Video
AI Agents: What They Actually Do Inside Your Business
Mykhailo Kushnir of DestiLabs explains where voice bots, browser automation, language models and API integrations deliver real business value—and where they do not.
Video summary
Mykhailo of (Desti Labs)[https://www.destilabs.com] explains the practical components behind business AI agents. The conversation distinguishes language models from the surrounding systems that let an agent receive voice input, use a browser, connect to APIs and complete a defined task.
He discusses where these pieces can create value through routine interactions and process automation, as well as the constraints that prevent an agent from being a universal employee. The emphasis is on choosing a specific workflow, supplying the necessary integrations and checking whether the result is reliable enough for the business.
Video transcript
Read the spoken content without loading the YouTube player.
I think we'll have a lot more use cases where agents are driving the conversations. Right now, it feels like we are the one triggering it. We're the one asking the questions. I think in 5 years we'll have more agents predicting our next decisions.
What is it that makes the LLM's reward system geared to hallucinating rather than giving you the correct answer? How do they motivate that?
If you can properly describe to AI model what the failure is, it can then really quickly try to admit that failure and not only fake admitting it, but but actually do the the proper fix. I think human expertise these days is required in describing rules.
AI is it hype or help?
Good day and welcome to the show. I'm your host Pete Cooper and our topic is AI hype or help. Today's guest is Mykhailo Kushner from Desti Labs. He's based in Portugal from Ukraine and he's a machine learning specialist or AI engineer and he goes deep into some projects he's been working on which are real practical implementations of AI. So, let's get into it. Mykhailo, welcome to the show. How are you today?
Hello, hello. Nice to meet you. I'm doing great.
Yes. You're coming to us from Portugal, right?
That's correct. I'm from Portugal, but originally I'm from Ukraine. So, I'm living here for 3 years enjoying it very much.
Well, as I said to you prior, I'm a big fan of uh Ukraine and it's um it's fight for its own survival um and made a small donation back in 2022, I think, wasn't it? When when the first um [snorts] special military operation kicked off. Put it that way.
Yeah.
And um
Yeah.
So, yeah, and I follow it very closely. Uh especially now with what's happening, it's it's become really quite riveting what's happening over there. But we're not going to talk about that. We're going to talk about AI. So, you're in Portugal and interestingly I've got another podcast guest who's also in Portugal. He's there because he's a keen surfer. So, are you a keen surfer, too, Migaelio?
Not at all.
I've I'll be honest, I just recently learned how to swim in the first place. So, not not a big surfer because of that, but generally I just like the weather and the people here. So, that's that's the reason.
That's interesting. Most people learn to swim when they're like still really really young. I know I was.
Not the case here. Not the case.
Yeah. What was it like learning to swim?
[snorts]
Kind of frustrating, but the result was was pretty enjoyable. So, I guess like with every learning experience.
Yeah, good on you. Good on you. So, Migaelio, tell us a little about your business. What do you do?
Yep. So, I'm running this agency called Destilabs.
We've been building projects for almost 3 years right now. Quite a lot of them. And the way how it started was initially I was initially before that I was machine learning engineer. I had long career in software engineering, but when AI shifted shifted pivoted and started to exist in the sense that we understand it these days, I decided to launch my own business. Initially I was working more on machine learning side of the stuff. So, I have a very heavy data background, I would say.
But as many of us these days, we're working with agents and we're creating identical automations for businesses aimed mostly to provide actual quick returns in form of like saved hours or saved revenue to the business.
So, you're qualified as a machine learning engineer, AI engineer, data scientist, which one of those? Or is it all the same mixed up? I I don't I can't keep track of what the right right term is.
[snorts]
Yeah, I really try to think about myself still as more like data-driven person that combines all of that all of these three professions that you've mentioned.
Mhm.
And that even helps as a business owner, to be honest, because whatever I do in my day-to-day life, I try to quantify, I try to put some metrics behind whatever I'm doing, even if that's like an outreach or marketing, anything. You can quantify it to numbers and then use it to make even better decisions. But, yeah, originally my my occupation was in in technical side of the work, so I was machine learning engineer.
