Summary

In this conversation, Antti Karjalainen discusses the evolution of RoboCorp to Sema4 AI and the shift towards agentic automation. He highlights the early adoption of generative AI and the development of AI agents. Antti explains that the direct personas for purchasing AI agents are often line of business champions, while AI innovation and governance are typically centralized within organizations. He also shares his personal evolution as an entrepreneur and the challenges and opportunities of working with AI. The conversation explores the potential of AI agents to automate complex processes, the importance of quantifying ROI, and the future of user experience paradigms in AI applications.

Takeaways

• RoboCorp evolved into Sema4 AI, focusing on agentic automation and AI agents.

• AI agents are being adopted by line of business champions and centralized AI governance bodies.

• AI agents have the potential to automate complex processes and provide significant efficiency gains.

• Quantifying the ROI of AI agents can be challenging but is crucial for adoption.

• The future of AI applications lies in innovative user experience paradigms.

Alp Uguray (00:01.479)

Welcome to Masters of Automation. I have the pleasure of hosting Antti again today. Antti, welcome

Antti Karjalainen (00:09.634)

Thanks for having me out.

Alp Uguray (00:12.148)

Since the last time we spoke, lots of changes happened. Obviously, agentic AIs kick off large language models out there. But also, the biggest change was RoboCorp is SEMA 4, which is a huge change taking the story from RPA to more of an agentic AI that can orchestrate automation flows.

Just to kick things off, what has been new since then? How is the Sema4 got incepted?

Antti Karjalainen (00:52.366)

Well, it's a huge and loaded question there right from the beginning. I started RoboCorp in 2019 to bring this Coded Codenative Automation to RPA. And I think we built an awesome solution for that problem space. And in fact, the beginning vision was

RPA is going to go into more complex use cases and over time, the limitations of that drag and drop style automation are going to come evident and you'll sort of become this expert domain of coded automation. Now we saw that happen with intelligent automation incorporating various other tools like document processing into RPA.

people building centers of excellence over years into large companies and a lot of specialist companies emerged to provide services for this. And some of the leading RPA vendors launched actually coded automation tooling and platforms just not too far ago. So I think we kind of called it from the beginning at RoboCore what's gonna happen in this field.

But at the same time, I think that we had a different evolution happening with LLMs starting from years ago with the first Lama 2 models that weren't too useful. then I think it was 2020, we got GPT -3, which kind of blew everyone's mind with the first demos. then 22...

the chat GPT moment and then it became evident that everything is going to go that way. So we started already while building RoboCore, we started investing into generative AI quite early on a few years ago. I started exploring, dabbling into it and kind of trying to understand how it will change automation.

Antti Karjalainen (03:04.11)

Really last year in 2023, we started heavily investing into that direction, kind of taking the company more towards the direction of agentic automation. You don't necessarily know always what's going to happen with technology shifts like that and how the architecture is going to play out, but directionally we knew where we were going.

We had built some interesting technology for AI agents, AI assistants, and we started talking with some people that I know from my past coming out of some of Cloudera leadership team, as CEO and the top product and engineering leaders from that company. they were exploring to build something with AI and in the data space.

building towards AI and automation, and it just clicked very strongly. And so we decided to combine forces and launch Sema4. And so Robocore was acquired into Sema4 AI to launch the joint venture, well, not the joint venture, but the real, company properly out of stealth. then we really focused on AI agents. And that happened.

late last year, early this year, came out publicly with that announcement. So it was a long lead up into kind of directionally going into AI and now really locking on to agents as the focus with Sema4. And you could think about personally, it's sort of a hard pivot in a sense.

But a lot of lessons learned in automation and of course we continue serving our existing global customers going forward. So, now getting to focus fully into agentic automation, AI agents is really, know, exciting time to be here.

Alp Uguray (05:04.77)

I'm curious because now like RoboCorp served very well to the automation COEs and centers of excellence at the customers. So when AI agents come and then there's been a lot of talk around it as well, who is the direct persona from businesses that are looking to buy it? Like is that again automation COEs adopting and tying

with RPA workflows or is it more like tying it to more of a CFO's office and then going by department and department? How do you see that evolve over time from RPA to more of an agentic AI that still needs maintenance, orchestration and things of that sort?

