What we talked about
Philip Samuelraj is the Founder and CEO of Techjays. He argues that the real opportunity is workflow reimagination, not simple automation: agentic AI can handle repetitive work while humans move upstream into review, approval, audit, and higher level problem solving.
Philip and Federico compare this shift to past disruptions like Excel replacing manual ledgers and ATMs changing banking roles. Philip says the biggest risk is people tying identity to old jobs, while the upside is redirecting human effort toward problems that were previously too expensive to solve.
Show notes
Philip Samuelraj’s father was a banker in India when bank unions were actively fighting against computers and ATMs in the 1980s, convinced the machines would destroy their livelihoods. That family history gives Philip a long lens on disruption, and he uses it to argue that every major technology wave redirects human potential rather than eliminating it. At Techjays he applies that conviction to agentic AI, signing contracts not on hours or headcount but on business outcomes, and only getting paid when they are delivered.
What we covered
- Philip draws a sharp distinction between “workflow automation” and “workflow reimagination.” All existing workflows were designed with a human in every step; enterprise software was built to visualize and report on what humans do. Reimagining a workflow with agentic AI means asking where humans actually need to be in the loop for review or audit, and then designing the system around that much smaller role, a fundamentally different exercise from adding AI to an existing process.
- The quality curve for agentic systems is exponential: getting from zero to 80 percent accuracy is relatively straightforward because foundation models are already capable. Getting from 80 to 90 percent requires serious engineering rigor and reinforcement learning infrastructure. From 90 to 95 is harder than the entire 0-to-90 journey. From 99 to 99.9 is harder still, which is why deploying a proof-of-concept that works four out of five times into production is a mistake most vendors make.
- Techjays structures its work as AI builders paired with AI validators. The validators are obsessively focused on data and outcomes, cataloging all possible inputs, defining what the correct output looks like for each, and running the reinforcement learning loop when the system fails. A human in the loop reviews those failures and feeds the learning cycle; without that loop, the system plateaus.
- The most common mistake Philip sees is companies slapping AI on top of existing inefficient workflows and then calling a 10-to-30 percent improvement a success. The second-most-common mistake is treating a proof of concept, where the model delivers something impressive in a single-shot demo, as equivalent to a production-grade system.
- On the “gold rush” of AI vendors: Philip is direct that Techjays is in what he estimates is the 0.1 percent of companies that claim to build AI and can actually do it. He has had to take over projects from other vendors where none of the prior work could be salvaged. His response to hype is to point to whether a vendor signs on outcomes, if they don’t, the incentive to deliver is not there.
- On AGI: Philip says he had a moment 11 months before the conversation when he tested a voice demo from sesame.com and realized within 90 seconds he had stopped treating it like a bot. He expects conversational agents to reach an inflection point, probably in a few years, where users prefer talking to them over a call center agent, and that moment will be the practical marker most people use for what “general” AI feels like, regardless of how academics define it.
About Philip
Philip Samuelraj is the Founder and CEO of Techjays, an AI engineering firm whose mission is to help organizations reimagine their operations with AI. He previously served as a program manager at Google focused on Gen AI for support intelligence, and earlier built his career in QA, testing, and delivery leadership at Cognizant, Bosch, NetApp, and Infosys.
- LinkedIn: https://www.linkedin.com/in/philipclementssamuelraj
- Website: https://techjays.com
Episode 127 of the PreVetted Podcast.
Full transcript
Federico Ramallo (00:00) Welcome back to the pre-vetted podcast where we spotlight extraordinary people and remarkable talent reshaping our world. Today we are joined by Philip Clements Samouraj, founder and CEO of TechChase, where his mission is simple and bold, helping the world build AI and build with AI. Philip has led teams across quality, program management, and product execution, including roles at
Google as a program manager for support intelligence, focus on Gen.ai for good support. Before that, he helped build and scale ambitious technology programs at Ervivo and Vivint. And earlier in his career, he grew through deep testing and QA leadership at companies like Cognizant, Bosch, NetApp, and Infosys. He’s also been recognized for excellence in testing and innovation.
and he brings a rare mix of engineering fundamentals, delivery leadership and builders mindset. Philip, welcome to the show.
Philip (01:07) Great to be here, Frederica. Looking forward to our chat.
Federico Ramallo (01:11) Awesome, I’m honored to have you here today.
Philip (01:14) The pleasure is all mine.
Federico Ramallo (01:16) So for people meeting you for the first time, how do you describe Tech Chase and what do you do?
Philip (01:23) Wonderful. So our mission at this point is how can we help reimagine the world for a better future with AI? What has reimagination been? So the world as it is, is structured around, the knowledge work is structured around people doing all of the work, processes that are built for that and technology that enables that.
