Your 2025 Starter Guide for AI Agents - AI for Business - EP 018

00:00:00: Speaker: Hello and welcome to This week's episode of the AI boardroom. Yeah, I think we should, it's not really weekly.

00:00:14: Speaker 2: It's a, yeah. This month's update on AI agents which I'm really excited about to talk about.

00:00:22: Speaker: I think we will try to bring it. Pick up some pace here. But yeah. For all of you listening, thank you very much for tuning in. And today we will talk about 2025. And I think Svetlana 2025 will bring us the year of the agents.

00:00:41: And with agents, we mean like proper agents, like AI deciding to use stuff and create stuff and make stuff happen without a constant human input. What do you think about this?

00:00:55: like the companies who embrace AI so who are not afraid in jumping in and actually developing and getting the value out of them.

00:01:08: They're going to win huge. And I think this is where you're going. But the companies who are still deciding is generally I, for me, risk averse still doing prototyping and not really sure when to release these solutions into the wild going to be limiting kind of the use cases and like the value gained from these systems.

00:01:26: But the companies that do embrace Gen AI and like these agents and they in their processes whether it's Supportive functions or their core part of the business are Going to win big because these models are maturing these architectures best practices are maturing. They've been in the wild being developed refined By a lot of the companies.

00:01:48: Embracing AI, if you've started to tackle those use cases, you've got to put them out into production and start getting value from them, because I think you're right, you can win big, because there's a lot of value captured in these systems that you can explore, exploit fairly easily.

00:02:17: Speaker: Yeah, the interesting part for me is now we like local models or like open source models or open weight models get pretty. pretty good. And actually China is leading the way with new models, which are capable of reasoning. They are not perfect. And even if they suggested with the benchmarks, it's not like open AI or one level, which is like the only reasoning model we knew up until like last week.

00:02:45: was questions And then you have the DeepSeek R1 and I think you have a third one somewhere there.

00:03:07: But yeah, they're all from China and they all work really well. They open source that stuff. I'm still eager to test it. And one of the models, I think that's the question is 32 billion, which runs perfectly fine on my Mac book. Yeah. Reasoning capabilities. And just to put into context, if it runs on my MacBook Pro, which is a fairly capable machine, but nonetheless, if it runs there, then you can run it like dirt cheap everywhere else where you can run it.

00:03:39: That's interesting. And especially if you think like how much you pay for, Or one because it's pretty expensive right now. And it's actually, it's expensive and you use a lot of tokens. So it's yeah. But yeah, that's I think this reasoning capabilities are going to be key for agents to do better work.

00:03:59: think a lot of value is gained through, as you mentioned, knowledge for knowledge workers, right? A lot of the reasoning capabilities are needed to create documents, create strategies, do research. And I think even with just even out of the box solutions, if you're embracing that, even building agents, there's so much capabilities that could be extracted through these search based tools.

00:04:25: I think knowledge workers are the ones that benefit. That was actually speaking to someone last week. And one of the things that came up with was like, who is truly benefiting from these systems? And you hear a lot about these wins and the value gains, and there's research that's published that.

00:04:42: as your kind of Coaches to help set those standards.

00:05:04: Either you delegate to them tasks that they have to go and figure out on their own. It's going to take them X as long and then hope that they find mentors to guide them on the right output or you give them access to these agents or tools that have these guidelines and these best practices embedded in them.

00:05:23: And that helps them up skill. And then do their work even more efficiently. So I think the reasoning capabilities are really groundbreaking for knowledge workers, especially again, we have new graduates, interns entering the workforce. And I think companies are reluctant to hire sometimes like the inexperienced staff, but I think they should be embracing them if you have, like, all the right tools in place.

00:05:46: produce. I'm a big fan of them.

00:06:01: Speaker: Yeah, I always I always compare it to a teacher that has infinite time and infinite patience, like if you think about like everyone who started in this first job, like back in the day, like his first professional job, like I don't talk about like summer jobs or something, but if you have like your first professional job, you mostly have some mentor or someone who's guiding you, who's assigned to you basically, and At least in my experience, like there's a lot of friction because the one who's assigned to you, oftentimes he doesn't have time or he's out at the customers a lot or something like you don't have him like at your disposal every time you have a question, which is not a bad thing because like figuring out stuff yourself is actually a pretty useful skill, which we should still embrace.