Antti Karjalainen (05:55.714)

Yeah, good question. I think right now what we're seeing is a lot of the work that we do is championed by line of business. So that will be where the use cases are really. I think the way we are building agents really brings it close and approachable for the business. it's not a distant thing like you throw over the fence a set of documentation and requirements and maybe two months later you get

get something working out of the box, but it is more of a tight collaboration, defining agents, the complete work for the line of business. so there's definitely that aspect to it. And then it seems that every company has organized themselves around AI in some way or shape where they have a central governance body

for AI and AI innovation specifically, and obviously agents fall under that. we don't know yet how companies will shape up around this. Some COEs are leading the charge into AI. In some companies, AI is getting spun off to a different entity on its own inside the organization. Typically, we see though that AI innovation and governance is centralized.

which is a good thing. We don't want to have a federated model where everyone is figuring it out on their own. And so I think people have really understood the risks if you do this the wrong way. So that's why the centralized nature and then it's, it's the line of businesses are ultimately that pull the use case through and, and realize the value of an agent.

Alp Uguray (07:42.754)

It is interesting, think definitely more centralized strategy around the governance and then how to manage even the workflows is going to be huge. What made true how you evolved as a person from RoboCorp times to Sema4 times right now? Because it's been a long journey. mean, RPA

like at the time was just starting, it grew, it's a huge industry. And now you're again at the beginning of the AI agents, right? Large language models is just bursting as well in industry keep growing. But as a person, as an entrepreneur having two startups that was right at the pinnacle, the moment of their inception.

What were some of your learnings? What did you see that others did not that stood out to

Antti Karjalainen (08:43.554)

Well, yeah, so starting with RPA, we kind of entered that game later. So we observed the market evolve and then, huh, we could do this different way. then there, we typically actually were replacing existing solutions in most places where we went. So we went against incumbents and replaced them. And that does an interesting sales motion in a sense. And so

With agents and AI more broadly, it's a green field where you don't have incumbents in there. You might have existing solutions that are sort of doing similar things, but you're replacing them with 10 times, 100 times better solution in the end. So it's not the same situation. Here, everything is so new that everyone is figuring it out at the same time in the industry.

And so when I say in the industry, I talk about people building LLMs, people building engineering frameworks, coding frameworks, development frameworks, people figuring out the architecture, the UX patterns is sort of the broad community of builders and engineers and scientists and everyone contributing to this is kind of figuring it out and it's breathtaking speed. So I think here, what is new to me is just

space of new breakthroughs and innovation. It's a full -time job nearly to just follow that and stay on top of it, which you have to do because you're applying a lot of this stuff in real time into an actual product. And the amount of research papers that I've been reading, I don't know when must have been my uni days when I used to read research papers. So in that amount, but it is breathtaking.

to have a research paper that's two days old and say, huh, maybe we could actually apply this in an interesting way. so that's one aspect of it, the technology. So leading the charge in something completely new. And then on the business building, the company building side, now there's a new set of co -founders.

Antti Karjalainen (11:02.566)

building the company with. that's always a new experience. Some very experienced people coming around the table with decades of experience leading and taking companies public multiple times over and leading them and scaling to hundreds of millions or even billions in revenue. so that's a new thing to operate in that kind of group. And then everyone's experience in leading companies is kind of showing in

we engage with customers, how professional the motion is from back in the days of having to learn it from scratch, from now to being done it multiple times, like how professional the customer go to market motion is and how early we engage with customers and how carefully we listen to customer signals when we pitch and then adjust the pitch and adjust the product narrative and adjust it. so, so test it super early.

and then really surgical execution of the product build out itself following

Alp Uguray (12:09.09)

And what do you think from the customer side is like they see as hype versus reality. Like, for example, they may come to you and then be like, I want to innovate this process, reinvent it. And they may be thinking some of AI's capabilities very maybe three to four years in advance. How are they thinking based on the conversations

that you have today.

Antti Karjalainen (12:39.79)

Yeah. So I think everyone has some sort of way of thinking about AI or agents or what they should or could do. I think the common kind of fallacy or the misread reading of AI is people say, yeah, it's AI, can learn. Like, no, it doesn't learn unless you very specifically train

As people associate AI with this self -learning, self -improving thing, which it can do, but it's hectic work to make it do that accurately and reliably over time. that's one, I think that's sort of a minority of people right now, but we tend to lead with showing instead of just telling. So we do practical demos really early on to level set people into

getting their hands on, eyes on, of actual working AI agents and then using that to drive the conversation. That's typically helpful. And I think in the industry more broadly, there has been a bit of agent washing going on. So everything seems to be an AI agent nowadays. And so I actually wrote, I'm finalizing this new article blog post about the levels of agentic automation and capabilities.

