, in a way can promote perhaps the capability of these frontier models against one another as they fight to be the leader in those spaces. Yeah, certainly serious subject matter, but also a bit existential. Let's get more to the slowdown angle here, which is maybe more pertinent to our enterprise audience. Anthropic co-founder and CEO Dario Amadei published an essay last weekend urging labs to slow the rate of capability advance.

6:01And OpenAI CEO Sam Altman and Elon Musk and many others publicly agreed. Altman even went as far as saying that OpenAI would not proceed with an IPO this year on safety grounds. Now, Amadei warned that a more capable version of the agent swarm behind recent incidents could within a year run a persistent botnet across the Internet. Now, putting aside any of the antitrust or market volatility questions that are certainly going to arise from all of this, for the enterprise, I think this is the first time that model vendors themselves have signaled a potential pacing change. And I'm curious, from your perspective, how does that affect roadmap assumptions, contract terms, and just strategic planning in general?

6:43Yeah, I think first we have to recognize that for the first time in a very long time, something that is both extremely innovative and, you know, potentially serves consequences of, you know, call it good and evil, depending on who's using it and how it's being done. And I do applaud what Dario is doing and the leaders are in terms of bringing these things to the forefront, exposing them to people in terms of how they think about it and how not just the message they're sending to the market, but as well as you know, and the listeners, you know, every employee in Anthropik is reading that memo too and thinking about how they're conducting their job and their function, you know, within the company and to do so more responsibly.

7:27That said, what I didn't hear from anyone who's making those announcements is what they intend to do within their organizations to slow down. or outside of, you know, Musk made some recommendations that I think at least he said, you know, some of the other frontier model providers were willing to do around third-party inspection of, you know, how would you get OpenAI to use or build harnesses that would inspect the reliability and the safety around anthropics models, right, versus the government. And, you know, the news stories that, you know, the frontier models and, you know, other key partners in that AI ecosystem are lobbying the federal government to stay out of it.

8:11We need an alternative then. And I think, you know, you know from your career, you know, not just with the pod here, but, you know, it's hard to proofread your own stuff. You know, when you look at things that you built, you just have implicit bias to either understanding what it is you were trying to do or reading through maybe some things that needed a little bit more attention and detail. So there's going to have to be some external standard that we're putting to these models, both frontier and open source, or the customers that I talk to are going to be hesitant to scale them from prototypes to applied AI use cases that are working across their enterprises.

8:53But we need an answer to that. And that's part of what I think we're all waiting in the wings here to hear a bit more about or understand better is what are those suggestions going to be? And when they say they're slowing down, what is that? What are they actually doing there? How are they incorporating safety into more of the core of what they're building? Yeah, another component here. I mean, one of our favorite sayings around here at WWT, as it relates to AI, is what isn't possible this week may be possible next. You know, even our own RFP assistant is kind of the poster child of that ethos.

9:24It was an AI power tool that we were ready to write off pretty early on, only realize a few weeks or months later, and this was several years ago, that we were on the right track, actually, thanks to model advancements. And today, it's one of several AI applications we've developed here that's driving the highest ROI. So it's actually turned out to be fantastic for us. How does the idea of a slowdown and how does that change the math behind how we evaluate AI use cases, knowing that there could be a slowdown coming and it may not be the don't, you know, just wait until next week type of mentality.

9:55Yeah, it's a fair question when I'm getting both on the call it kind of how do I grow and differentiate the business as well as how do I secure it? And my short answer right now is most of the organizations, you know, including even I'd say some spot use cases inside of worldwide technology is that the room between where we are from a data and our customers are from a data readiness perspective and what the models are capable of doing, we've got a couple of months of runway there to keep up, just to keep up and closing the gap that's there. And so, you know, I don't see any immediate impact to enterprise adoption or our own adoption over the next three to six months.

10:38If anything, I see the imagination of organizations continuing to grow as they see one or two things that go really well. RFP assistant to respond to RFPs turned into Proposal Studio, turned into an access and corpus of knowledge data that we're now integrating into not just the ATC and the digital assets we have there, but our e-commerce platform around buying decisions. So these things, that flywheel effect that you've talked about in the pod, there's so much room between where we are and what's possible that I don't see an immediate impact in the next three to six months. I think the constraint that remains out there is still just around data center capacity, where are people getting power?