00:06:45: to like you, you want to, you wanna argument the process and not replace it.

00:07:05: So because, but I also think for I, I have a lot of experience now with using AI for coding and. I would love to have more agentic behavior. And the more agentic behavior I have, there is some agentic behavior already out there in coding. The more natural it feels because now I feel like. I just only I only go in there and do like product descriptions and it codes all the stuff and I love this.

00:07:33: Like I really I can't even state how it, and even there's like small things you have two AI tools and the one is just showing you the output instantly, like renders it out while the other you have to switch to your brother, even that little bit of friction like breaks the immersiveness and I had this, was it on Monday?

00:07:53: had a good case and it just worked. Like I build it out or was it even an hour? Like really worked pretty well. And I built like in an hour what I've, I would have needed a week back and figuring out a lot of stuff along the way, which I never would need again just for this one project.

00:08:17: And now I just drop in the knowledge into the AI and the AI just makes it happen. And that's really beautiful.

00:08:23: Speaker 2: And I think with just setting realistic expectations also, I think our audience may wonder, okay, this is great that we've heard about the benefits and it's awesome.

00:08:32: We've heard some of these things, but like, how do you choose the right use cases to apply these agents, right? So it's not just about putting technology for technology sake, just because it's so beneficial, but you have to apply it in the right places. And also with the right mindset.

00:08:46: Speaker: So it's you have to acknowledge this is like a new type of doing things.

00:08:50: factor. And we had to first learn how to use it properly. Like just remember how they took six, seven years until like everyone And the websites caught up to be more mobile friendly.

00:09:10: It's still like still up for a lot of them. It's still out in the open. And that's what I always try to emphasize this. There's a paradigm shift and we have to embrace the shift to make it useful for us.

00:09:22: Speaker 2: And I think with maybe going with The elaborating on the mindset, I think bringing these solutions you have to embrace, and I think you and I talked about it privately before this recording, but when it comes to Gen AI solutions, they are prone to hallucinations still.

00:09:41: code into this AI system and then you knew what was wrong and you knew how to exactly address the problem that gen ai.

00:10:09: Kind of the situation you put them in, but without that proper kind of maybe experimentation that you've done and like the experience you had, you would have been such so much more problematic for you to address those issues. So I think it comes down to you become. AI will do majority of your work, especially if you're using it for supportive functions for knowledge workers to use, but there are some things that you still have to upskill your teams to decipher.

00:10:36: They'll do majority of the work for them, but it's really behind the human to evaluate the final output because. It's you who is behind the wheel and not AI completely. So you can't delegate 100 percent of it. You're delegating a portion of it. That's where the efficiencies and enhancements are happening.

00:10:53: see if it works. And sometimes it's like actually reading through the output and evaluating if it's Fits what you're, you've asked it to, and then does it pass the sniff test?

00:11:11: Speaker: I'm shocked that, the, what are the IDE.

00:11:14: Speaker 2: I used to, no, I took a course in in programming. I'm telling you I, I know enough to be dangerous about coding, so I can,

00:11:22: Speaker: That I heard already from you. Yeah. No I want to I would love to add to this, that it's like, It starts to show that if your people get the gist of it they start getting creative and the use case problem solves itself.

00:11:37: whatnot.

00:12:00: It will be the same, like everyone has to be able to use this new technology. And it will be a lot faster in an adaption or the adaption rate needs to be a lot faster than it used to be.

00:12:13: Speaker 2: I think one, one thing I just want to emphasize it, but I think with what you said that I don't want to underscore, but domain expertise as you rethink these workflows is so important.

00:12:22: When you design, like you can't delegate a process, like improving a particular workflow or a process with agents. If you don't bring the domain expert who's actually executing that process. Into that design picture like they have to make sure that whatever you're designing or developing will actually fit the use case or have someone on the team that is a proxy or.

00:12:43: adoption will suffer because people will be like, Oh, it doesn't fit my use case.