So I'm trying to build out the framework to have for people to thinking about agents so that similar to self -driving cars, everyone knows that there's from zero to level five self -driving zero is your old Honda with stick shift and everything and hands on the wheel and eyes on the road. Whereas level five is you're reading a book behind a self -driving robot taxi wheel.

So we're trying to basically frame it out the same way as that with agents so that you have different levels of capabilities. A co -pilot does this versus an agent that can reason, plan, and act, and reflect, can do this. And then there's things that we can't do yet but are within our reach if we use a bit of imagination.

Alp Uguray (14:54.774)

That's interesting. if we were to apply the same principle to like agentic AI and think about that, right? Like self -driving car example that you gave, like the ultimate end is that someone just chilling at the backseat reading a book or like listening music. What would it look like when large language models or agentic AI scales to that kind of

comfort level so that we can adopt it maybe for both consumers and for businesses.

Antti Karjalainen (15:31.318)

Yeah, so there's been a debate in the industry about, you know, can LLMs reason, like properly reason or not, and it has been proving in lot of experiments that they can't do logic, like simple logical relationships. They just can't grasp, like, auntie is the guest of Alp today, and then, you know, who's hosting auntie today, and it can't figure it out. And so,

those kind of relationships, it can't map in this world model, depending on the complexity. So I think people who say that AGI, like true, like general intelligence is here based on LLMs, they're kind of extrapolating too far away in the future. But what I think of level five agentic automation would be where it can actually solve problems and

work that it has not seen before. And so that sort of behavior requires a bit of creativity, like original thinking, and then being able to apply logic consistently. And if you have even a glimpse of that behavior in an AI model sometime in the future, we can then apply engineering into it and multiply that reasoning a thousand times over and get to systems that are

pretty smart pretty quickly. so then what it looks like is that you could basically say that you could have like an accounting agent and it sort of knows the rules of accounting and can search and can have converse with you to have your opinion because it's sometimes opinionated and a strange invoice comes in and it goes, let me take a look at this. And that kind of sort of looks like this.

and apply original thinking into it without having a very competent coworker. And then very soon you would actually go into like, super competent. It's way more competent than any one of you guys. And then you have a thousand of these in a box and you can just scale them out endlessly. So what does that mean for human work? I don't know. We don't know if or when we're gonna get there.

Alp Uguray (17:43.01)

Good.

Alp Uguray (17:50.844)

Thank you.

Antti Karjalainen (17:56.366)

with that level of behavior. Certainly, I think not scaling today's LLMs. We need another architecture

Alp Uguray (18:04.62)

And it sounds like then that architectures, like one of the biggest component is giving a direction on reasoning capability. So that it will reason as an accountant or it will reason as a, I don't know, software engineer so that the outcomes or for outputs that it provides are very specialized. So if we think about that way, then there will be maybe hundreds of specialized

AI models and then there'll be one model that tap into their intelligence or like trigger it at times. So it could be an interesting time because then if you can count on their reasoning and they reason really well, then there will be maybe this will democratize the entrepreneurship in a way as well because there will maybe companies won't be just 10 ,000 employees, 100 ,000 employees, but maybe there'll

500 startups that come out of it that have three employees each or four employees each. Do you see it's like a same way as well, like how it can evolve?

Antti Karjalainen (19:15.96)

Yeah, people are predicting like when is the first time we see a three person company hit billion in revenue or whatever it is like. I don't know. I think when we think about reaching that level of AGI, it's still out in the future. We don't have a line of sight towards that. So I tend to not speculate too much about that myself.

Alp Uguray (19:22.588)

Yeah.

Antti Karjalainen (19:44.398)

because I'm really focused on solving today's business problems with today's AI and obviously knowing what's around the corner so we can expect and build towards that. But AGI, to me, it's not around the corner yet. But yeah, I mean, that kind of world where you have endless access to skilled workforce everywhere, for everyone. Yeah, it'll democratize a lot of things, for sure.

And I think a lot of this conversation around AI is revolving also around AI taking jobs and stuff like that. think the same narrative applies as to automation in previous days. It's like you get to more interesting stuff and you'll get to work on things that matter and are meaningful. And we are at a point in history where people are working longer days than ever before. And we have all this technology around that.

did that come about? I would want to work eight hours a day and be confident that that's enough instead of not. So I think there's still ways to go with AI.