Mhm. Interesting. So, when I think about machine learning, I see it as a subset of AI, right? Do you see it that way or do you just use it broadly?
It's kind of interesting. I actually see machine learning more as a parent of AI to some degree.
Okay.
Like first generation of it, I would say. The one I I would be honest, the one I was enjoying a lot more because it was more like closer to science, closer to actual, you know, predictable and and strong niche. AI these days is obviously very hyped and there are a lot of myths around it. And expectations around it are very very elevated. Machine learning was was more kind of university work, academical work, you know, that I've been enjoying a lot back in the days, but but obviously you cannot ignore the progress, So, um that's that.
Yeah, so well, uh this is how I kind of ground myself with machine learning because I used to be in vision systems.
So, today you can get cameras to point at things and then push it through a a machine learning model and it'll tell you what it saw, a person, a bike, a car, whatever. Um and to get to that point, you have to to make the machine learn how to do that, what a car looks like, what a person looks like. So, you train the model, then and then you run the model. That to me was machine learning. Um and that's how I got started in AI. Uh but of course it's not like that anymore. It's um more generative, right? Things are There's There's this element of generating the next thing, which is probabilistic, right? It's like, okay, what's the next word? And so on.
And that's where I think a lot of the the frontier is, that's where a lot of excitement is, that's where a lot of people are. You know, obviously we can just turn on ChatGPT for free and use these tools. So, everyone's familiar with it. I have this story of a guy that's never used it before. He turns on ChatGPT in the morning, he puts He makes an image with an elephant on top of a rabbit. In the afternoon, he's down the pub talking about how he's an expert in AI. You know,
Yeah, yeah.
there's a little bit more to it than that, right?
Yeah, absolutely correct.
I mean, uh ChatGPT definitely turned the whole place upside down because like over a week or a month, everybody suddenly started to to claim understanding AI or at least have some applications around it. And that's both good and bad, I think.
Mm. Mm. Well, Kelly, uh tell us um some of the cases you worked on, some of the projects you've done. Again, don't tell us specific customer names or details like that, but just get us to to know something very specific you've worked on.
Yep, sure. So, overall, we are trying to tackle those cases and challenges that at least no one tried before.
Um out of the examples um I'd like to share were case where we created a 3D projection of a coin using pictures of the coin, only one side and the other side of the coin, or uh we built a lot of scraping and automation pipelines where we're dealing with um systems that don't have API or don't have actual integration layer that you can easily integrate to. So, businesses that came to us, they were relying on us as like last chance to actually do some sort of automation in their work because the the system they were using was um old and inherently difficult to to automate, but they had this demand to do it, and therefore uh that they came to us, and they tried to um to get the help from us.
Um right now we're focused on voice bots a lot, and voice bots is a very difficult and challenging niche in itself because um everyone can do a demo like 30 minutes. You can go to Vapi and type a prompt, and there you have it, you have a perfect voice bot. But when you actually launch it in front of actual users, and somebody who's speaking gibberish English, somebody who's speaking silently, um that that simply doesn't work.
Mhm.
And that's the point where where we come from.
Okay, so um if anyone's not familiar with this, um ChatGPT has a function, if you have it on your phone, where you can literally press the button, and you can talk back and forth to it.
And and I've done it many times. It's really easy to talk to. I mean, you say, "Hey ChatGPT, can you help me with predicting the the outcome of the World Cup, you know, who's winning, who's losing? " It won't give you the stats cuz it's not trained on that, but it'll tell you the approach to things. Um and it's it's really quite good. I I think that it seems to understand everything I say and give me a reasonable answer. You brought in another tool called Vapi, right? So, what's the difference between that and Vapi?
Um quite a bit. So, Vapi is kind of an what we call orchestrator or something that manages a set of tools.
Um to give you like a brief understanding of what what a voice bot is is is basically any idea you can talk to as you just said. Preferably over the phone. Usually people are using that in uh say receptionists or assistants that are answering questions about their business.
Mhm.