00:13:03: It doesn't actually fit what I need, needed to do. So a Gen AI or a Gen Tech system for, let's say, a knowledge worker that works in HR versus a coder, completely different use cases. And you can't just say Oh Jenny, I will or agents will figure that out. Like they're intelligent. You have to optimize these systems to specific use case and you have to validate with the domain experts, SMEs or whatever, who actually perform those functions, which you target and they will bring them along on the journey as you build, but then also help them have them validate what you're actually building.

00:13:36: I think that's a big failure point is like where organizations like dedicate these. solutions to like these disparate silo teams and they don't know what to do with it. They're just build the technology with some assumptions. They deploy it and then they wonder why it hasn't been adopted or why they the users are not happy with the output.

00:13:57: to any implementation, whether you use the domain expertise during the design process or like that, that requirements process, You have to bring the domain experts who are representative of those workflows or that specific use case in helping you build the solution and actually validate it.

00:14:19: Speaker: Yeah, definitely. So let's dive deeper into the agent topic. Microsoft had this Ignite event last week and they showed something they call co pilot actions, which I found interesting because it says I need to do this task every week. And that's basically the goal of the task.

00:14:42: most you will the most easy way to get people into building stuff out.

00:15:04: And they have the copilot AI engine on the backend and it's just working for them. In their, like, how do I say, in their frame, the frame that Copilot builds for them, basically. And I found it interesting and Maxfield also added the Copilot Studio agents, like they call it agents, which have now also the option to be triggered autonomously.

00:15:30: And they start to roll out stuff in the business suites, which they built out of the box agents. For example, for order management, like you get emails, it's connected to email account and automatically creates an order and just says Hey, this is what I got from the email. This is what I created. Is it fine or not?

00:15:46: important is human in the loop, like this whole notion of the AI doing autonomous work, but it's able to ask your stuff if it's not certain or if if there's some critical, but for example you create an order and before you like post this order or commit to it, you might, But Want to have the option to check the order and actually check what the customer is getting.

00:16:28: Like you don't want, at least not now to have this be like completely with you without human intervention. And and I think that's something everyone should Like if you plan on a AI agent systems in some way, shape or form being like autonomous or like trigger regularly every day, like on their own, think always about this human, the loop part, because that's like an integral part of getting this reliable.

00:16:55: more complex your agent or the more complex the processes that you want to automate, the more you have to make sure that what you give on it and instructions has like the least amount of ambiguity possible. Because people talk hallucinations and might be a bit philosophical, but I'm like, if I take a random human, I don't give him the right information.

00:17:26: just guess, which always got me furious.

00:18:04: Because it's I'm not paying you for guessing stuff. But that's a different story. What I want to say with this is is hallucinations really a problem of the models themselves, or is it more like a side effect of us not using them properly?

00:18:19: Speaker 2: Big instructions. You mean?

00:18:20: Yeah.

00:18:20: Speaker: Yeah. Because language is like the worst. Think to have precise extractions. And yeah, that's what I'm always like, because look at every prompt ever built, there is still some way to interpret this the one way or the other. So yeah, you have to get this I call it value of ambiguity. You have to get this as small as possible, but because you don't get this out of the system.

00:18:44: You have to have this human in the loop as an integral part.

00:18:48: system? It's a CLAW, Chad GPT, and yeah, I totally agree if you put in a vague prompt and you're basically it's a robot, right?

00:19:08: So like when you tell a robot to do something, it's like very deterministic. Like you told me to answer your question, I am going to answer it. Whether I like it or not, I'm going to make up some things, if my life depended on it. Yeah. Cool. That's where hallucinations come in, because if your question is broad and it's determined to give you an answer and it doesn't know what it's going to make stuff up because you told him to to give you an answer, but in use cases where, there are custom solutions that you're building, you can actually like, let's say custom or customer service agents.

00:19:39: If you're talking about voice agents when you can put in guardrails and specific instructions for when you don't know and resist the temptation of making stuff up. If it doesn't exist in your knowledge base, do not respond or tell them like to try a different prompt or to try. A much more specific answer.