Alp Uguray (20:55.85)

In terms of like the way you like you're like as a co -founder and someone an entrepreneur within the field, like have you adopted into your workflows a Sema4 or like a similar agentic AI to do certain things for

Antti Karjalainen (21:14.168)

Yeah, for sure. It's like literally today I used some AI to help me write that article I was mentioning here about the levels of agentical automation. I think it comes in subtle ways. It's like helping me understand information faster and helping me be more creative in writing. And I use it here and there. I wouldn't publish anything completely written by AI, but help me get another

way to phrase this thing here. It just unlocks you really nicely. And then internally, we are using a bunch of agents in our company, in this fairly young company, we are already using agents. Typically, they are in the software development space. So we have, for instance, an agent to go through a new Python package that we want to incorporate and go through the due diligence of, you know, is it legitimate package? You know, is the license valid? Should we use it?

you know, any dependency issues that it might bring in and all that sort of thing that, and then write up a report about that. So these kinds of things people are just putting to use cause it's so, so good to have it around and it's easy to build and run. So, so I think, you know, companies of every size will, will employ a bunch of agents as, this technology gets more broadly available for people.

Alp Uguray (22:40.286)

In terms of this unlocked totem in my mind, especially the BPO times, where they get all the business processes are outsourced and an RPA came in and then became the stitching software on top of the business processes to automate them. And now AI agents came and then it makes it much easier

centralized data in one place and had reasoning on it, the cognition capabilities that in the past was not even been dreamed of. So for as a technology capability, and I will unpack this question here, but at the time when you have the RoboCorp, there were some problems that the customers have like invoice processing is one of them, like we talked about earlier.

And then there's like other financial loan predictions, like loan issues and stuff like that. What do you think was not possible at the time with RPA that today now you look at it with AI agents and you can say, yeah, we can now automate this or replace or augment it end to end. And what are some of those use cases that come to your mind today that you are like,

damn, now it works. Now I can integrate it.

Antti Karjalainen (24:04.808)

Yeah, there's certainly new types of use cases that we keep uncovering by the day and the week. And then there's the fact that how long does it take to build something? Whereas you could actually build an agent in a matter of a day that would have taken you a week to build a similar one in RPA. And so I think there is some differences in the build experience itself.

how do you put together the agent? But when you think about, like you mentioned, invoice processing, I've seen use cases that involve very complex invoices that are hundreds of pages long. And the rules to reconcile those invoices against an internal ERP system are page after page of written up text.

So with RPA, you would have had to try to codify all of those rules somehow so that you can build a logic that goes in and somehow recognize it. when you actually kind of on the surface level, yeah, seems doable, a lot of work. When you go into the details, you see that, okay, there's a lot of rules that humans apply intuitively into this reconciliation task, whereas an invoice could

mention the customer's name in slightly different way than it's read up in the ERP system. And you can just codify all of that. Whereas an LLM is really good at just reasoning about that and say, yeah, looks close enough. It'll pass. And so being able to ingest hundreds of pages of invoice into the agent and then have the agent reason about it and match the data against your backend system, your ERP.

Alp Uguray (25:41.025)

you

Antti Karjalainen (25:59.374)

and then come up with a report saying, here's what I found, these lines don't match, here's what's the difference, you have $10 ,000 discrepancy in this invoice. And I think it's because of these lines in a matter of minute, which for a human would have taken probably at least a day to crawl through all the hundreds of pages of data, find the discrepancy there. So I think just these examples alone,

Like no way we've never approached that use case with RPA to begin with, let alone be confident with Solvitt, but these are actually working with AI agents really well.

Alp Uguray (26:44.66)

With RP, it was always that we will have to have a more specialized ML model that's trained on that invoice template type. then once that we receive a different type of an invoice, the model crash automatically. So it does then allow actually capturing those edge cases that we couldn't before and manually we had to configure.

