Ep 398: How AI Agents Can Bridge the Gap to the Future of Enterprise Work

Resources:

Join the discussion: Ask Jordan and Scott questions on AI


Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineup

Connect with Jordan Wilson: LinkedIn Profile

Try Our Free AI Prompting Course: Register for our free Prime, Prompt, and Polish AI Course! 

Backward Working Approach for Efficient Solutions

For productive and efficient solutions, adopting a backward working approach can be advantageous. Instead of experimenting with new technologies without clear objectives, it is crucial to focus on solving specific business problems. AI agents offer vast potential in enhancing productivity, indicating that collaboration among companies may be necessary to leverage AI for complex task execution.

Necessity of AI Guardrails

These AI guardrails extend beyond addressing inaccuracies or AI "hallucinations"; they are also crucial in maintaining brand integrity. To align AI outputs with company ethics, cultural views, and customer perceptions, enterprises need AI agents that represent them similarly to their best employees.

Becoming Multimodal

In the futuristic landscape of AI, interactions through voice, video, and AR will become commonplace, leading to significant shifts in service delivery. This shift will necessitate a rethink of data privacy and governance in enterprises.

AI's Impact on Sales and Customer Engagement

AI's increasing involvement in sales, especially in automating initial customer outreach, can dramatically enhance human roles, support account executives in building customer relationships, and enrich real-time sales education efforts. The ideal future sees AI agents conducting entire sales processes, demanding an adaptive approach from enterprises.

Humans, AI, and the Changing Roles

Short term, human roles will mostly encompass setting up AI systems to ensure correct data integration and guardrail setup. In the longer term, as AI agents become more autonomous and capable, human roles will inevitably evolve.

The Risks and Mitigation in AI Systems

Given the inevitable errors and inaccuracies in AI systems, it is crucial to implement strong accuracy and observability safeguards, especially for crucial sectors like healthcare.

The Evolution of AI Agents Reasoning

The transition from level 2 to level 3, involving AI agents reasoning with each other, brings significant opportunities and challenges. Tracking the accuracy of these interactions introduces an entirely new challenge of quality assurance.

The Use Cases

Increasingly, AI is being employed for intricate tasks like arranging tow services, showcasing the potential benefits of AI in customer service. Quick and accurate responses can lead to greater customer satisfaction. However, it is imperative to address the challenge of implementing guardrails for multi-agent environments, recognizing their potential of becoming a barrier to adoption by 2025.

In conclusion, we must acknowledge the prevalence and mainstream integration of AI agents in enterprise work. They are having an impact, are transformative and capable of integrating and acting on various systems. But with every boon comes certain risks that have to be considered and effectively mitigated in time. This will ensure the advantageous utilization of progressive AI technologies, benefiting business enterprises significantly as we move forward in the future.

Topics Covered in This Episode

1. Current State of AI Agents
2. Challenges in AI to AI Interactions
3. Guardrails in AI
4. Humans’ Roles in AI Integration
5. AI Agent Use Cases
6. Future of AI Agents

Podcast Transcript

Jordan Wilson [00:00:17]:
About 18 months ago when I spent every day telling you all, hey, AI agents. You need to be prepared. They're coming. They're coming. You guys probably thought I was a little bit nutty, but here we are now in the late fall of 2024, and AI agents are not coming. They're not some weird AI powered future. They are here. Right? So between actual large language model companies that have pushed this live, such as Anthropic Claude with their computer use tool, and also all the names in big tech Microsoft, any day now rolling out their autonomous AI agents in Copilot Studio, Salesforce with their Agent Force offering.

Jordan Wilson [00:01:04]:
AI agents aren't some weird figment of our future imaginations. They are not only here, but I think they are now becoming mainstream, and they are here to stay. So I'm extremely excited to talk about that today and how AI agents can bridge the gap to the future of enterprise work. Alright. What's going on y'all? My name is Jordan Wilson, and welcome to Everyday AI. Before we get started, have to quickly shout out our partners from Microsoft. So why should you listen to the WorkLab podcast from Microsoft? Because it's made for leaders who know they must adapt to stay ahead. WorkLab is the place to find real world lessons and actionable insights to guide you and your organization through your AI transformation.