00:19:57: if you're thinking about out of the box tools or even like thinking about like agents, there's probably limitations for like guardrails. That's always the case. So either you go out of the box, which is quicker to market, quicker to implement. They're probably cheaper as well, but you have limitations on how much customization you can make to those workflows, but then you can go the custom route.

00:20:20: A little bit more expensive, more time to actually develop, but you have so much control. You can actually design the system to do what you need. You can apply it to specific data. You can apply more advanced techniques that are specific to your field. You don't get that flexibility with out of the box solutions.

00:20:38: So it's always like a build versus buy analysis for what is it, That you're trying to do in what time frames and then like how much budget you have what are you trying to get out of this, you always have to measure it against some of the metrics and objectives you're trying to set and then figure out what is the best route to get there.

00:20:54: boxes, but sometimes it doesn't. And it calls for actually bringing in some experts to build those solutions. And we can there's another episode that we can probably do about how to build and prototype and we will determine whether it's the right even.

00:21:17: Use case and like even project out some of the costs and the timelines. Yeah, we can talk about a little bit more, but just wanted to mention that is also a route that gives you more control.

00:21:27: Speaker: Yeah. So for from what I, how I experienced it, like I use out of the box solutions to, to verify some idea that I have with AI because building it out in a complex System from the get go is mostly not suitable, like just from a time frame.

00:21:44: it's fine. You can just try to prompt and no conversation of matter, get to your side output.

00:22:06: But one thing I wanna also emphasize is that you don't get anywhere close to the benefits without the box solutions compared to custom custom made solutions. With one exception. And I think that's still out in the open. Can we build. Out of box systems that are customizable enough to be good and to be highlighted.

00:22:33: And I think Copilot is What Microsoft does there is like this out of the box solution, which can be deeper integrated and which can be customized. Yeah. Meanwhile, like with the new tooling they presented, the customization options will be really versatile. Yeah I have better hopes for copilot.

00:22:50: strong partner with OpenAI,

00:23:00: Speaker 2: I think this is maybe going back to who gets it first and who will benefit. I think just going back to Microsoft and probably Google Workspace.

00:23:10: Like those are enterprises already. And both of them are enabled. One with Gemini, the other one is Copilot. Let's. So if you're still exploring the space and trying to figure out what value can you gains that is like a low hanging fruit for your organization to just subscribe to the AI or the copilot versions of the system and just upscale your teams on how to properly use them.

00:23:33: And then you can have if you're a Google workspace type of company, like there's Gemini that you could use across different systems, like whether it's Google. The sheets, the word documents and things like that, like the presentations, the slides, like you can use a lot of it and then build that up that expertise.

00:23:50: some ideas and oh you know how I wonder if. This could be solved with, let's say, a custom solution, but you have to get that experimentation out of the way to actually build an idea, like a mind map in your head okay this out of the box solution or what Jenny is really good at, and then maybe not so great at, and evaluate it.

00:24:20: But yeah. Whether that's like an unmet need within your organization that can be solved and most likely could with a custom solution. So if it's hallucinating, or maybe you need something that's more customer facing and things like that, probably these out of the box solutions that are built for enterprises are more internally facing.

00:24:39: So those are great use cases for custom builds. And I'll say, Another thing that kind of why it makes me think like probably the more predominant use cases are really internally facing, but that's not where in fact your value. And I think you say this a lot too, is like you're sitting in the pot of gold.

00:24:59: like, how can you actually take that data and then make it usable to. Capture more revenue, build better business models out of them. So that doesn't come from out of the box solutions. It's actually like building custom things that utilize your data that you can do something with to increase the value of your organization.

00:25:17: And there's different ways to do it, but there's been again, some research that has been done that majority of the companies who are. or the leaders who are applying AI in any of their operations, they're actually applying it to improve their services and operations that are capturing more value.

00:25:36: ROI on those investments.

00:26:01: So those things that you should evaluate for custom solutions and things that are more internally facing HR, think out of the box types of things. Those are lower costs, but easier also to demonstrate like quick value, but also the ROI. You're not going to get as much ROI out of support agents that you build that are again internal operations and things like that.