Antti Karjalainen (27:10.37)

Yeah, and I think what's kind of people haven't really, maybe this hasn't been in the talks so much, at least where I'm following people's conversations around AI, the multimodal models that we've seen come through in the past few months. I think those will provide some really interesting breakthroughs in the space of document processing and document intelligence.

seen so many RPA use cases that just struggle with complex document types. And it's funny, I kind of laugh at this at times, but we come up with the most groundbreaking novel technology breakthroughs and then we apply it to invoice processing. Always it seems to be that, but it's a well understood space and it's a good starting point for a lot of things to just prove it out. I think beyond invoice,

processing and stuff like that. Yeah, I think there's so much opportunities beyond that. And we've seen when we go to our customers and run through a first agent implementation, then they start coming up with, yeah, I have this process here and this workflow here. And all of a sudden you have a dozen use cases to go through and evaluate.

Alp Uguray (28:12.05)

That's hilarious. That is whole truth.

Antti Karjalainen (28:38.744)

So people need to see one thing working really well in a unique new way for them to unlock the thinking and then, you know, the use cases are endless. here, obviously, invoice processing, it's still an interesting thing to solve. But now we're solving hundreds of pages of invoices, not just one page of like, you know, reading the vendor name or

Alp Uguray (28:51.218)

Hahaha

Alp Uguray (28:59.276)

Yeah

Alp Uguray (29:04.271)

We'll get there, we'll get there one day, there are annoying noises.

Antti Karjalainen (29:08.184)

By the way, funny thing, and this is why it's actually funny, is that I live in Finland. I've lived in Finland for my whole life. And here we actually have electronic invoices. So nobody sends anyone PDFs or paper invoices. It's all electronically exchanged with fully structured data. And that's throughout everything. Nobody sends any PDFs.

Alp Uguray (29:31.522)

Structured data.

Antti Karjalainen (29:38.702)

Yeah, the world isn't ready on that side either, but luckily we have now pretty strong technology that we can finally take people from reading scrutinizing 100 pages of invoice on a PDF to working on more meaningful

Alp Uguray (29:54.43)

I think that's interesting because the solution exists, Maybe it's more on the government and regulation side to apply a standardized way to manage invoices, but people don't do

Antti Karjalainen (30:04.642)

Yeah, but then again, we are dealing with enterprises and the realities of enterprises are that they are multinational companies operating around the globe. So you might get a shipment of, know, nut samples from Indonesia somewhere and then try to get them to send over the right electronic invoice in the right format might not be tomorrow's reality.

Alp Uguray (30:32.65)

Yeah, I mean, some of them are just handwritten as well that like somebody just sketches something and then

Antti Karjalainen (30:39.126)

Yeah, exactly. It's global trade, so you can expect some of that.

Alp Uguray (30:45.838)

There's an interesting thing you mentioned because like one of my mentors, I was speaking to a few months ago and I was designing this product interface with Figma and stuff. And I was thinking about talking with him about how to drive the most value from an LLM. And then he made an interesting point that like if you had the all worlds

intelligence, what would be the question to ask about this specific use case? And I was trying to just make the user experience easier, right? Like apply those Airbnb principles of three clicks away. So the coding is right, like user journey is right. So it made me ting. And then I went and asked both cloud and chat GPT, here is

wireframe that I have and here is the user journey I imagine and I want to design an interface that is similar to Apple and how Airbnb would do it and can you come up with a few ideas? And ChatGBD gave some ideas more verbal and then cloud was able to fully sketch an image

Antti Karjalainen (32:08.163)

Nice.

Alp Uguray (32:10.514)

I was like, yeah, I'm going to copy paste this into my workflow. I was thinking at that time, like to your point, right? Like it's like we were asking you to do invoice processing, but in hidden, in hindsight, there are like so many things that it's like tapping into an intelligence and how do we find the right questions to ask to it is also is an

Antti Karjalainen (32:13.432)

No.

Alp Uguray (32:40.098)

like an education, right? Like that people need to learn.

Antti Karjalainen (32:42.74)

Yeah, exactly. And now what I see from the first use cases that are being uncovered with AI agents is that people easily gravitate towards like the sort of RPA style type of workflows that have been in the Harper for a while, but they haven't been able to crack them.

I think that's sort of the broad category is how I think about it is this save, so drive efficiency, there's protect, which is around compliance and compliance will be a huge area for agents. There's so many different use cases in that space that are difficult to implement first of all with any other technology, but second, they have the potential to protect massive losses

you know, realities that might come for not being compliant, banks, financial institutions, and pretty much any company. And then there's expand, which is like creating new lines of revenue, new business opportunities. And that's sort of the last one is something that we'll get to over time. When people have agentic systems to their disposal, they are starting to apply them creatively. It's like, how do you actually create new revenue opportunities?