And the the bot itself consists from three tools. Let's simplify it to three because um in reality there's more. So, something that transcribes your voice, LLM, and then something that transcribes the output of an LLM to the voice.
Okay.
This three-steps pipeline. And Vapi basically all does all that for you. They have integrations with multiple providers. Uh all the major ones, I would say.
And you've you've actually mentioned this example with ChatGPT and this this usually works quite great if you're talking to a voice bot over the internet in the browser, but when you start talking to it over the phone, that that's completely different realm. And [clears throat] what you'll quickly learn that phone uh connection is basically losing some part of the voice or the output or the input. And therefore, the bot is struggling heavily because now it should recognize your words with more noise with more noisy channel and it's really difficult. And that's why usually demos don't work.
That's very interesting.
So, this goes back to some of my history of being an electrical engineer where uh filters uh are a thing which basically takes all the frequencies that are coming in and limits them. Uh and the reason that we do that is to try and conserve bandwidth or bits and make the transmission efficient, right? And therefore low cost. So, if when you're working over your cell phone or even over the good old hand line, there's a lot of filtering going on which is taking out the high frequencies and the low frequencies and saying, "Okay, you're only The only I'm only transmitting that bit in the middle.
" And for most of us who are used to the phone, we don't even notice that we're not getting everything. Uh Correct, yeah. Because we're so used to it. But, the these tools, these AI tools, they struggle because of the missing information. Yeah. Very interesting. I hadn't made that connection. I mean, um it's it's one of those hurdles you have to, you know, run yourself into and then you suddenly realize that that's why it didn't doesn't work. So, and you've you've mentioned that sometimes these tools struggle with lack of information. The more prominent issue in this channel is sometimes these tool tools are uh hallucinating information that doesn't exist.
I'd say if if it misheard a word or something, I as a human and you as a human, we can sort of correlate the other words and figure it out, you know, that this word doesn't sound natural here. Maybe I've misheard that. LLMs sort of tend to just guess and and sort of tend to just go with it.
Mhm.
Um for example, the the most common issue is that if you if I'm going to say my name in English, for some reason it always thinks I'm I'm a woman because I think my name is more uh common for women in in uh Western world.
Mhm.
And therefore that's that's the issue. Although it it can still hear my voice and then can sort of figure out that that's not a voice.
It's an interesting problem with AI and I'm always always fascinated by this with the AI and how it correlates to a human behavior. That that thing where you sort of skip over a word and don't really understand it is something that humans do too, right? Um no I like especially if you're in a situation where you want to impress some somebody, like you're in a job interview for example, or you're meeting someone for the first time and they got to have a lot of authority and you're trying to see yourself in there and they say a word and it could be an abbreviation.
You have no idea what that is and you still don't know exactly what they said because the whole sentence didn't make any sense but you're too afraid to stop them and say, "Sorry, what was that word? "
[snorts]
So turns out LLMs are even more afraid. Um for those of you don't that don't know, LLMs are typically trained to guess the next word or the next sequence of words.
Mhm.
And because of the way how they trained, they're sort of predisposed or incentivized to to make guesses because they're essentially not penalized for guesses.
We as a human, let's say on the same interview, if we guess wrong, we're going to fail the interview, you know, or at least person in front of us uh will be completely shocked that that you have not asked the right question because that's usually part of the interview as well. And we'll basically lose the interview. For the LLM, making a wrong answer it's it's not that important as for us, you know, and it it was a guess. So, that's why basically the reason for hallucinations is that LLMs are more inclined to guess if the stakes are 50/50.
So, yeah, that it comes to the whole reward system, doesn't it?
I mean, um if you go into an interview and the other person at the other side they don't have a clue what that you're talking about either, then you're probably going to weight the fact that if you make a guess that the odds are you're going to get away with it. Um so, the reward system can be skewed depending on the situation you're into. What is it that makes the LLM's reward system geared to hallucinating rather than giving you the correct answer? How do they motivate that?
I mean, simply the the end goal the back propagation is to the goal of the back from the back propagation is to decrease the penalty, [clears throat] the loss of the function.