Jordan Wilson [00:01:48]:
That's w o r k l a b, no spaces, available wherever you get your podcasts. Alright. And, technically, another place to get your podcasts is your everydayai.com. Right? We have nearly 400 episodes no matter what you care about. AI agents, marketing, sales, customer service. We have it all there, bringing the world's leading experts on the show, to share their secrets with you and help us all prepare. So if that sounds like something you wanna do, make sure you head over to your everydayai.com. Sign up for the free daily newsletter.

Jordan Wilson [00:02:23]:
We will be recapping today's show, and you can check out a whole lot more as well as, normally, we bring this to you live. We are debuting this show live. It is prerecorded, but if you want the AI news for today, it's gonna be in the newsletter as well. Alright. Enough chitchat y'all. I am excited, to bring on our guests to, for today and talk AI agents and how they can really bridge the gap, for the future of, enterprise work. So please help me welcome to the show. We have Scott Beechuk, the partner, partner at Norwest Venture Partners.

Jordan Wilson [00:02:57]:
Scott, thank you so much for joining the Everyday AI Show.

Scott Beechuk [00:03:00]:
Hey, Jordan. Thanks for having me.

Jordan Wilson [00:03:02]:
Alright. Hey. Can you tell us a little bit about, Norwest Partners and and what you all do there?

Scott Beechuk [00:03:07]:
Yeah. So, Norwest, we've been around for over 60 years, global platform, venture and growth equity, invest across enterprise, consumer, and health care. And, we're here in North America, California, also in Israel and, India.

Jordan Wilson [00:03:24]:
So tell us a little bit about kind of what caught your, attention when it came to AI agents. You know, kinda like I talked about, they're not necessarily new. Right? And I think they've been a part of this, almost sci fi, storytelling for many decades. But, you know, from your vantage point at Norwest, when did you start really looking at AI agents and saying, okay. This is, an area that we need to be invested in and paying attention to?

Scott Beechuk [00:03:53]:
Well, you know, it's funny when you mentioned sci fi. I mean, Isaac Asimov used to write about this stuff in the sixties, and you're absolutely right. It's, you know, it's been a long time coming. We've been investing in AI for over a decade. Machine learning, deep learning, transformers came about, and here we are in, the generation of generative AI. I think, you know, AI assistance, with ChatGPT and sort of the early versions of transformer based, GenAI were were really exciting. I think we started to start to see some of the things that were possible, but here we are, I think, in the next major, chapter, of this unfolding, and AI agents now are more capable than ever. You said it earlier and which is absolutely right.

Scott Beechuk [00:04:35]:
They're not coming. They're here. And we are just on the precipice of unlocking a huge set of new capabilities for enterprises and, consumers.

Jordan Wilson [00:04:45]:
Yeah. And and I can understand because, you know, for the average person that maybe doesn't pay too close attention, things are happening fast. Right? You you try to keep up with what the big companies, you you know, the Microsofts and the Googles and the Open AIs are doing with their AI tools, their platforms. And now all of a sudden, you know, everyone is talking about agents. Right? How how do you think that this conversation has transitioned so quickly from, you know, traditional AI and then we had the, know, kind of, quote, unquote, ChatGPT, or generative AI wave at the ChatGPT moment in late 2022. And now everyone is talking about AI agents. Scott, why?

Scott Beechuk [00:05:28]:
Well, you know, I think whenever a major technological shift happens that could unlock a ton of, like, life changing capabilities, you get the best minds in the world all piling on it. And so you're talking about millions and millions of developers all over the world that have really leaned into these big platforms. Start with, you know, OpenAI, and now you have Anthropic, and you have Coherent, Mistral, and a whole variety of open source projects that are pretty interesting. And when you put that many smart people on a single platform in a single architecture, you start to unlock the art of the possible, things happen really, really rapidly. I think it also helps that a lot of the larger LLM, third party providers, they're starting to provide some pretty robust frameworks that we can work with to actually develop agents. I mean, OpenAI alone has this amazing, you know, evolution that they keep unlocking more and more and more for developers. And so, you know, you can you don't have to squint too hard to see a future where we start to build software that completely changes all aspects of how we live our lives and do our work. And AI agents takes it to a whole new level because now developers can start to not only, you know, build simple applications, but now we can actually build code that integrates with other systems, can draw on datasets, can learn, can fine tune, and now can actually take action on other external systems, enterprise systems, consumer systems, systems of, you know, action and data.