I actually have one company I've been working with who has a new product in mind that they're going to build with an agent. I'm not going to spoil it for everyone here to keep their secret, but it's sort of going from that save, to grow over time. And we are just at the first phase, first inning where we're seeing people just apply it to like, okay, I got to save so and so many hours of this team's working time

this

Alp Uguray (34:37.058)

So in a way it takes off a more of a product centric approach versus process centric approach because then in a way, right, we, when we think about changing it with the new ways of thinking, then we may just dump the process to trash and build something new with a product. So do you also see some of the customers and again, without spoiling.

how they are approaching it. Like, are they seeing it more like, yeah, I want to take this and apply it to this process or like to your point, let me take that process, break it to fundamentals and build a new way of an interaction mechanism with an agent, chats, bots or such and create a product offering so that that way

If you have a product offering, you may have many products for many different processes. So it beats up also SaaS business as well, right? There are so many products out there for any type of processes. But how do you see that?

Antti Karjalainen (35:53.004)

Yeah, I I think there's both going on. So just apply it to the same stuff we've done all day, every day in the past and try to make it quicker, better, faster, cheaper, and then really breaking down your work into fundamentals and having the AI do that. think looking at just AI companies more broadly, I think the more successful folks in the field are figuring out how

not sell you software, but the work product of that software. So it's basically like a service as software, somebody used that. And so instead of like, hey, here's tech, what do you want to do with it? Here's a work product that happened to be completed by an AI somewhere. And so I think that model of being able to focus and hone in on the outcomes, the work products,

that the AI agents will complete for you is the winning recipe in this field. And so for a enterprise AI agent company, we of course primarily focus on what outcome for the CFO's office. Was it like a fully reconciled invoice would be like one measure of outcome? Like how do you get to that result? There might be multiple paths.

reimagine some of the workflows in between because of this new technology capability. But that's really the focus in the end always, the work product that you complete.

Alp Uguray (37:30.946)

Just like all capturing a good ROI to business case as well so that it validates investment and then going through it.

Antti Karjalainen (37:39.352)

Yeah, and we as technologists sometimes forget that people don't want to have agents for agents sake, right? So you and me, we might want to have them just because they're cool. And it's nice to play with new tech, but people actually have a job to do. so, you you get that agent to complete that job for them, not just partially make them more efficient. It's like agents are actually software that do the work for

Alp Uguray (37:48.479)

yeah.

Antti Karjalainen (38:09.12)

not just make you more slightly incrementally more efficient on that

Alp Uguray (38:13.89)

Yeah, that is true. That is true. I mean, the ROI side of it is also been discussed right now as well, like how to capture the value that an AI agent brings because sometimes it's very hidden into the way people work, Like just the example that I provided on figuring out user interface designs, that probably saved me a month of work.

but nobody really like goes in and then be able to capture

Antti Karjalainen (38:46.53)

Yeah, I mean, my example of making my final conclusion chapter slightly more punchy. It's like, how many hours did I save? I don't know. Maybe my article would have sucked a bit more if I wouldn't have the AI to help me out rephrase it. But I think what we're seeing here in the first phase, like last year, was a lot of experimentation. Everyone was just excited and pressured, to be honest.

like what are you doing with AI? we need to do stuff with AI. I'll just try it out. And so people were in larger companies, they would either enable people to have access to chat GPT or then various copilers. And that would be something that they would, okay, we're doing AI now. And a lot of folks realized that it's super hard to quantify

the savings, the efficiency gains, that the benefits ROI of having a co -pilot with you, because it's sort of not completing a full work product in the end. It's still like making you slightly more efficient somewhere or completing something for you. Like in software development world, we have seen actual studies and it's typically interview studies of developers asking like how much more efficient do you think you are?

co -pilot or various AI assistants and people say up to 40%. I can believe that. I think depends on the quality of the engineer. Like I'm super like I'm four times more efficient with AI than let's say some of our senior engineers here who know everything out of the head. So it can see mileage way very, but what I'm trying to say is

Copilots are super hard to quantify versus an agent going back to that work product is a fully onboard employee, fully reconciled invoice, a written out report on your compliance status on trade export, like you name it. Those are distinct work products that we can actually put value on and they're easier to then quantify and measure. And so I think we're going

Antti Karjalainen (41:07.458)

get there this year where people are starting to come up with case studies on actually deploying agents to production and having some really awesome numbers to share.