00:26:27: So that helps someone. Yeah. And if you want to hear more about like some of these best practices and like things that we're finding happy to do another kind of focused episodes on use cases that have higher chance of showcasing ROI. Maybe also

00:26:42: Speaker: what role. Data place into a specific use case.

00:26:47: key milestone for me, like to understand, okay, data and examples and best practices and stuff.

00:27:05: And that in the right way. Combined to prompt, which is dynamic, which is suited to the request and the situation and the context that the result, but that's highly valuable. And and that's what that's what, that's why I always say you're sitting on a put of gold for example, if you're a marketing agency and you have data about what ad copy worked really well and what didn't.

00:27:31: basically I don't know what, when the last time, when was the last time I actually looked at docs, I mostly like just load them into the AI and say please do this, like the docs say if it's some, something that's really interests me and that I might benefit from doing it myself.

00:28:18: I still go in there, but most of the time because that's the thing, like these whole frameworks and code, the whole coding world is changing so much. And it's the same for marketing. So for sales, like circumstances change there are certain paradigms that work. These are basically the your best practices that's.

00:28:35: What you experienced, you can program this into your AI, and then you have to have be able to get the right data for the right in the right moment. And that's where the tuning gets in. And that's where you also build out agents that have access to different data sources and functions and can execute stuff on their own merit.

00:28:55: word actually comes from. They get agency to actually decide what function to use when. And yeah, and there's a lot of tuning involved and that's why custom made solutions at some point deliver the most value because you can tailor this like little details that simply matter.

00:29:17: It's the same, like if you have your own sales team, which only sells your product all the time, every day, or you have an external sales team. Yeah. Which is good at what they doing, but they're not your people. Yeah, I think that's

00:29:30: Speaker 2: I'll give one, maybe other example that I think you can get in marketing pretty good ROI, cause you want to keep probably the This type of data proprietary, so you don't want to share with an open source or, open out of the box type of tool, but think about brand and creative kind of brainstorming that the marketers do.

00:29:49: secret sauce of. Your company like this is something that you want to keep internally, probably on premises. But use the generative kind of creative capacity and the reasoning capacity of Jenny.

00:30:12: I to come up with these more innovative solutions that are hyper. Targeted towards the persona and the brand guidelines that you've described. So you're creating these again, guardrails around like the brand guidelines, best practices, risk and things like that. But it'll do majority of that work for you for those guidelines.

00:30:31: And it's something that you can use repeatedly and you can AB test different outputs and see what, what can work. So something that would have taken a creative team to brainstorm I don't know, over a few days and test options, you can come up with hundreds of these scenarios that you can test that small, bigger scale on a daily basis.

00:30:52: just testing with which messages that better resonate. And then what is the highest quality that. Pulls and converts the right audience. So I think that there's a lot of power again, in capturing more, most of that value from, I think, and being able to demonstrate tangible ROI behind some of these solutions.

00:31:18: Which makes it exciting.

00:31:20: Speaker: But this will be fairly generic if you don't feed with your data. Correct. Yeah. Okay. I think. We're that closes it for for this week's episode. I think, yeah, we should definitely jump on the next episode and yeah, talk about the use cases and the role of data a bit more think, yeah, that might be really helpful to y'all.

00:31:40: Yeah.

00:31:41: Speaker 2: Maybe one call to action. We noticed that a lot of folks who listen to us, hopefully this is informative and helpful, love to get you to subscribe to our channel because that gives us some validation that what we are talking about is valuable to you and it encourages us to do more of these episodes.

00:31:58: pressing the subscribe button we'll totally appreciate it and I guess we'll see each other next time.

00:32:04: Speaker: Yeah. And maybe on, on that merit. I also try to go on, go live every Friday, so you can just like, yeah, bombard me with your questions. And yeah, because this is the podcast is always like asking and it's like, It's in its nature and yeah, have having have the option to ask questions live.

00:32:23: Yeah. Give some alternative consumption methods. Yeah.

00:32:29: Speaker 2: Totally valuable. Alrighty. Thank you all.

00:32:33: Speaker: Thank you. Bye.