With the AI revolution, what we see is that the technology is so far advanced and there are so many capabilities that knowledge work as such is disruptive. All the way starting from how we build software. And that’s why we call, we not only build AI, we build with AI. So now the time is ripe where the whole world can be re-imagined with AI.
all of knowledge work can be reimagined. So whether we are you’re a software engineering company or a manufacturing company with a large operational team, all of you can reimagine how your engineering does or your operation is done in an AI first, AI native way.
and completely revolutionize the way you drive outcomes, whether it be saving amazing amount of cost or driving phenomenal amounts of revenue.
Federico Ramallo (02:48) Amazing. Yeah, I that there’s a lot of hype around AI, but ⁓ I agree that the ⁓ knowledge workers as they are right now, the workflow is going to be disrupted. ⁓ I don’t believe this idea of ⁓ that AI is going to take a lot of jobs. I think that those jobs are going to be reinvented in new ways. People will have to develop new skills for sure, right?
Philip (03:02) Yep.
Absolutely, that’s the whole reimagination, right? I was talking earlier this day with a customer and he came up with an amazing example. So there were a lot of accounting jobs and there were people manually working on ledgers, right? On paper. Then computers and Excel came in and it’s not like, ⁓ people can’t really hold on to know I am great at what I do in this manual ledger.
Federico Ramallo (03:33) Right.
Philip (03:46) Your job morphs into how you can use Excel to do the work and accomplish what you’re trying to do. What these will open up is that the mundane work will be taken over and so many other avenues will open up. Knowledge work is just the beginning. Say knowledge work gets reimagined with AI.
Look at all of the creative output of people. So now the next frontier in knowledge work would be there are so many problems that we have not solved for because they have been too expensive to solve or the return on investment is not there. Those becomes the become the frontiers of research, the frontiers where people actually solve problems and move humanity forward. That is knowledge work. The second part is the physical world.
Federico Ramallo (04:38) Right.
Philip (04:42) Robotics is at its very early stage. So robotics plus AI.
there is huge amounts of disruption that is waiting to be had. So again, which will change the way people do physical work in certain avenues, and then again, open so many areas of potential where we’ve always thought these are problems that we cannot solve for to now these are problems that if you solve for they make sense and they give us a huge return on investment.
Federico Ramallo (05:16) Right, right. You can get, ⁓ you just need to think of human productivity in a whole different perspective. And there’s so many new opportunities now coming up.
that we are just starting to see the dawn of it. So we’re not really seeing the whole picture yet. Following your example on the manual lectures, if you were to tell to those people at that time,
you’re going to be able to use a computer to automate a lot of your workflow, right? It wouldn’t make sense in their minds, right? ⁓ It’s a paradigm shift that needs to take place, right?
Philip (06:05) Exactly.
See, my dad was a banker in India and I still remember these conversations when I was a kid, right? The whole idea at that point of time were like, computers are king, they’re coming, they’re going to take away all these jobs. And there were these unions in these banks who were all against computerization. Can you now think of that? Right?
Federico Ramallo (06:27) you
Philip (06:29) They were all against ATMs. They said like, ATMs are gonna take tellers jobs away. You should not modernize. Can you even think of that now, right? So that is this paradigm shift. So as every single paradigm shift comes across, I do believe strongly that humanity moves forward for the better, like we build a better future.
While there is some amount of disruption and change, I do believe it is just that we channel human potential differently rather than the same old ways that we are traditionally doing. That is this fundamental shift. And that’s why our mission is to help reimagine the world for a better future with AI. And that’s how we want to do it.
Federico Ramallo (07:07) Right.
I can give you another example that now that I was hearing you, I remember for some random reason. ⁓ So the elevators, they used to have a driver, human driver, know, moving the lever to go floor to floor, right?
Philip (07:30) See you.
Federico Ramallo (07:34) So they would ask you, which floor are you going? And you would say, whatever. it would do the dynamic routing in his head. And this was in a government building that I saw this elevator. Many years later, they changed the elevator to a more modern elevator with buttons. But for union reasons, the union pushed to preserve the job of that.
person. So you would get into the elevator and the same person would say, which floor do you want to go? And he will press a button to take you to that, you know, which I found so interesting, so funny because, you know, his job was redundant, right? ⁓ And kind of pointless, right? But he was collecting a paycheck. So, you know, he was happy, I guess, ⁓
Philip (08:24) Yeah, I do
believe at that point, right? You think about this, Friedrich. That is a huge waste of human potential. There are so many problems in this world that needs to be solved. So many problems that we can solve for. While instead, because we are many are stuck with this idea of the old world, we still are solving for problems in the past instead of solving for the future.
Federico Ramallo (08:33) Right.
Philip (08:53) The same like for example, ⁓ we all want, you know, clean air, clean water, know, great food, even fundamentally there are so many problems to be solved. Which these resources can very well be redirected to.