Scott Beechuk [00:07:00]:
And I think that in and of itself almost is is like a a a huge universe of possibility that is that we're that we're just about to see come online.

Jordan Wilson [00:07:13]:
I think everyone rushes to the pros of AI agents, but with promise, there's also the peril. There's potential downsides. Scott, can you quickly just walk us through and and maybe just talk about what are those, you know, ups and and the pros and then the cons? Because I think people just rush to what is possible or what could be possible, but you gotta have guardrails. You gotta think about safety.

Scott Beechuk [00:07:40]:
Yes. I think that's an important question because you start to think about what types of human knowledge tasks we can automate, with AI agents, and certain things come to mind that are that are somewhat obvious, like tier 1 customer support. You know, you got you got tens of 1,000 100 of thousands of people all over the world answering the phone, answering the same question every single day. How do I reset my password? How do I get access to my account? Well, those types of things, you you it's you can wrap your head around that pretty easily. But then you start to move up the stack and you start to think about, well, how do I automate more complex thinking? Things like selling, or you move into some regulated industries like health care. And how do we start to predict, you know, clinical, types of applications, that AI could handle? And you start to think about all the things that could go wrong. Right? So, I mean, if I give, if an AI agent gives a customer the slightly wrong information for how to reset their password, well, maybe not the end of the world. But I start giving let's just say I call up my insurance company, and I'm stranded on the side of the road in a blizzard, and I need a tow truck, and that AI agent isn't getting me that tow truck ASAP.

Scott Beechuk [00:08:54]:
I gotta stand out there for an hour. I I could you know, that's that's a little more uncomfortable. And then you take it one step further and you think about calling up an AI agent on the phone or chatting with 1 online, asking for, you know, some some medical advice, and it starts to hallucinate and it starts to give you bad medical advice, well, that's a whole another level of risk. So the the more sophisticated and the more we move up the tiers of knowledge automation, the more important it is for us to think about how to build the proper systems of guardrails and just accuracy, observability, into these types of systems.

Jordan Wilson [00:09:35]:
Speaking of of tiers and levels, I think this also helps frame why today's conversation is especially timely. Right? About 2 or 3 weeks ago, OpenAI CEO Sam Altman kind of acknowledged, you know, they had released or, you know, leaked out their kind of five levels to HEI. Right? Level 1 was chatbots, level 2 reasoners, and level 3 agents. And he kinda said, hey. We've achieved level 2 now with reasoners, right, with this new, o one reasoning model. And then he said once that happens, the agent's stage comes pretty quickly. It's not as big of a gap between steps 1 and 2 as it will be from steps 2 to 3. You know, when we think about Scott, AI agents being able, to reason with other AI agents, What does that unlock for enterprise?

Scott Beechuk [00:10:25]:
Well, it can unlock a lot of good things, and it can also unlock a lot of, the need for a whole new way of looking at quality, quality assurance across the the enterprise. Because you can imagine a simp a single AI agent reasoning with a single human being, okay, we we can understand if it were to hallucinate or were to get something wrong, easy to course correct because we can, you know, sometimes, you know, just simply go back and change the prompt or the query, that we send it. But imagine a world where AI agents are talking to other AI agents, and they're hallucinating a little bit every step of the way. Let's say we're stringing together 10 of these things. And as a human, I submit a request to a network of agents that are orchestrating together and reasoning with one another in in order to, achieve a more complex task. Well, the problem there is that it's like the telephone game when we were kids. You know, you whisper into somebody's ear one one concept, and then that kid whispers into the next ear and the next ear, next by the time you get to the end of the 10th kid, you know, the whatever the the original person said, completely different. Well, the same thing could happen in a network of agents, but in a far more profound way.

Scott Beechuk [00:11:38]:
Right? Because these agents could be could be solving something that goes completely off the guardrails and and completely, you know, produces something that not only is incorrect, but might not even be relevant. So these systems can have this, exponentially, a steep error curve. And so the idea of being able to trace the accuracy of the agent network as they communicate with one another is something that's new and something that some companies in that category are starting to try to figure out, how do we in in the future, how do we actually trace AI agent networks, the way that we trace complex service oriented architecture in in classic software?