11:24And we are seeing all kinds of unique things, you know, where the public opinion around data centers, you know, generally negative and trending more negative over time, balanced with the fact that none of this AI, all these great AI circumstances we're talking about, they don't occur without continued growth evolution, which needs power space pooling in these data centers. We're seeing customers, NCPs and others that are going to far reaches of the world where hydro is more available or other resources or repatriation of manufacturing facilities now in the data centers. So we're seeing all this movement, call it at the base layer of this token world that continue to move along very quickly.

12:14And so I don't see a slowdown there. I don't see anyone backing away. If anything, particularly on the security side is how am I using AI to defend against threat actors that have instrumented AI for not the grand purpose of what we've seen in the news around agent swarms doing hacking, but just nefarious organizations that have instrumented models to go and do more persistent penetration testing, better secrets gathering and interpretation of those secrets for the purposes of access. That's where I see us and our customers spending more of their time. Yeah, I want to get to a little bit of a counterpoint to that slowdown, Mark Palantir CEO, Alex Karp, writing on X.

13:04He was talking about what model companies or what he thinks model companies are actually asking for when they call for regulation or why they might want to be nationalized. He said, quote, it's a weird twist. They're asking for societal regulation to get out of the first line of defense, which is if you build a technology that can destroy 10 percent of the world, that has civil and criminal liability attached to it. Now, you know, what do you what do you make of that? How does that actually, you know, lend itself to this conversation right now? Yeah, I think going back to a bit, well, exactly where we started in some of this, you know, is the, not just the, what are you doing with my data?

13:43Where is my data going from prompts to training and, you know, inference and everything in between? but at what point are the behaviors, whether instructed or implied within these models, when are they your fault and when are they the model provider's fault? You know, and at what point do instructions become so good that they're rock solid or vague enough that, you know, liability is up in the air, right? And, you know, when you and I hit a fender bender we don't get to call General Motors or Ford and say, hey, your car should have known I was gonna run into that fire hydrant or that parking bus.

14:27It's who's operating that. And I think the challenge in this world is that you have human augmented activities here and the augmentation capability is the piece that I think Alex is talking to there. And those lines are not drawn. I don't think they're well understood. And, you know, the evolution, I always go back. Everybody gets and understands is, you know, when you do accidentally rear end somebody, you don't get to go back to, you know, to the provider of your level three or four autonomy in your vehicle and say, hey, it did an emergency brake for me. And now I got a, you know, I got a deductible.

15:08So we've got a lot of work to do there in this. I just still do not see the consequences yet being severe enough that it's slowing anyone down. I think we're asking better questions. I think we're being smarter about when to use Frontier versus on-prem. And we've seen some huge investments in the, call it, privatized world of both inference and training that I think we've predicted on this pod would be, you know, here and now in terms of the balance of on-prem demands. And, you know, NVIDIA has made some big swings there too in terms of acquisition that they announced. So we're going to see more of a mix with this around where is your data?

15:57Who's got access to your data? But in the end, I think we're going to have to draw the lines as we're on the road here a bit. Yeah, absolutely. I mean, it's not just an accountability conversation either happening right now. We're also now getting into usage. Manthropic just recently announced Claude Docs, which is coming out and competing and potentially raising the stakes for what Microsoft has. how should we think about, you know, from an enterprise standpoint, who's getting to use these tools and what do you think some of these model providers are thinking there? Yeah, I think overall, the more choice we have, generally speaking, the better.

16:34And so, you know, it is a complicated web. You cover it every week, you know, between the relationship. And I think we saw some of this. The best example I can give is in the cloud world, right, where we had, you know, we had partnerships being announced where we would scratch our heads and say, gosh, those feel directly competitive or, you know, how is that going to work? And then you had frenemy statuses, you know, across different partnerships that were going on between private, public cloud and the software and hardware that connected those things. I think we're in the midst of the same journey with the AI providers and the, call it the capabilities that have been built on top of there.