Federico Ramallo (09:11) Right. Right. And the reason I was mentioning this example is because there’s a group of people that are going to push back and are going to try to hold to the old world, right? ⁓ Where their job is going to become redundant, right? ⁓ To a certain point, right? ⁓ So. ⁓
I think that there’s going to be a group that is going to resist, there’s going to be a group that is going to embrace, and those are going to be the ones that are going to take the most advantage of this new reality. ⁓ And then there’s a group that is going to be in between, mostly because they don’t quite understand what needs to happen or what other things they need to do. So, you know, it’s a…
Philip (09:50) Yes.
Federico Ramallo (10:02) it’s going to be an interesting transition to see happening, right?
Philip (10:08) Yes, yes, Virgica, I do believe it is going to be an interesting transition. And what I see also is that, I think this is also proved in any disruption or any innovation wave is that people sometimes tie their identity to what they were doing. And when a disruption comes in, people are not able to move forward because they’ve solely tied their identity to what they were doing.
Federico Ramallo (10:35) Right.
Philip (10:36) while they are very
well capable of learning this new wave and going over and delivering it.
Federico Ramallo (10:42) Right, yes. So you’re thinking there is an existential crisis in this process. Right.
Philip (10:50) Yep.
Federico Ramallo (10:52) So what kind of teams and products are you most excited to build right now?
Philip (10:59) So anything in the frontier of problems that we thought that we could not solve for, those are the most exciting. For example, what we see with agentic engineering is that people, ⁓ think CloudFair today or yesterday, they came up with, hey, we’ve rebuilt Next.js in a week or so on, right? So if you look at it, there are like problems that are…
kind of deterministic at this point of time, which have a very good documentation, modus operandi and all of that. And those are very well solvable. One of the things that I am really interested in is solving, are those are the things on the frontier. For example, there is a business that has to estimate and bid and provide bids, a logistics company estimate bids and win them.
Can we truly build an intelligent bidding machine that actually makes sense of what the market is about, what customers we’re coding to, take all of the lessons learned by the best bidders in that particular organization, and genuinely solve for it so revenue goes up? Say for that another problem, like say there’s an operational team.
a large operational team doing a lot of mundane repetitive tasks, but you do that with AI, with agentic AI, it becomes incredibly fast so that companies have a lot more leverage. Yeah, they reduce costs and they are able to grow. They have operational leverage without having to add to headcount. Can we delineate that? These are the kinds of problems that I’m so interested in solving because they truly make an impact both on the top line and on the bottom.
⁓ and that is what businesses are looking for.
Federico Ramallo (12:52) Right, right. ⁓ That’s very interesting. And what about workflow automation? I mean, you kind of talk a little bit about that, right? But ⁓ how much potential is in workflow automation versus, know, Gen.ai, assistants, whatever you want to call it, right? To assist ⁓ humans on their work. ⁓
Philip (13:21) Yeah, at this point, I would still call it instead of workflow automation, workflow reimagination. And here’s why. Why I call it workflow reimagination is all workflows have been designed with humans stuck in each and every step of the way. Any piece of software that has been built is a means of codifying what it meant for humans to do the work.
to also visualize each and every step, provide reports to up, provide reports down, provide reports sideways, and it is built on the structure that humans are going to do all of the work. When you reimagine a workflow based on what software can do at this point of time, Based on agentic AI. So then you reimagine as AI first and AI native.
What that brings forth at this point of time is where do humans need to be in the loop? Where do humans need to review? What do humans need to audit? With a path towards, if this can be fully re-imagined, humans just need to review slash audit. That is a paradigm shift with respect to how AI re-imagines workflows.
I do believe this is the time to reimagine workflows and we are seeing, we are already delivering for our customers, reimagining say just workflows slash processes, one or many.
reimagining just entire orgs inside of companies like I say like we’ll reimagine operations, we will reimagine HR, we will reimagine accounting, we’ll reimagine IT and the likes and also reimagining entire companies. What does it take to truly set up a company across all of these divisions AI first and leverage all of the benefits?
So this is where the reimagination is happening and it is an amazing time, a time of our lives.
Federico Ramallo (15:36) Right, because I mean, ⁓ when I talk about my experiences scaling my company and ⁓ bringing more employees, you face the N plus one problem of ⁓ big systems, mean, similar things happen when you add more.
highways to a city, right? Or you add more lanes to a road, right? ⁓ I mean, the system complexity increases exponentially because every node becomes ⁓ possible, you know. ⁓
bottleneck or issue, right? And in the case of the cities, you increase the roads and now you have more traffic, right? Which happened, you increase it to have more flow of traffic, but then you have less flow of traffic because no more people use it, right? So ⁓ I find something similar happening when I bring more people into the company because now instead of, if I have to make a decision and talk to my other part of the brain and then make the decision, right?
happen
so much faster than if I delegate a task and I have to talk to somebody that has to talk to somebody, right? ⁓ So I can see how, ⁓ you know, if we reimagine how these processes are going to be run, then you have agents talking to agents or, you know, how we want to implement it. ⁓
that would allow us to duplicate the same concept that I’m talking about that, you know, everything happens within my mind, right? ⁓ And it’s, you the communication is much faster, the decision making is much faster, right? ⁓ Is that, does that translate to what you’re trying to convey?
