How AI Agents Will Revolutionise Work by 2030 — Suman Kanuganti

Scaling a company means more people, more tools, well, more of everything. But you don't get just more of good stuff, you get more of bad stuff as well. More meetings, more scattered knowledge, chaos everywhere, or at least more chaos. So the argument in 2025 is that the solution isn't a headcount or another platform. It's capturing your team's know-how, structuring it, and training AI personas that can work alongside you. Here's a couple of things you can expect from today's discussion: how to capture and store team's know-how, why role-specific AI personas beat generic assistants, how parallel workflows remove bottlenecks, and we'll touch on AI governance for enterprises. Hey, what's going on everyone? I'm Mike Sirius and this is Mastering Tech Growth. Every week we chat with industry leaders to bring you no-fluff tips on how to grow. If you enjoy the show, please do me a favor, leave us a review and drop us your biggest growth questions in the comments or visit us at masteringtechgrowth.com. Today I'm joined by Suman Kanugandi. Co-founder and CEO of Personal.ai. He's a two-time venture-backed founder recognized by Forbes, Smithsonian, and New York Times for his work in AI innovation. At Personal.ai, he's building your digital mind. AI personas trained on your knowledge and voice designed to boost team performance at a fraction of the cost. Welcome, Shimon.

Glad to be here, Mike.

So when I say Mastering Tech Growth, what comes to mind?

Master in— say it again.

Mastering Tech Growth, what comes to mind?

Tech Growth. Oh, okay. Um, it's about, uh, revenue per employee, right? The idea of how do you continue growing without thinking about adding a linear capacity of the headcounts and ensure your EBITDA margins are, you know, net positive because that's a milestone within the startup environment while still keeping up the momentum with the growth activities or the business.

That's such a scientific answer, Shiva. That's just one like chung chung chung recipe. Other guests, you know, they go like theoretical and maybe it's people, you know, maybe it's tools. You just went like, uh, deep level revenue per employee, EBITDA margins. So is that something you as a CEO track all the time to make sure that you are in the right spot? Or is that like, let me rephrase this, like from all the things you can choose, right? What influences tech venture growth? You got marketing, you got sales, you got innovation, you got maybe toolings, maybe employees, maybe team culture or organizational culture, maybe it's purpose, like from all different elements that have direct or indirect influence towards business growth. If you choose one, what that would be?

Customer obsession.

Customer obsession. I love that. Okay. Can you give us more like what is a customer obsession to you?

You are a business when customers love your product and they're benefiting from the value that you are creating. So there is a transaction that is happening between you and your customer. Unless and until the customer problems are fulfilled and they are extremely happy and they are growing with your company, you know, the rest of it doesn't really matter. How much of our marketing you do without customers really getting the value, it doesn't really help. You know, partly the reason we also see within the markets a sudden spike and increase in a specific tool, uh, shoot up the charts because there is low acquisition cost, but at the same time, you know, retention and stickiness will become a problem and Immediately you have a, you know, retailer problem, right? Customer obsession is really critical for Personal AI, you know, one, because the core fundamental issues and the problems that we solve are very personal in nature, very private in nature, very precise in nature to the customers. And that basically defines every other function in the company, either it be marketing or sales strategy or pricing. Are how many architects do we need, are what kind of developer engineering capacity that we would have to plan for depending on what kind of customers, you know, we are seeing growth in.

I love that. That's a great answer. We had customer obsession with other guests as well. It is— you want to say that like, this is almost like, duh, like it's like you're solving someone's problem, someone's paying you, like it shouldn't be that difficult to to realize that you should be obsessed with providing value to the customer. But it's not like when you talk with other founders who don't come from that domain of dealing with customers, maybe they're more technical, maybe they were building tools all their life, they don't think that would be their biggest differentiator. They somehow think maybe it's a tool, maybe it's a technology, maybe it's a cost of maintaining code or something like that. But the customer obsession at the end of the day is the recipe, right? I love that.