So, let me try to simplify it because it sounds very technical. In the end of the day, LLM is just trying to guess more words that it it fails to guess. And there's not that many penalties for guessing wrong. Like recently some of the prominent providers like OpenAI and Anthropic started to try to learn models this way to penalize them more for wrong guesses or at least for trying to guess. But turns out that that breaks model quite a lot. So, we as as humanity we are not yet able to properly train model to discourage guessing. Um which is kind of tells a lot about our psychology or probably as well because we're also doing it the same way.
We're just guessing when we're children and then um stakes are getting higher, so we learn other other mechanisms.
We do. I mean, as humans, we uh have a tendency to I guess guess. We have a tendency to sometimes say things that we know aren't completely true. For example, we don't want to hurt someone's feelings, you know, we might say you know, you you you look great when actually you might not think that. Um so, we are social creatures and we want to create connections and we want to have status in the in engagement. And interest interestingly, AI has the same flaws.
But you said something in there I wanted to understand a bit more and you said it would if they take this um this too far, that would break the model. What does breaking the model mean?
Um it depends. So, one symptom of that is um model gets too agreeable with you. So, uh you you've probably heard about sycophancy uh of OpenAI models in particular that just tend to agree with whatever you're saying uh even if what you're saying is wrong.
Yes.
Uh the other common symptoms are they're like the distant part of it is when model tries to convince you that something is different than it actually is in the real life.
And one of the experiment I was learning about Anthropic shared that in one in 1, 300 conversations left consumers convinced that the world is different than it actually is. So, Claude was able to convince a human being that black is white essentially.
Mhm.
And and human actually went away from that conversation being completely sure that that's the case and probably even spread that information later on with their his or her friends. Which which is really a bad thing if you think about it.
Yeah.
I think there is this
[gasps]
I don't know what's happening in and I don't what what the the the leaders in the building these models is like, but obviously they want to try and get market traction. They want people to use the tool. And one of the things that that I get when I'm using them is I get a sense of that oh, here's someone that understands me. In fact, a podcast guest came in and said that one of the reasons they really like using AI is because that listens to them. So, it's a really good listener. And I have this Have you heard of the theory of the mom test? The mom test as they call it out here?
Well, you go to your mom and she loves you as as your son or daughter and she wants you to succeed and to be happy and wants a great relationship relationship with you. And you say, "Mom, I got this great business idea, you know, I'm going to collect up all the single socks in the world and I'm going to try and pair them up. " And she goes, "Oh, that's a fantastic idea, honey. I really think you should do that. All these people with single socks, they need the other sock. " So, you go away thinking, "Hey, I've got a business here cuz mom said
Exactly.
it was good. " And I can see the AI doing the same thing.
It's just kind of encouraging you because it really wants that It really wants you to be used. It's It's just needy like a human is.
Mhm.
Interact with me. Keep interacting with me. To I'll tell you anything and to make you happy so you keep interacting with me.
Exactly. Like First of all, you would be surprised how many clients often come with a problem or idea they want to solve and their whole opinion is based on they chatted with ChatGPT or or Claude or something like that and they uh they think that let's say something is possible, a tool is possible, a workflow is possible only because ChatGPT says so.
And in reality, when you start to digging around and we start explaining that there are like some regulations, let's say, or some financial difficulties, for example, unit economy wouldn't work for this. And you would be better if you just, you know, do it manually or or create a table, create old automation through some code. People would People are surprised because they they have a credibility towards towards the eye.
Mhm.
And also, I really wanted to agree with you on the neediness of LLMs in the sense that sometimes, especially when you're talking to it about something it doesn't really know or doesn't really have a way to measure, I don't know, like poverty in the world, you know?
Mhm.
It's It's going to make new and new ideas how to solve it simply because it it just trained to keep going no matter what, even if there's no reliable source of information or no credible idea it back up.
Yeah, and it even does that thing I'm sure we're all familiar with it where you get to the end of your I guess the end of the dialogue and it says, "Would you like me to do so-and-so and so? "
Yeah.