Jordan Wilson [00:12:35]:
Alright. We have to take a real quick break to tell you about WorkLab from Microsoft. So why should you listen to the WorkLab podcast from Microsoft? It explores the questions business leaders are asking. How can they guide their organizations on their AI adoption journeys? How can the technology help them create new products and business models and maximize value? How should they help their teams reskill for this new era of work? And why is it important to be completely transparent about when and how you utilize AI? Find the answers on WorkLab. That's worklab, no spaces, available wherever you get your podcast. So let's get back to the show. I think that's an important one, and I wanna get back, to this, zoomed out overview. But maybe if we can, Scott, let's maybe zoom in here quickly and talk about some AI agent use cases.

Jordan Wilson [00:13:36]:
Right? Because I could go all day on the theoretical, and then I wanna get back there. But maybe it will help our audience, along a little bit. If we can talk about, actual use cases for AI agents. And I think one of the not saying lowest hanging fruit, but one of the easiest ones to, kind of walk along that journey with and really understand is in customer service. Right. So, as I understand, Norwest has an investment in the company, Replicant, you know, working in the customer service space. And I think it's one of the, areas most ripe for disruption from AI agents personally. But can you walk us through kind of like what that use case looks like and maybe tell us a little bit about Replicant?

Scott Beechuk [00:14:16]:
Yeah. So Replicant, was a company that was started before November 22. So before we had access to things like ChatGPT. And what they do is they act as a tier one support agent on the phone. So if you call up AAA Insurance in North America, there's a pretty good chance you're going to get Replicant and Replicant AI customer support rep on the phone. And they are trained by millions of calls and millions of interactions with customers. So, hopefully, they become as good or even better than your best support rep that you might have otherwise had. But Replica today is a different company than they were back in 2019 because now we've got generative AI.

Scott Beechuk [00:14:58]:
We can make the conversations even better, more robust, and more accurate, and they can also do more complex things. So you can imagine a scenario where I'm on the side of the road and I need to call, you know, a tow truck. We were mentioning that earlier. Well, that's a scenario that involves a lot of integrations to other systems because I'm calling up. I need to figure out that it needs to know where I am. So, hopefully, we we can use GPS. We can tell it exactly what happened to the car, what type of tow truck service I might need, where I need to go, And then it can integrate, and it can actually make outbound calls to the tow truck services. And that is a true that that is what we call, you know, true AI agent integration in action.

Scott Beechuk [00:15:41]:
And, a little bit more complex, use case than we've seen in the past, but that's what's happening today, and that is, like, AI agent 1 point o. We're we're there's gonna be far more complex use cases coming up.

Jordan Wilson [00:15:53]:
Yeah. Oh, absolutely. And I think, that's a very, good use case to talk about and walk through because it's probably beneficial for everyone. Right? Kinda like what you said. Probably a lot of these, you know, customer service, humans are answering the same questions probably over and over, and then the humans on the other end are probably waiting in a long line and, you you know, there's probably some some disconnect there along the way. But, you know, going back to this kind of two sides of the coin because, I mean, you can immediately see how that is huge. Right? In in theory, you know, AI agents can understand human language, in the same way that a human can, and they can in in, like, sometimes more accurately also route a customer's query. Right? Because they have a company's entire, you know, knowledge base essentially in their training data where if you're, you know, talking with a brand new customer service human, they might not really know how to, you know, handle more complex queries.

Jordan Wilson [00:16:56]:
But then can we go back into the guardrails a little bit? Right? Because, going from, generative AI and being able to work with unstructured data, the promise is huge. But even for when we're not talking about, you know, multi agent environments, how important are those guardrails for enterprise companies? Because I assume that's going to be one of the biggest hang ups for enterprises in 2025 not adopting early.