Philip (17:23) Yes,
exactly. So you think of this across like any transformation that we do, right? It’s across the people, the processing, the process and the tools, right? So now because of the technological shift, there has been a huge improvement where we can build agentic AI systems that deliver amazing amounts of value. Now,
for humans to be reviewers slash approvers slash people being in the loop they are solving higher quality work.
than lower quality work. All of the lower level work, more than lower quality, let me rephrase it, higher level work than lower level work. All of the lower level work is automated. It’s the higher level work that people have to do. They have to understand problems deeply and then solve them deeply. So.
This, as you mentioned, one of the effects is that the coordination tax of having like so many people in so many processes and that exponentially growing that actually goes off.
That is one effect of it. The second and the most important is that efficiency gains go through the roof. The speed at which you are able to do business goes through the roof. And so many opportunities that you have are totally uncovered.
Now you’re looking at, that’s a problem we can afford to solve. That’s a problem we can afford to solve. This is a feature that we can afford to build. This is a new area where we can expand on. And that’s the whole art of the possible that expands with agentic AI reimagination.
Federico Ramallo (19:08) Interesting. So how do you test this system?
Philip (19:15) Yeah, so…
What we have found in all of our agent-academy limitations, right? It is the structure of, say, if it is a workflow and we are reimagining it, building a reimagined workflow is not the complex part at all. It is ensuring, as in any process, any workflow or any process, there are a set of inputs, there is a process,
and a set of outcomes slash outputs. This is all, right? On a very simplistic note. Can you define or look at all of the possible inputs?
Now, the system is built for all of these possible inputs are regenerating the best possible output. And agenda systems are probabilistic systems, right? They’re non-deterministic. So now that is where, you know, our entire systems becomes, our entire software engineering process becomes a group of AI builders and a group of AI validators.
The AI builders build the systems and the AI validators ensure that this system performs to whatever level of quality that is required in the field based on all of the inputs, possible inputs that can come through and the outcomes slash outputs that are needed to be generated. So it gives needs all of these AI validators need an obsessive level of focus on the data and the outcomes.
Federico Ramallo (20:52) So the more expectations you can set, the more accurate the system is going to be.
Philip (21:00) Exactly.
Federico Ramallo (21:02) So it’s up to us humans to provide those sets of expectations.
Philip (21:09) Exactly.
Federico Ramallo (21:11) Right. does the system, is the system being able to learn and suggest new expectations or is that something we need to keep adding into the system?
Philip (21:22) We build a system with the learning component into it. But what happens is that when a system handles, say, a part of data, that it does not process well, that is where the reinforcement learning loop comes in, where there is a human that actually manages this is not actually handled, why it is not handled, and then the system learning loop compounds it.
That is where we get to say, anybody builds an agentic system right now, out of the box it is like 50 % good, 60 % good, because the foundation models are good. What does it take to 80 % quality? People get there fairly easily. From 80 to 90, that needs really good harness, really good agentic engineering and reinforcement learning.
90 to 95 is even more harder than 80 to 90. 95 to 99 is even more harder than 90 to 95. And 99 to 99.9 is harder than getting from 90 to 99. So that…
is where the compounding effective data comes in with respect to all of the different edge cases and how we can handle it. And so that is where agentic engineering becomes a lot more important. Do we have the right frameworks in place? Do you have the right reinforcement learning loop in place? And are we making sure that we have observability built in and quality metrics for all of these agents that are built in so that we make sure that quality always remains high and are in production grade.
Federico Ramallo (23:01) Right. Right. That’s very interesting because the yeah, I mean, the closer you want to get to the goal, I think it’s a quadratic, ⁓ you know, increased level, the harder it is to get right. ⁓ And the learning.
Philip (23:17) Yep. Exactly.
Federico Ramallo (23:24) Reinforced learning can do a certain thing, it’s up to us to tell, to use a criteria to say yay or nay, to what is the best. ⁓ Because ⁓ you could get answers that are not the right answers. So fixing those would take a human in the loop.
Philip (23:33) Yes.
Yes. Yes, fixing those would definitely take more than that. The analysis of that data and why the system actually failed, all of that takes that human in the loop at this point in time. Yep.
Federico Ramallo (23:48) Ahem.
Right. ⁓
So basically you would have one agent talking to another agent to validate the output of the first agent, right? And then as a third step of validation, you would have a human validating the recommendations of the second agent basically. Right. ⁓
Philip (24:11) Yep.
Yes.
Federico Ramallo (24:22) that’s based on the current capacitors of the models because this is evolving very quickly.
Philip (24:28) Exactly.
And that is what I’m very, very excited for, Federico. Right? All that we are talking about is based on the history that we’ve seen and where we are currently with the models. The model capabilities are growing by the day. And we do believe that the foundational model capabilities become better and better as we grow.
and agentic engineering there are going to be so many gains in agentic engineering structure, agentic engineering infrastructure with lowering costs in it for infrastructure so there are like huge huge gains to be ⁓ made.