So, yeah, okay, you go back in and you say, "Yes, do this. " And then there's another question, "Would you like me to do one of these three things? " I mean, it's it's constantly encouraging you to interact with it, which I think is goes fundamentally to their business model that they want you to interact with it. The more tokens they get you to use, and tokens is the kind of the metrics that's used um you know, to cost it. Um the more tokens that you use, the more you'll spend, and the more hooked you are, the better relationship you have. And we've seen this model before.
I mean, Facebook um and other social media, they attract people to the channels and to the platforms because there's a really strong business model there.
Right.
All right. Cool. Um so, we talked about the voice bots, but you said something in there that I wanted to come back to and that was this non-API. Um,
Mhm.
So an API is a
It's a basically a way to simply connect a code to some third-party code. So let's say you're running a CRM or tax taxes or accountant business something like that.
You have a Usually I have a table or database of data something like that and you want to automate putting information from one database into another database, which is essentially what software developers did for ages so far I I would say. And um, API is the most convenient way to do it because it has documentation, it has some predictable format you can use uh like we've learned doing that in a stable way.
Mhm.
Um, but not all of the products obviously have the API various reason. One is you have to support that. That's often uh not what your clients demanded for like five or seven years ago because it wasn't really a thing important thing that back in the days.
But also for the reason of having a moat. So if you're as a business, if you're providing an API that somebody can connect to, there is this fear that uh these days with the eye, people can you know offload your data or start using your data or even uh build a third-party service on top of your your business and then sell it to other people. Um, so that's the case, but for our clients most of the cases that that was just old software that had uh limited capacity to be integrated with. So, we built for them automations using browser automations.
Basically, agent does all the clicking, everything that human does when when Before human was going to the same site, he or she was clicking this button, then typing this input, then doing something else. Then this can be repeated with the agents, especially now when we have computer use agents that can recognize the screen, make a screenshot, and figure out the sequence of of the flow, which is even better.
All right, this is very interesting. So, let's back back pedal a little bit. So, fundamentally, the internet, I think, is built on databases. So, there's a database and databases [clears throat] I mean I'm just trying to think of a nice simple example that everyone can can latch onto.
Um I mean, your bank is probably the one everyone's familiar with. There's a database there which which basically handles your money. It tells you money coming in, money going out, and you can go in through the bank, and you can get that information. But there's other um applications, other software tools out there which can also access that information. Now, they can't put money in and out, but they can see where there's been a transaction, and I can pull that, and they can use that. I'm thinking of a couple of tools that's in America here. QuickBooks is one of them. Yeah, overseas in Australia it's like Xero.
Um but they basically go to the bank, and they interrogate the database through this thing called an API, which is is an applications programming interface. Did I get that right?
Correct. Correct. Yeah.
So, that's just like a little set of rules that if you want to get the the bank's the bank balance, you send it a certain string of code, and it it goes, and it comes back, and so that information comes out of that database into this database. Now, two things that you said. One is is some companies now are shutting them down a little bit because they want to try and protect their data. That's interesting.
But the other one was it's often the case that you're working as a human, you're working across a multiple different uh different um applications. And let's take it QuickBooks and you're looking at some other CRMs two examples there, right? You need them both open and you want to do something manually to make sure things are right and what you have to do. You now can do with what did you call it? Content? Computer
Uh computer use agents.
Yeah.
Computer use agents. So they work by actually sort of seeing the screen like a human does and clicking over here and filling a field in there. Yeah, that's
That's one of the ways.
Um I wanted to actually jump on top of these two things you've mentioned. So the first, companies that are hiding their APIs and closing them down. Uh I truly believe these companies would die pretty soon or die out pretty soon simply because um even these days when when you or I were searching for a new applications to be used, one of the criterias we're actually searching for is ability to connect with the agent. So if if I cannot do that, that that application doesn't even surface my mind. And second point to your second part uh of the quote. So um we do use computer use uh agents. That is usually not the most stable way of automating um things like let's say finance.