Scott Beechuk [00:17:23]:
I agree. I think guardrails when we talk about guardrails and AI agents, I think a lot of people just assume it's just about hallucinations or it's just about getting something slightly inaccurate, which LLMs tend to do. Right? These are nondeterministic systems. What they output doesn't always agree with the query that you thought you intended to ask. The the output sometimes is a little unpredictable, but it actually goes beyond just inaccuracy. For some companies, it actually, can become a brand issue. And so, you know, how do you how does your organization view, certain, you know, ethics issues? How do you talk about certain political issues? How do you think about culture? And how do you want your customers to view, the sort of the empathy or the sympathy of your particular company or your brand? And those things are a little softer and but they also matter. And so guardrails are really designed to solve for all of those types of challenges to make sure that agents represent an enterprise the same way that your best, most, sort of enabled, compliant, employees, would do whenever they interact with your with your customers.

Jordan Wilson [00:18:41]:
Yeah. I think that's a I think that's a great point. And and and, also, the capabilities, right, of these agents are changing quickly. Right? Like, you kind of already referenced, you know, a lot of these companies existed pre generative AI, so their capabilities were much less robust. Right? It was much more, you know, binary 0 and ones deterministic, and now the capabilities are in theory, quite limitless. So, you know, as we look into the future, which I know it's hard to do, I'm not gonna ask you to bust out your your crystal ball. But for decision makers right now at enterprise companies, what are some of the most important things that they need to consider? Right? Like, what you just said, with the the the brand issue. Like, how AI agent can actually be a brand issue is a huge and I think an important call out.

Jordan Wilson [00:19:32]:
But what else should enterprise leaders making decisions on, should we go all in on AI agents? What else do they need to consider?

Scott Beechuk [00:19:39]:
Well, one interesting thing that's coming that I don't think a lot of enterprises have fully, grokked yet is the idea that the future of AI agents is multimodal. And so we've been sort of accustomed for the last couple years to chatting with our AI assistants and now AI agents and chatting, you know, using text or using voice in some cases. And but but where we're going is you are gonna interact with we are all gonna interact with AI agents via voice, with video. We're going to be able to send and receive images. We're going to be, eventually, we're going to be wearing, you know, AR glasses, that I think will become ubiquitous before we before we know it. And when we have systems like this to interact with agents that are more than just the the ChatGPT, text window, I think that changes all kinds of things about the type of service that we can deliver over AI agents and the types of concerns that enterprises are gonna need to have about, you know, how to govern those systems and then thinking outside the box. Because if you wanna be a competitive enterprise in the future, you're gonna have to embrace and support these multimodalities of interacting with AI agents just the same way that we did look. When when we're chatting over video right now, in different cities, in the US.

Scott Beechuk [00:21:08]:
We we think of that as second nature now. Well, in the in the near future, we're gonna think of it as second nature to have these multimodal experiences in AR, VR, video, real time. We are going to have devices that we carry with us that are starting to come online now, like Meta's, you know, Ray Ban, glasses that are taking video and taking audio. And then you've got different devices that we wear now that are that have all kinds of sensors on them. And all of this sensor data, all of this intake will be available to AI agents, and so we're also enterprise are also gonna have to think about privacy. Data privacy has been on everyone's mind for the last couple decades, but its meaning, the meaning of what we do as enterprises with our customers' data and how we allow them to control the use of that data in exchange for some value back is gonna take a whole new meaning.

Jordan Wilson [00:22:07]:
Yeah. I think, even even the thing with with meta, right, seems, simple enough. Right? Oh, okay. You know, you have some some ARVR type glasses that interact with, you know, llama that's it's but then you also see some new advancements. Right? Like Meta previewed the Orion version of those glasses and much more capable. Right? So it's not just about, you know, computer vision and interacting with a large language model. I think it's much more than that. But, you know, I'm wondering, Scott, though, since your your company, you know, at NorQuest, you, invest in multiple companies that are in the agent space.

Jordan Wilson [00:22:46]:
Could you give us a little bit behind the scenes of maybe what you see, coming next? And, you you know, a great example there is it's it's no longer just tax. It's multimodal. It's it's AR. You know, it's AR, maybe VR. Right? But, where are these companies kind of shifting and setting their sights on? Because, ultimately, that's going to give us indicators as business leaders where we should be focusing on as well.

Scott Beechuk [00:23:10]:
Yeah. Another big area, and I'm sure, you've been following this, I I know a lot of your viewers are probably following it, is the area of sales and engagement with customers. We talked a little bit about customer service, but I think it's very different. When you're selling a product, hope and let's just talk about a complex product, to customers. You know, there's a lot of different, again, tiers of how we do we engage with customers and we sell. Sometimes that initial tier, the initial outreach tier, we call the SDR, the sales development rep. Well, there's a lot of companies right now that are starting to try to figure out how do we automate that SDR tier. And that's yeah.