Federico Ramallo (25:08) Right.
So, I mean, ⁓ I’m a little bit skeptic on the capacity of the models, Of the current capacity of the model, mostly because of how much hype there is, you know, in the world. Everybody talks about, ⁓ in five minutes, I created an agent that, you know, it’s generating a million dollar in revenue or, you know, I’m just paraphrasing, but, know, people say crazy things like that, which…
I read it and I’m thinking that I don’t think that’s necessarily true, right? I mean, there’s so many things that could go wrong with that, right? So, you know, what approach would you give to a skeptic so he can build trust on the models and the system that you built so ⁓ he can trust and delegate more into it?
Philip (25:43) Yeah. Yeah.
So I’ll address this twofold, One, as with any… ⁓
Federico Ramallo (26:09) Yes.
Philip (26:12) technological shift or new wave, right? There is this gold rush. You always have these people who are the charlatans, the snake oil salesmen, who come around and bluff and who kind of position themselves as somebody great and all of that. I don’t even fault them. It is in every single field. Any single field there is some growth that it also attracts these kinds of people like these grifters. ⁓
I don’t pay any attention to them. What we are focused on is truly what are actual business gains? What can be driven with AI at this point of time, with agentic engineering? Can we, that’s what I said, can we reimagine a business process or workflow with AI?
Can we reimagine a bunch of business processes or workflows with AI? Can we reimagine an entire operational org with AI? Can we reimagine a company with AI? That is totally possible at this point. And that’s what we have unlocked. And that’s what we are going after.
Federico Ramallo (27:08) Great.
Philip (27:26) And I know all of these people come over with like, we’re gonna do this in one week. And I built this in one day and I made like $100 million or something like that for clicks. That’s not, you business folks or C-Speed leaders are not buying those. Yeah.
Federico Ramallo (27:36) You
Right, right. Yeah, I
when I read about your approach, I find it very interesting because you are ⁓ de-risking the implementations and you are providing ⁓ real business benefits, right? And… ⁓
Philip (27:58) Yes.
Federico Ramallo (28:00) At the end of the day, on every organization, there’s a maturity towards how much they can implement at a time. ⁓ So I see your ⁓ approach where you accompany these companies throughout their maturity process, and you leave them, but ⁓ with specific goals and ⁓ gains that are measurable.
and also the risk in the situation because it’s easy to say, we can automate everything and then, you know, it can blow up, right? ⁓ So that’s what I find refreshing and interesting about your approach.
Philip (28:41) Thank you very much, Federico. And see, we operate with generally mid-market and for S &P 500 companies and so on. There isn’t, ⁓ I know there’s been some amount of FOMO ⁓ and people have built, I spent some, ⁓ three million dollars here and there down AI with everybody claiming that they can build AI.
What we have demonstrated is that we are easily from TechJs, we are easily in the 0.1 percentage of companies that claim to build AI and can actually build AI. And when you say like the same de-risking part, we sign off on outcomes. We’re not selling hours, we’re not selling bodies, we’re sign off on outcomes of what we can actually deliver and deliver those and for those outcomes of what we are paid.
Right? And so we totally, there’s customers, we know that their success is our success. And that’s how we structure and build ⁓ our deals.
Federico Ramallo (29:46) That way you are fully committed to the success of the company and of your clients.
Philip (29:51) Yes.
Federico Ramallo (29:53) And I think that’s amazing because that, ⁓ you know, ⁓ clears out any charlatanery out of your approach. Yeah.
Philip (30:05) Exactly, right? So if we don’t deliver, we don’t get paid. It’s not like, we claim
that we can build something magical and we start charging them like, you know, every, every week and every month. And then at the end of the six months, we’re saying like, sorry, we tried, we couldn’t get it. Which seemed to be what a lot of other companies seem to be have done. And it’s very interesting, also sad in one sense that we have had to pick up projects.
Federico Ramallo (30:11) Right.
Philip (30:31) that were done by other, you know, some vendors. And we realized that none of the work that they had actually done could be even reused. We had to just trash that.
Federico Ramallo (30:42) Right, that’s such a sadness because that company ⁓ trusted in that vendor, invested the time and the money and the effort, and then they had to start fresh.
Philip (30:50) Yep.
Yes.
Federico Ramallo (30:56) So ⁓ talking about mistakes, what other common mistakes you see when people try to add AI into products?
Philip (31:06) Yeah, so these are common mistakes, right? ⁓ And I do believe us consider, know, ⁓ fortunate for me to learn it, you know, inside of Google before even ChadGPD came out. And also for us to, you know, because we are in the frontier and we are learning, we’ve also built, we’ve made a mistake, so we’ve learned, we’ve quickly pivoted and so on. So we are already battle-hardened to go over and deliver. One.