17:15So you go back to a year, a year and a half ago, right? You know, Cursor was one of the biggest consumers of Anthropic. And then Anthropic comes out with Claude Code. They're now competing and Cursor is still using and accessing Anthropic models. You had Microsoft, you know, came out with their version of Co-Work, which they publicly announced, backended by Anthropic's models. And now you got Anthropic coming out, you know, with Claude Docs. is going to directly... Provided we don't start seeing frontier model companies cutting off access to their ecosystems, I think this continues to be a productive thing for all of us in terms of the choices that it derives, where it's going to start to get heavy.

18:02And there are examples of this is where a model provider decides that a third-party partner is too competitive to their core and all of a sudden you can't use that model through that third-party tool. And there are examples of that that's come out for a lot of different reasons. Could be competitive, could be philosophical differences on how something's being conducted, could be the excessive demand that some of the frontier models are under. But I think if we started seeing that ball of yarn start to unravel, I think it becomes unproductive for the consumer and limits our choices as enterprises.

18:42Yeah, well, I mean, regardless of potential for a slowdown or usage or anything like that, I mean, every new model release now seems to come with its own big warning sign in regards to their ability to control the agentic technology that they've created, mythos, fable, now Astra. I was reading a new investigation from Meter, which is a research firm that evaluates frontier AI models for capability and risk. They found roughly 1 ,200 agents sent more than 70 ,000 unsanctioned messages through a hidden artifactory channel. A hundredths of those were involved in the hugging face incidents. Now, this seems to impact anyone running agents inside of a corporate network.

19:19So how do we prepare for this rapidly evolving cyber universe where every new model is just bringing seemingly new risk? Yeah, I think this is a really good lesson for us to think about as we go from, you know, I'll call it workforce experimentation with AI to and through to applied AI solutions at scale within our companies. And, you know, there is this whole sort of thought in battle between, you know, so much of what we do today within organizations, especially if you think about a business process flow from the RFP example in sales to think, you know, finance and collections and payables, etc.

20:02It's a long set of steps. And within those steps, they're mostly deterministic, like you or someone in that workflow is looking at something and you're deciding what to do with it, you know, and that deciding what to do with it is pay the invoice, don't pay the invoice, it's extend the terms, defer the terms, it's, you know, in the in the The RFP assistant world, you know, it's deciding to respond to something or not. And then which components do we need to include? And so the lesson there from my perspective and what we're learning as we've implemented these things internally and for our customers is that, you know, as we spin up these agent swarms, we have to pay real close attention within each agent or within the groups of agents, where are their functions, how, you know, we don't need a billion parameter model access for specific agents or agentic functions.

20:57And the more that we can build context around those functions, but link them together to produce a business outcome in a, call it, deterministic process sort of focus, that's what needs to be done. Going in and prompting and saying, I need you to help me create an agent swarm to decide what I pay and what I don't pay from, you know, payables or what I respond to or what I don't respond to from a response standpoint is their response. You need someone who's giving the context of what we do, why we do it. And then also, at a minimum, given the scale of what's going on, providing some level of either human supervision or human inspection.

21:40We have to treat and find ways to collect the data that these agents are doing and inspect them like they're employees. And I think that's going to be the tension in the system is the opportunity to let something kind of go on its own in the scale that that's going to create versus some level of human supervision, even if that's in retrospect to go back and look at the thousand things that that thing or sets of things did today. Nowhere is this more present today than in the development context where you have developers building out software. I was just on a call right before this. We have some really, really smart, capable developers doing work internally and also for our customers.

22:25At the end of the day, if they don't set up the right architecture and the right parameters around the agents, the coding, the test cases, the data collection and telemetry that come out of these, you know, the longevity and supportability of that software just continues to degrade and get lower and lower. And so I think if I have one tendency, as you've learned from me, is how do I glean from the past what is now in the present? I think we can look at the journey many of us has taken with the software development, the AI native engineering purposes there and say, how are we applying that to other agentic business functions?

23:05and ensure that the beginning of that is not, hey, I need to create a process to go in and do this, but we're almost working sort of ground up, value stream map, what's happening, what are the parameters and rules within each step. Yeah, and just to add to that, Cisco president and G2 Patel, a great friend of us here at WWT, he posted on LinkedIn recently, we're past the point where we can remain safe and secure without the use of AI. And he called AI a prerequisite for safety and security because his attacks become more sophisticated and machine scale, human scale. He says defense simply won't be able to combat the attacks due to the sheer scale of differential.