Scott Beechuk [00:23:51]:
Again, you can wrap your head around that because it's kinda like a tier one support agent where, you know, the outbound is and or the inbound intake is kind of you could you could almost like, it's a it's a bounded set of tasks. But then you go up one tier into the account executive tier where someone actually has to build a relationship with a customer, and you have to have a long running set of interactions with somebody and build trust. Now that is a way more interesting, problem to solve with AI for me. I think we're gonna go through 2 phases. Oftentimes, AI agents go through the phase of human in the loop where that AI agent is actually supercharging a human being. And if you're account executive trying to, build trust with a customer, and you're trying to, you know, get educate them on your product and your company and help them help them fall in love with your brand, you know, AI agents now can, you know, listen in on those sales calls. We have, for example, one of the companies that we invested in very early on was a company called called Gong and, again, before the latest GenAI revolution. But, even today, Gong is more relevant than ever because they're teaching account executives in real time how to sell better, how to be more effective, and how to make better use of every customer's time.

Scott Beechuk [00:25:15]:
But you can imagine where that all can go because in the future, you actually some companies are gonna adopt AI agents that do the selling and actually conduct the demos and actually facilitate multiple meetings over time with customers and build trust with customers as an AI to human interaction. And that that's that's that's an area I think everybody should keep an eye on.

Jordan Wilson [00:25:42]:
I have a question or 2 to follow-up on that, but real quick, have to give another quick shout out to our partners from Microsoft. So why should you listen to the WorkLab podcast from Microsoft? It explores the questions business leaders are asking. How can they guide their organization on their AI adoption journeys? How can the technology help them create new products and business models and maximize value? How should they help their teams reskill for this new area of work? And why is it important to be completely transparent about when and how you utilize AI? So find those answers on WorkLab. That's worklab. No spaces available wherever you get your podcast. Alright. So real quick to follow-up on that last point that you made, Scott. So as an example, you know, a tier one STR, you know, oh, that could be an autonomous AI agent and doing the demo and answering questions.

Jordan Wilson [00:26:34]:
So where are in the short term, right, I can see humans really kind of just overseeing, right, and setting these up and putting those guardrails and and more data. But, you know, past that, where are the human roles going to shift and change, and and how is the kind of, quote, unquote, human work going to change once we have, more capable and more robust AI agents?

Scott Beechuk [00:26:57]:
Yeah. Well, I mean, one area that we could talk about is r and d. Because the idea of building software is something I think will persist at least for our lifetimes. And but the way that software is built is changing really rapidly. Today, we have, you know, great products in the market like Microsoft's Git GitHub Copilot for helping writing code, and you've got a lot of other companies out there like Cursor and others, a new company called Zencoder that's also playing in the game. And these are companies that are helping junior developers and sometimes even more senior developers really accelerate the ability to get code into production. But it by doing all of this and building AI agents in in this new way, it introduces a new conundrum. And the conundrum is what happens when AI agents are talking to other AI agents, and we're stringing together a network, a web of AI agents that are solving far more complex problems.

Scott Beechuk [00:27:58]:
In other words, I ask, my AI agent system or my product to help solve a big problem for me. Hey. I want to as a as a you know, in venture capital, I want to invest in a category. I wanna find what is the best company in this new world of AI agents that's solving the the marketing problems of the world. Well, that's probably a problem that a single AI agent alone wouldn't be able to help me solve. It's probably one AI agent talking to another AI agent. So one's doing the reasoning. They're they're trying to figure out, okay.

Scott Beechuk [00:28:33]:
What is the domain that we're really talking about? What is the size of that universe? Then another one goes out and starts to do research on a lot of companies, the start ups, but they can't find them all. So they go to it talks to another AI agent and says, here are the companies that I found. Go find all of the private companies or the stealth companies that we don't even see online right right now. So that's all these strings of different reasoning abilities. But the problem is that AI agents, like human beings, are inherently inaccurate. They're not perfect. They're, again, nondeterministic systems. So in classic software development, if I'm, an an engineer, I will use a tracing ability.