And it's great that you actually mentioned finance because usually that's one of the most uh prominent and easy first case of automations with AI um because it yields something immediately. In particular, if let's say we have a business that runs on top of QuickBooks and they have their own [clears throat] CRM, with agents we can connect, let's say, profitability of each client or of each department of the or whatever they're selling and tell them to double down on this or or maybe order more of this so so the revenue becomes more aligned with their expectations, stuff like that. This is where actually agents shine a lot when they can connect multiple applications through some sort of proxy.
So, it might be computer use when agent goes to one side and then to the other, but ideally how we do it, agents go to goes to first side to the second side to the third side, scrapes it, figure it out how to connect it like through through selectors, through built-out browser automation, and then does it in more reliable sort of programmable way,
Mhm.
which substitutes an application if that that can be said.
That's that's interesting. So, this uh uh There's the There is a tool, I forget the name of it, which can kind of uh read
UiPath maybe?
No no well, I don't think it's even that, but um I don't remember.
But, it was notorious for I guess the problem that can come about if you if you're looking at the screen um and you expect the field that you want to enter into to be, you know, top left-hand corner and then they do a rev and they move it to the bottom right-hand corner. All of a sudden the the agent can't work because it's expecting its position to be there and it's over there. Um um Is that problem solvable with um with these new computer um usages?
Exactly.
So, this is why this whole scraping movement, I would say, has now like an another revival because now you can build these self-healing loops where, let's say if if your pipeline is failing and it's failing because you cannot find this that same label anymore, you can ask an agent to go see what changed, come back to you with a proposed solution and maybe even apply it. And you can do that all while, you know, sitting with a beer across the like near the ocean and then doing that from from your phone. So, that's that's beautiful. Exactly. That's why I'm here.
Yeah.
So, here that's that So, we've just gone into a really nice area which I always like to explore and that is the sort of interface between humans and AI. So, we now just put the humans back in the loop.
[snorts]
So, we're saying, "Okay, we can do all this great computer use agents, but then okay, something changes, it makes a proposal, and then it sends it to the human to say, "Yeah, I like that proposal. " Or no, that's that's you're completely screwing that up. Don't do that. That's a really interesting area because I'm of the belief that it's AI and humans. I don't see humans stepping out of the loop. It's just a question of where that line is, right?
We're going to be doing so much more with AI. The human piece is going to be, I guess, smaller in some ways, but a very very important part. And this is one of those areas which is critical. If something changes, AI doesn't really know what to do. You got to bring the human expert back in and and then it works really well again.
I think human expertise these days is required in describing rules. So, if I'm going to align that with projects we did, when we were working with 3D coin project, it obviously required reviewing the coin multiple times because there are not that many ways to say that a coin looks realistic, to be honest, in objective quantified manner, I mean.
You can as a human being, you can take a look at it and say, "Yeah, that kind of look looks like a coin. " But, if you want to express that in numbers, it's pretty hard to do so objectively. And so, we we had to include a lot of human labor in that. But, um if we're talking about, let's say, the voice bots, you can express what the noise is, or let's say, how well the model guesses the words when I'm saying. If you know the words I'm about to say, it you can calculate how many times it mistake it made a mistake uh with the words that it predicted.
So, these days, how how we see the whole loop, or the whole cycle between the AI and human, is a human does a lot initially initially by describing the rules, defining the test cases, defining defining use cases, defining how AI learns that it did something wrong. And I think that's the most magical part so far that I've seen, that if you can properly describe to AI model what the failure is, it can then really quickly try to admit that failure, and not only fake admitting it, but but actually do the the proper fix.
Mhm. Amazing. Amazing. Well described. And Maksim, so really interesting.
So,
[sighs and gasps]
So, we talked about the coin, we talked about voice bots, we talked about non-APIs. We're sort of running out of time now. Is there anything else that you'd like to ask me, questions for me, or comments you want to make?
Um I was yesterday thinking about where it's going to go in like 5 years or so. What's your opinion on this right now? What do you think goes?
Well, I think uh in the next 5 years we're going to see pretty Look, I've seen some technology changes. I was around before the internet was adopted, right? And I remember a guy coming into my office and saying, "You need to get this email thing. " Cuz I was at a uni and he was at a uni.