The biggest mistake that people do is instead of reimagining their workflows, their process and orgs, they slap AI on top of existing workflows, which are usually highly inefficient. They spend a lot of money and say, oh, we gained 10%, 20%, 30 % here and there, but they don’t really see, you know, that is a ton of money wasted. One. Number two,
without having the barrier to entry to building A is zero, right? You have an open API call, you know, open API API call and everybody calls, we have an AI software. Nobody’s building, you know, general software anymore. Everybody’s building AI. And because the barrier to entry to building is very low, everybody builds something. They’re so excited by this proof of concept or, you know, the model’s just delivering something single shot.
Federico Ramallo (32:14) Yeah
Philip (32:32) and they think that is the end of it. Not really understanding the same part about like, hey, building quality is about being obsessive about the data, understanding all of the inputs, then building great processing and delivering phenomenal outcomes. You can get to that 80 % and call it done, but that 80%, you cannot deploy in production.
Federico Ramallo (32:54) Right.
Philip (32:55) It works only four out of five times, right? That’s not good in a business. You need to get it to 99 % plus 99.9%. How do you get to that? Those are all like common mistakes that people make, ⁓ that we have seen. And we very often have to advise the customers as well against that kind of decision making.
Federico Ramallo (33:18) Right, it’s a lack of vision of what the AI could do on the business, right? And ⁓ what you’re talking about is understanding all the possible use cases, which is a lot of work because it’s basically, know, I always remind when, you when you play lottery, right? ⁓
Philip (33:23) Yes.
Federico Ramallo (33:42) You know, it sounds so easy, but when you do the math, right, it’s so unlikely that you win. Right. So ⁓ I don’t remember the numbers right now, but, you know, you know, it’s one in billions, one in millions. Right. It is more likely to get hit by a car or by an airplane or whatever. ⁓ But anyway, people still believe in that. Right. So there is this irrational behavior. ⁓ So.
Philip (33:49) Exactly.
Hahaha
Federico Ramallo (34:10) It’s easy to find the most common use cases, but then finding all the use cases, that’s where it gets harder. Because at the end of the day, if you have an automation that you cannot trust, because you know…
Three out of five fails. It does correctly, but two fails, right? Even if it’s four out of five, right? Still, we have a high percentage of success, of failure, right? So, automating is basically building that reliability, right? I mean, on cars, they talk about…
that cars are unreliable, there are some brands that have unreliability issues. But if you compare that with cars in the 70s, the percentage of reliability, we’re talking about 3 % versus 20 % in the 70s, right? So we as a society,
demanded more on the cars, car industry, to do better. And yes, there is a percentage that fails relatively to one another, right? But most of them are reliable, right? Plus in a car, if it breaks, you can go to the side, right? But here, the…
The comparison, I guess it will be better with airplanes, right? Airplanes cannot fail, right? You cannot go to the side of the road with an airplane, right? It has to work, right? And that means that every part that you put in an airplane needs to be tested thoroughly, right? Because you could have a small issue of cable getting brushed and then, you know, that part fails, right?
Philip (35:57) Exactly, Friedrich. So let me add on to that, right? It’s about what that’s why understanding what problem you’re solving makes it really, really important. So say you’re working in something creative or a generator.
like you’re working on the estimation side and so on. So today I met with a customer working on generating estimates and he was like, hey, if you give me 90 % and greater accuracy, that is better than my average human reviewer or somebody that I need to train.
So that becomes the barrier, which is nine out of 10 times. We will very easily be there. We’re still like, oh, we’ll easily be greater than 95 % in the first launch, and then we can improve to 99%. So there, there is a huge margin for error because you can have humans in the loop, they can review, they can update, and so on. Again, it’s the same.
Federico Ramallo (36:50) Right.
Philip (37:00) How do you get to 99.9999 % or 100 % like whether it is medical, related to human safety, how do you make sure you get there? So those are completely different problems with respect to how we approach agentic engineering.
and to have not apply the same pain strokes across, which people generally do that because they don’t wrap the complexity around all of those things and say, if you can do this, then you must be able to do that, which is not the case.
Federico Ramallo (37:33) Right, right. They make correlations that they shouldn’t.
Philip (37:38) Exactly.
Federico Ramallo (37:40) And how do you work on the security aspects of it? mean, how do you ⁓ work to make sure that the… We talk about expected outcomes, right? But ⁓ how do you work to get, in case you get unexpected incomes that would trigger an alarm, right?
Philip (38:02) Agreed. purely from an AI perspective, right? So that’s why we build observability, monitoring, and proactive security in preventing prompt injection attacks and the likes. Right? So we build all of those things. From an overall general security,
Security standards for general software has not changed. Like we follow the OWASP standards, the latest and greatest, make sure all of our applications are, you know, can pass any compliance test, whether it is SOC 2 and the likes. So that’s what we anchor to. What we are seeing purely from a security perspective is that a lot of the transformations that we do,
are done under a B2B, kind of like a walled garden. So those things are easier. For example, like say that we build an agentic system for an operational team, where only these 100 employees are ever going to have access.