Scott Beechuk [00:29:13]:
I'll be able to trace if I'm making different services, different classes, different functions, different you know, maybe it's a serverless architecture. I am actually following from API to API call exactly what's going on. The bits and bytes that go in equal certain number of bits and bytes that go out, and I can observe and I can monitor the accuracy of all of those interactions. And those are if we find inaccuracies, we call them bugs, and we set somebody to go solve it. But in the world of AI agents network together, that takes a whole new meaning because what is an AI agent to AI agent interaction? Well, there's no API, for those to talk together. Today, they may talk to one another using JSON or some other, you know, sort of technical structure, but they might not. They might actually use human lang English language. Or who knows? We've seen experiments where AI agents talking to one another invent their own language to talk to one another just because it's more efficient.

Scott Beechuk [00:30:12]:
So how do you, as a software developer, observe the interaction between AI agents in a system like that and figure out, okay. Well, there's a a 1% inaccuracy between the first two steps, a 2% inaccuracy in the next two steps, and so on. And by the time the whole system is ready to present its results, the thing has, like, gone way off the guardrails, and you're in, you know, another universe. It's not you know, the inaccuracy has gone has compounded throughout the system. So we're gonna need to develop more robust systems for monitoring, checking for accuracy, observing all these things, and then ultimately fine tuning and and and and putting the right, guardrails and iterate recursively back to fix the prompts into each one of those notes.

Jordan Wilson [00:31:02]:
I love and I I could dork for hours just about the fact that, right, these AI agents are creating their own language to to talk with each other because maybe English is is is not, efficient enough. But, Scott, we've we've covered a lot in today's episode from, you know, kind of the history of AI agents, pre generative AI. We've talked about now some some use cases, some potential problems and the promise. But as we wrap up, what is the one most important takeaway that you think people need to understand when it comes to AI agents and how they can kind of bridge the gap, to the future of work in enterprise?

Scott Beechuk [00:31:41]:
Well, you know, I think what we need to do is we need to think about working backwards as we build companies. So I I'm gonna take the the perspective of the the the founder the world of founders who are building, startups. And so what the the opportunity is to make life far more productive for all of us humans here, make it way more efficient for us to get things done, and that's what technology has always done. But oftentimes, we make the mistake of working forwards instead of backwards from what we want. If we work forwards and we say, well, somebody dropped this cool, you know, LLM on me. Let's go see what we can do with it. That's a great academic experiment, but those taking that path doesn't always lead us to the best conclusions. What I think we all need to do is we need to really think carefully about our own businesses and say, what are the things in our business that we would love to solve, that we would love to automate? Maybe it's in the front office.

Scott Beechuk [00:32:40]:
Maybe it's a back office, finance type of of problem. And then if we work backwards from the the the the the goals that we have, we will have clear business cases, and then we can go out and either build those systems, you know, with knowing the the intended result, or we can seek out solutions or partner with other companies because I've never seen so many new companies getting formed right now, especially around the area of AI agents. So there's gonna be a lot of incredible technology and incredible products coming out, and I think a lot of those companies are gonna need to work together so that we can actually orchestrate AI agents from one company to another to solve our more complex tasks over time.

Jordan Wilson [00:33:24]:
Mhmm. It is definitely a hot space, and it is moving very quickly. But, Scott, I think you helped us all understand it a lot better. So thank you very much for your time and sharing your expertise on the Everyday AI Show. We really appreciate it.

Scott Beechuk [00:33:40]:
Thank you, Jordan. I really appreciate it too.

Jordan Wilson [00:33:42]:
Alright. As a reminder, y'all, we covered a lot there. I know that the AI agent space and how quickly it's move moving can be very confusing. So don't worry. We're gonna be breaking it all down, all of the best insights from today's, episode and a lot more on today's newsletter. So make sure you check that out. If this was helpful, don't don't be greedy. Share share this with a friend because we all need to understand where this, you know, AI agents, entity is heading because it's fast.

Jordan Wilson [00:34:12]:
Speaking of fast, you're gonna learn fast if you go to your everydayai.com. So thank you for tuning in today. Please join us back tomorrow and every day for more everyday AI. Thanks, y'all.