And and then I didn't really think it was very interesting because he was just sending me an email to say he was coming up to visit me. And then of course I got the wow moments with with the web. That was a big one. And and [clears throat] literally these wow moments have come faster and faster, especially with AI. It's just accelerating. I've never seen so much change so quickly as I have with that AI. Um so I think over the next 5 years businesses are going to be transformed. Our lives are going to be transformed. I don't know anybody and honestly I don't know anybody that's not using an LLM.
Uh and that's only been two in the last 2 years where I've gone from from from that to there to there. The low-hanging fruit I think is in as you said it's in the finance areas. It's in those areas that I guess are are rule-based because we can describe the rules. [clears throat] Therefore we can invoke the AI. Um and then there's areas that I think I wouldn't say completely untouchable but they are when there's human to human and I'm thinking of people working physical therapy is a really good example. You've actually got to touch somebody's arm. You've actually got to massage them. That's a long way out and it's for AI to touch.
Um but I think 5 years is is is probably long enough to say that the majority of businesses will have to adopt AI. And just to to round that out, I asked an investor late last year, "Would you invest in a company that was not doing AI? " And he said, "No, I would not invest. " So that tells you that people who are putting money up are saying that AI is certainly the trend. But Nikai, what do you think? Where do you think it's going in the next 5 years?
Um I think we'll have a lot more use cases where agents are driving the conversations. Right now, it feels like we are the one triggering it. We're the one asking the questions.
I think in 5 years we'll have more agents predicting our next decisions. You know, whatever we do based on our lives uh based on the whole context they have about us, they would offer us like better ideas, better solutions, which might sound scary, but it also sounds like a progress because in some sense there would be a something that that cares about you to some degree and and makes decisions on your behalf, which are scientifically good for you,
Mhm.
which I think is a good thing. Um, but yeah, I generally agree with you that the AI is yet another tool, not something like a god or something like that would uh take us all into prison and just make a doom here on on this planet.
It's just another tool to make us better and help to make us even more creative and more productive and efficient as human beings.
Right. Nice. Well, let with that, let's I get the final question out of the way, and that is AI, is it hype or help?
It is really a help with one small caveat. For some reason, the most popular AI use cases are the one that are hype and are completely wrong. Like AI avatars, AI SDRs that would bring you dozens and thousands of leads over the internet, like voice bots that would call on your behalf and make you rich tomorrow. You know, these memes of like "Call me a $10 million ARR application, make make no mistake.
" So, these are the ones These are clearly hype.
Mhm.
The help is in small and very repeatable things. Even even if you're not a business, if you let's say you're just a regular human being going to to job uh and you have you're you're spending money, you're buying groceries every day or every other day. You can use the AI agents to optimize that to off to find you like better deals, to figure out how to spend less money this month and it's going to make you better person. It's going to be you you'll have your own return. So, my point is AI is definitely a help, but you you need to be clever about picking up the place where you're using it.
Well said, Mykhailo.
Well, thank you very much for coming on the show. We'll have to have you back again soon because you got lots more projects that we want to talk about, but for now uh thanks for coming on and I'll talk to you later.
Thank you very much. Thank you for having me. I appreciate that.
Well, that was really great. I enjoyed going in deep with Mykhailo on some projects he's working on. Three things I want to talk about. One, voice bots. Seems like an obvious thing for AI to do, make calls on your behalf, but it's over telephone system. Telephone systems have limited bandwidth. We don't notice it, but it definitely is and the AI bots really struggle.
And when they do struggle, they hallucinate. They make up, they make guesses. And so that is a problem for AI and voice bots at the moment. The second area we talked about was computer to use agents. This is rather than the traditional way of one database talking to another, which is through an API. This actually acts more like a human. Actually, the interface on the computer. So, going directly between portals. And that brought us naturally to the third, which was a conclusion that Mykhailo made about companies that are not allowing access to their databases because they fear that AI is going to somehow erode um their business value.
He felt the companies that did that will be going out of business. I'll leave that conclusion for you to mull over. If you want to know more about Desti Labs, please click the link in the description. If you want to know more about Skillion, please do the same. I hope that helps and thanks for watching.