23:45So, you know, he called for a need to fuse data sets of security and observability as it gets harder to distinguish a malicious attack from an act by an agent that exercises poor judgment. I'm curious, based on what you said and what he's saying, how do we just how do we start to actually build and work towards that type of architecture and mentality? Yeah, this is, I think, where it would be one of the areas that I'm most excited about what AI can do for us. And I think it certainly applies in the context of security, but by extension, it also resides squarely within reliability and availability, which, you know, in the midst of AI, everything else, we, you know, the CIOs and leaders of these platforms within our customers, we can't forget at the end of the day, things have to be up and running.

24:34And most of the time, the reason that they weren't was because some human being did something they shouldn't have and stuff broke, right? It's the morning after the maintenance window. It's the morning after the patching. It's, you know, those sorts of events are what generate some of this. Back to the security consequence. the the number one reason that a bad actor or threat actor is able to reside within an environment depending on what you read the average is there is anywhere they've been in your environment for anywhere from three to six months and you didn't know it right forget much less how they got in but you didn't you didn't know that they were there and and the reason you didn't know that they got in.

25:19And the reason you didn't know that they were there is the amount of data that is logged within enterprises and other organizations. It's immense. It's massive. And the ability to thoughtfully interrogate and correlate across that data was a problem that was by most extents limited to machine learning as mechanisms to aggregate and identify those gaps. Well, the transformer technology and logic and what's been introduced there, along with the pattern matching that can go into the outputs of those things, I think has created the opportunity to react on the perimeter and identify within much, much faster.

26:05And so when I hear G2 talk about those things, I look at, you know, a vast set of capabilities within their portfolio, their AI threat defense platform that they've built out that's looking at these different endpoints and correlating that information together, that helps us identify and react in a much faster way. Where that translates into the reliability and availability, the amount of time spent when something is down or out and not performing, same circumstances apply. How do you interrogate massive amounts of data, draw correlation between events, not even to a completely deterministic end, but you're able to lead the witness?

26:54Those are powerful and I think are exploited more in a positive way today than ever. And that trend is going to continue. It has to, to G2's point, because the threat actors are going to be doing it for that negative consequence. You can be assured of that. Yeah. Well, I mean, even despite concerns, like you mentioned, they have to continue. And in fact, they actually are advancing just this week at Dreamforce. Salesforce introduced AI Force. How many times can I say force in a single sentence? But AI Force is an interface layer exposing its business logic, security, and workflows to external agents.

27:29That was from their website. And according to Salesforce, AI Force will bring the data, workflows, business logic, semantics, permissions, security, and governance that already run in the business to any AI interface that we might be using. I mean, Salesforce, according to CX today, they're betting that the CRM interface, excuse me, becomes less important than the governed system underneath it by positioning itself as the layer that connects models and agents to the information controls and tools needed to do work inside any actual enterprise. I mean, do you see this? Is buying shifting from seat licenses towards unlimited agent access and action execution pricing?

28:08Is that where we're headed here? Yeah, it's a confusing world today. And I'd say this is existing within the workforce productivity toolkits, whether it's embedded into platforms like Salesforce, Microsoft, ServiceNow as a workday. Every major SaaS provider or platform provider is contemplating the same circumstance. And to at least the ones I've named and the ones that we've worked with, every one of them has a per seat and a per use model, depending on which product you're talking about. And when I look at what any of those organizations are building in terms of, call it an agentic layer into AI ready data, I look at an organization, I say, you as an organization need to be building something similar for yourselves.

29:03you need to be thinking about what does your version of AI Force look like because you're using five and a half to six SaaS apps that have a lot of different data in there for you. And whether it's a Percy license today or it's going to be by the drink in the future, access to your data and what it's going to cost you is something you need to be thinking about. And I think that's where a lot of the interest in the on-premise and open source or open-weight approaches has started to gain steam and momentum. People are starting to see their bills come in. It's not just about the use of frontier models.

29:49It's about the access to the data that is in other places. And the meter is running in those circumstances. So I don't think we'll have one that wins the whole thing. I think your ability to control not just your spend, but your outcomes is going to be having or thinking about building an architecture that mimics what some of the SaaS providers are building, but in the context of your data, your assets, and the most valuable pieces of information you have in your company. Yeah, I mean, tokenomics certainly top of mind for every organization out there, something we've all got to figure out very much in real time.