Alp Uguray (41:18.422)

Yeah, and I'm looking forward to that as well. I know there are some customer success stories a lot in the customer support process workflows, like tying an agent into handling the tickets and then resolution types and then navigating there. It's just that conversation with a customer who just wants to update their bill or phone number in their account, but it takes...

Antti Karjalainen (41:41.112)

Mm -hmm.

Alp Uguray (41:47.778)

a lot of money and lot of human time to do it. So we have one last question. I know that we could go for hours, so maybe we'll do another round later. So now in the podcast I'm asking at the end of episode to every guest, what would be the one question they would want to ask to the next guest?

So if you were to think about one question and it could be anything, what would it be?

Antti Karjalainen (42:24.082)

gee, this might be the moment where you need to cut and edit the show for a while. Are you going to have AI guest or what kind of guest are you thinking about?

Alp Uguray (42:35.97)

It will be a human guess, it non -AI guess, just to help with the thinking. It's like maybe what are you curious about these days? And other than like research papers, what is new and integrating into AI agents, like more from a broad aspect of

Antti Karjalainen (42:57.674)

I think if we stick with the industry that we've been spending time here on, I would ask what kind of user experience paradigms they expect to see in new types of AI applications beyond chat. What would that look like? And be a bit creative on what they would like to see.

Alp Uguray (43:28.054)

I love that question. really do. That's been actually something that's been on my mind as well, because I've been just looking at how cloud and chat GPT has been evolving over time. And obviously Instagram's co -founders

but their announcement of artifacts, shareable products, whereas chat GPD is still a linear chat interface. Yeah, I'm very curious about that as

Antti Karjalainen (44:02.178)

Yeah, think we are now in a phase where we have sort of started getting these capabilities to our hands. Like I have an agent full studio on my desktop here, but you don't have it because it's still not launched. so we kind of are getting there where these capabilities are so broadly available and kind of end to end for companies to be able to build, manage and run.

AI agents. we're getting there. I think the next phase will be the really exciting innovation phase where we're going to see a lot of interesting use cases, case studies with numbers, showing the efficiencies coming out. And then we'll see an innovation also on how you want to interact with agents and AI more broadly. And that's what I'm kind of looking forward to as one of the things that I'm looking forward to in this AI world.

We started out with RPA and been talking about agents. There's so much happening here that I couldn't be more excited to be working in this space myself personally.

Alp Uguray (45:09.628)

Absolutely. Likewise as well. And then the amount of exposure to business impact and personal impact of these solutions is massive. Like seeing it every day, how actually people are thinking of it is huge.

Antti Karjalainen (45:26.676)

And I'm naturally skeptical engineer from my mindset. So I'm not going with all the hype that is out there, but it took me a while when I started first, it was maybe early last year I saw first agents scratching my head and looking at them and where could this lead and is there any practical applications? And now it's virtually new application every day.

that you're finding out. But it took a lot of really solid baseline technology development on the AI model side to actually unlock that. And then this stack has been building up from the software development frameworks, from the infrastructure up and up and up. And now we are finally getting to the application layer where we're starting to see really interesting things come alive. So that's super exciting.

Alp Uguray (46:23.51)

Yeah. And then for an ending note, I'll mention that I've been reading about this research paper the other day, and then they were mentioning how AI agents can evolve to become the sole intermediary between a user and the internet. So like how we go today to certain apps to find information or certain websites to find or execute.

some tasks and information. So like the theory that they have regarding user experience was particularly focused on this agent system will be the sole intermediary between us and knowledge at large. But of course, there are certain risks and things of that sort.

Antti Karjalainen (47:13.688)

Yeah, I think more broadly, let's say going back to the enterprise space, I would bet that, I think AI is going to disintermediate a lot of different applications and platforms. I won't name names, but we are not thrilled to work with certain types of enterprise apps day in, day out, and being able to have a digital coworker to just take care of that for

Alp Uguray (47:21.666)

Still dealing with invoices.

Antti Karjalainen (47:43.128)

people will love that experience.

Alp Uguray (47:45.58)

Yeah, I think so too. I so too. On that side, was a pleasure to have you again, Antti, for a second time. I really enjoyed our conversation and maybe sometimes we'll have another one too.

Antti Karjalainen (47:58.53)

Awesome, thank you so much Al, it's good being here.

Alp Uguray (48:01.154)

Yeah, thank you, Antti. I'll stop