The challenge then becomes when you use it, you open up the endpoints to larger consumers and the whole world. So both of these need very different security approaches.
Federico Ramallo (39:19) Right.
Philip (39:24) All right. Yeah.
Federico Ramallo (39:24) Yes, because it’s a level of trustworthiness
you have on the other side.
Philip (39:30) Exactly. And most of the AI transformations that what I’ve seen, you know, and what we do are for businesses. So there is a trusted network already. So people who are only authenticated usually get into those systems. And there are very few places where you actually process input from the outside world.
Federico Ramallo (39:48) Right, having an authenticated user with an authenticated, ⁓ with an audit log of what the user is doing, eventually that gives you accountability and that allows you to trust more on that user, person or system.
Philip (40:07) Yep, exactly.
Federico Ramallo (40:09) Right. if, I there’s always the, know, the, ⁓ you know, if that account gets hacked, if that system gets hacked, right? There’s always that, but you’re working on an environment of trustworthiness enough for ⁓ the expected behavior of the user, right?
Philip (40:34) Yes, absolutely, Fredrika. So we’re definitely ⁓ working on that. Even with respect to say, gets access and so on, I think standard security protocols work, right? With respect to, hey, how you have user access control? How do you handle breach? ⁓ How do you have an audit log? All of those things are very standard security practices that we’ve always had building applications.
Federico Ramallo (40:59) interesting so
⁓ You talk about this ⁓ new ecosystem of automation. How close are we to have a ⁓ system like Jarvis on Ironman, where you can have a conversation with the assistant agent, whatever you call it, ⁓ where you can ask high-level complexity tasks and would understand everything and figure out how to do it.
Philip (41:18) Yeah.
Hmm, interesting. So I’ve not watched Iron Man, but I know Jarvis because it’s coming. Comminate right? So. Intelligence as a whole actually depends. And for me I did. I don’t see it like there is going to be this one definition of AGI or ASI because people’s definitions differ, right?
Federico Ramallo (41:39) Yeah
Philip (42:02) What I was like struggling to think of, you know, how you can define AGI or ASI in that.
But there was this one incident last year that happened. This has been, I think last March, 11 months back.
sesame.com had launched their demo voice mode and at that time that was like amazing, mind-blowing. So I had a demo, just one demo and I was like you know click and I was like talking to it five minutes and then it’s like in 90 seconds in I stopped talking to it like I was talking to a bot and I was almost conversing to it like a human would. And there’s something flipped in my mind. So
What we’re looking for from capability perspective on any mode, whether it is multimodal engagement or any level of information is where we start preferring talking or engaging with this more than we would with respect to a human. Can it become give a trusted input more than it can give a human being? There isn’t this change that happens. You understand? Like with that bot, say,
I do believe in a few years we’re gonna have like conversational bots that you would prefer talking to more than any call center agent.
Federico Ramallo (43:31) Right. Right.
Philip (43:33) You get it right?
Say they might be you’re calling the airline ⁓ and they are actually responding to just, you know, a ticket cancellation this and so on. But this conversation bot gives you a phenomenal experience.
Now, instead of going through the IVR, waiting for it, then talking to an agent, you just have had a really good conversation. Now you would prefer this over going to and talking to that agent. So that is what I believe is the inflection point in having how AI would actually take over lots of, you know, ⁓ lots of workflows.
That is the inflection point where you start saying like, ⁓ this is artificial general intelligence, artificial super intelligence, and we are on the path for it. We are never going to have like a general, I do believe, an agreement on all of those things because there are always problems in the frontier that need to be solved. ⁓ The models are all, you know, just only as good as the data that they’re being trained on. ⁓
So they reflect some of the same biases, the same issues. So if you’ve just as a consumer, if you’ve chatted with the model, they will always say you did a great job, kinda, you not necessarily ⁓ the truth and seeing what really is the truth. So AI has quite a bit of way to actually to go before we can be a really good from AI can solidly detect cancer. It can come up with, you know, ⁓
Federico Ramallo (44:53) Right.
Philip (45:12) say unique solutions to cancer-solving and all of that. So that’s the path that we need to take and there is quite a bit of way to get there. That’s my opinion.
Federico Ramallo (45:24) Amazing, amazing. ⁓ So the ⁓ interfaces that you’re building now are conversational agents that can do ⁓ what we used to do in lower level tasks, right?
Philip (45:37) Mm-hmm. Yep.
Federico Ramallo (45:39) Yeah,
Jarvis on Iron Man, Iron Man would say, ⁓ just find me a new material that could withstand whatever blast I got on the last scene, right, or whatever it was, right? In a very high level ⁓ complexity that requires a lot of thinking. And the robot would say, yes, I found this material.
Philip (45:57) Yeah.