30:32And it's anything but easy. I mean, here's an interesting article that I was reading. It was from the Wall Street Journal. They were polling executives at their Wall Street Journal WSJ Technology Council Summit in New York earlier this week. And they were saying that AI is delivering outsized value in some cases, but that in the majority of cases, it can be wasting more money than it's worth. And companies are having a hard time distinguishing which is which. And here's the most interesting quote that I found in that article. It was David Hsu, CEO of AI Harness Company Retool. He said, quote, We believe that about 90 % of the tokens you're spending are probably negative ROI.

31:10It's just the last 10 % of tokens that are really driving a lot of that ROI. Now, WWT's token bill has surged from a few hundred grand per month earlier this year to well over$2 million at this point, and that number may be even growing as I speak. it's a major discussion point amongst you know the executive team with yourself and every company that we we deal with but what's going on here how are we going to have to think about that as we head into kind of the back half of the year into 2027 are we talking token limits are we talking just most of what you said in your last answer just give us a level set yeah i think about a couple things on the the tokenomics and the spend side i think first is we're trying to feed into curiosity we think curiosity is is going to win here in in terms of the best ideas again related to our strategy.

31:56We're not reinventing our company strategy, but we're saying, how do we feed into that and use AI, you know, to feed our strategy? Building those things from the bottom up, there's going to be waste in that process. You know, it's people using the tools for the first time. It's iterating through ideas. You know, I'm sure as you, you know, think about the pod or you're building out content and other things, it's rare that you get what you want the first time. You're iterating and iterating and iterating. We're also learning how to better prompt, include better context and skills around these things, all to say, I want curiosity and I want some appreciation or recognition that when we first start off, we're going to be a little clumsy with this.

32:42And so where we are is, okay, how do we eliminate some of the waste that's just associated with accessing data, the design skills that go into producing, whether it's PowerPoints, reports, et cetera. If we can contextualize that into a skill or into a markdown file that somebody uses that we know is sort of the best way to do that, that's where we're really saving in this. It is a little bit about picking the right models. I mean, you know, Fable for the right purpose versus, you know, Sonnet or Opus. You have to be thoughtful about where you're using those, but the real savings in our experience comes from the architecture.

33:26Do you have data that is AI ready? And does that data have the context that it needs to be produced and reported in the right way versus someone who's running a query against something and then continues to prompt and prompt because they don't know why they're not getting an answer or why it's not including something. because the instructions for that particular, it's just churning. And so I tell our teams, our customers as well, if you're asking me to save$100 on tokens, I'm going to get 90 cents of that by building out a data-ready architecture, harnesses and context around data and applications that I want to AI-ify and build out in that way.

34:09And so we've been tolerant of that. That said, we've also seen situations where without limits, without capping what, you know, in our world is a workforce user versus a power user versus a developer, they can run through those ceilings really quickly. And so having some limits to that spend and then using those limits as people approach them or, you know, need to breach them to understand is what you're working on valuable? Is it helpful? And, you know, the second lever, if I've got 10 cents left, you know, there in that$100 or in that dollar that I'm trying to save, the$100, I think I said, that I'm trying to save there is you have so many people who are trying to do the same thing.

34:55you know, whether they're on the same team or they work in different groups. So we've created a skilled library that we publish out to the entire organization to say, if you're looking on, you know, for reporting on sales, customer research and analysis, here are the sort of certified sets and skills. And by the way, you can evolve these. They're not completely locked down. You can build on top of them, but gives folks a better way to get started. They're not starting from zero. They're starting from something that we think is best in class in and around repeated business functions that we know, you know, we've got hundreds of people doing on a daily, weekly basis.

35:37Yeah, certainly a balancing act and one that organizations are going to have to grapple with for some time here. And Mike, as I understand, you're having a balancing act as well, giving us all your insight with the news and balancing parents weekend out at a university in Ohio. So we appreciate you spending the time here. Go easy on the frat parties tonight. Enjoy the rest of your weekend. Yeah, wish me luck. Thanks.

36:06you