Federico Ramallo (46:08) you know, in this corner of the world that it’s going to be amazing for this. Right. I already ordered, you know, the amount we need to build a new suit. Right. ⁓ Which, you know, the movie was, I think, 1998 was the first Iron Man. I don’t know. But, you know, at that time, everybody was blown away by, you know, the concept. Right. ⁓ So that’s why I go to that. Right. The the idea of.
Philip (46:31) guess.
Federico Ramallo (46:32) being able to have this conversation with somebody, with an agent that understand the context and the intention and can split that into smaller tasks and then go and achieve them, right?
Philip (46:46) Yes, yes, Virgina.
Federico Ramallo (46:49) That’s the dream.
Philip (46:51) That is the dream, but it is quite always forward to have something like that. And if we have something like that, the world will never be the same again.
Federico Ramallo (47:02) Right, right. So, so
far the agents, we can talk to the agents and they understand the context so we can pinpoint specific tasks for them to do, right? That’s the current stage of evolution, guess, Yeah. And analysis, you can ask to analyze.
Philip (47:13) Yes.
Yes.
Federico Ramallo (47:22) a lot of data and we’ll do that very quickly and find the correlations. ⁓ And like if you have a database where you have to be very specific on the query, otherwise you get the wrong answer.
Philip (47:35) Yeah, the beauty is that the models can handle unstructured data. And most data in the world is unstructured, right? Databases are not commonly in that they are structured. So that’s the beauty of the models as well, when you’re handling data.
Federico Ramallo (47:39) Right.
Right.
Right, we structured the data because we ought to, right? We needed to.
Amazing, So we’re running almost out of time, but I want to ask you one last question. What advice would you give founders that are trying to build the next AI thing, right? ⁓ And what advice would you give them?
Philip (48:14) Absolutely. So this is something that’s a pattern that I’m always seeing, right? The advice that I generally give is there are so many problems out there in the world that are waiting to be solved and people would be willing to pay money if you solve that. So focus on the problem deeply.
deeply focus on the problem. Your solution should be with first principles thinking. So you’ve solved it really, really well. You’ve obsessed over the problem and you solve that really, really well. And you solve that tastefully.
As long as you do that, there is always going to be value in that work, in what you do.
Federico Ramallo (48:56) Right.
Philip (48:57) Coding may come across and become a commodity. Engineering work might become a commodity, but deeply solving problems and having great solutions with taste, that is still going to require a lot of thinking and people will be willing to pay a lot of money for that.
Federico Ramallo (49:16) Right, right. So in your opinion, people say that SAS are going to be dead because now you can vibe code whatever set platform you need instead of hiring a subscription. But you’re talking about understanding deeply the problems and being able to provide solution for it.
So what are your thoughts on the future for SaaS companies?
Philip (49:47) I do believe it will all evolve. ⁓ Say the era of, I think, maybe SAS charging hundreds of thousands of dollars for being a tiny wrapper, they’re all gone. They’re all gone because the cost of building software comes across. What will happen? This is my hypothesis again on SAS is that.
Federico Ramallo (50:01) Right.
Philip (50:12) Still, people who are paying, $20 for a project management tool aren’t going to build it, host it, maintain it. For them, they would happily pay $20 of subscription. But it’s the businesses who actually pay a lot more. And when they pay a lot more, they are not going to be any more going to be paying a value for a beautiful portal that they have to input data into and get extra out of it. It’s about how much value…
That piece of software drives their business. It is the same, right? We help re-imagine companies. How much value are we driving? Are we helping them save phenomenal amounts of costs? Are we driving them, you know, exponential revenue? If we create value, we can capture some of that value. If you don’t create value, and that value creation can basically be done by somebody inside the company or someone, then you are not adding any value.
So it is value creation and value capture.
Federico Ramallo (51:11) right.
Right, so some SaaS that are providing very little value are going to be replaced by vibe coding or any other alternative. And that’s why I think that your ⁓ recommendation of deeply understanding the problems resonate more because SaaS companies that are deeply understanding the problem and are providing a solution, they probably will need to adapt, but they will be able to carry on business, right?
Philip (51:41) Yes.
Federico Ramallo (51:42) Amazing. Philip, are running out of time. I truly appreciate you being here today. Any final remarks before we wrap it up?
Philip (51:51) Lovely chatting with you, Federica. It’s been a great chat. I hope our listeners ⁓ get lots of value. And what I would leave to say is like, hey, the world is going to be reimagined. There are only going to be two kinds of companies. Companies that reimagine themselves with AI and companies that go out of business.
And as the world reimagines itself, we want to be the people who help the world reimagine. It’s a fantastic time to be alive, a fantastic time to build, a fantastic time to solve problems.
Federico Ramallo (52:25) Amazing, amazing. And we’re going to leave the links for people to reach out to you on the note of this episode. So, you know, if anybody’s looking to automate their ⁓ workflows with AI, improve how they operate and thrive in the AI world, they should reach out to you. ⁓
Philip (52:49) Thank you so very much for having me on, Friedrich.
Federico Ramallo (52:54) Thank you, Philippe.