Yeah, I mean, all those kind of support your desire to provide value, right? Because at the end of the day, if you don't offer the value to the end customer, what business are you building?

You could probably market it and shove it people's throats. I see that happening quite often. But at the end of the day, you don't have that stickiness, right? You don't have that retention. Which is taking place. So you're always going to be burning money on ads to get some conversions, but the lifetime value is just, it's just not there, right? It's not sustainable business model. Okay, that's cool. Cool. Let's get into it. So the very first topic is how do you capture team's know-how? So let me ask this. Actually, let me quote this. So I got this report which was done by Asana, which is people behind like Jira and Confluence tooling. So they did that last year. So they saying that 53% of workers time spent is spent on busy work. So that means they communicating, chatting, and guess what they're chatting about? Well, they're usually looking for information of how to do their work, or they chasing for a status updates on their work, right? So if we ask this, Roman, like, what's the first step you would do if you come to your organization to actually capture, you know, the scattered team's know-how into a structured knowledge vault, which you can then train AIs on?

It's a very in-depth question depending on, you know, what company stage you are in. You know, some customers that we deal with, they already have lots of structured data except that it is not captured. There is no value that is being created from that structured data. Some of the customers are, you know, the data is everywhere. It's in Slack threads, it's in channels, it's in documents and multiple different tools. The most interesting thing is, at the end of the day, still it comes down to really smart people who are gathering and keeping up with all the context that is needed to drive certain decisions or drive certain answers to these problems. Now, if you think about what do we do with AI, we are pushing AI to be collaborating with people alongside with them. Because if we can think about the most standard way of creating structured data in folders, but it may not work for everybody. But what works for everybody is definitely the people and their expertise. So are we doing this more to solve the problem of what that structure looks like, or are we simply able to create AI is based on these expert people because they are the ones who are making decisions on what that appropriate context looks like. So where I'm getting to is organizing the information and the structure based on your employees and the roles that you already have decided, because there is a specific strategy and organizational strategy internally, may go a long way sometimes rather than trying to attack the problem from a technology lens on, oh damn, I have 20 million places where my data is, I want to put all of them into one place. you know, for some magic to happen, which normally doesn't happen.

So it doesn't work like that then. So I can't— I would not expect them to just dump everything into a bucket, a folder, and just go like, oh, it's gonna now work for everyone. Like, from this we now can— that just doesn't work like that. Okay, exactly. So you mentioned— yep, you mentioned it. Finish your thoughts, finish your thought.

Uh, I was just— I was gonna finish my thought by saying, um, So we encourage thinking through from a lens of the role, um, a human or an expert AI is supposed to perform, and the decisions and the tasks that human or an AI is supposed to perform, and what goes into that specific role, right? Very similar to how you would train a human, you would train the AI with that specific context. And our architecture supports that mindset, that line of thinking.

Okay, that's a fairly simple way to think about this. So effectively, you have an AI agent, right? So that's going to be fulfilling a certain role within the business. For the sake of argument, let's say it's a customer support role. So they will be doing that role. So then you ask a question, you know, what skills or what responsibilities that role will have. And effectively what information you need or AI persona needs to actually fulfill those responsibilities within that role.

Yes. And if you want to build on top of it, the most efficient systems and specifically within the AI and the training is when you incorporate the idea of continuous training where people are already doing their work. For example, in that particular scenario, Mike, if there is indeed a human who is able to respond that specific response, and that goes back into the AI as a mechanism to continuously train, I think that's a place to be in. That way we are not thinking through the lens of, oh, I need to automate this training that is separate from my existing workflow and how I behave and how I interact with my customers. Any other use case for that matter for my teams.

Okay, so is that like a must-have requirement with Aising that it should be continuously trained or continuously training? And if we can't provide that, that's just not a good use case for AI persona?

There are hardly few number of use cases where the information doesn't change, right? If it is set up for loss, okay, the loss are stagnant and it doesn't change for a long period of time. And yet, you know, even the loss example is, you know, evolving pretty rapidly. So you got to keep things up to date, otherwise the persona is already outdated, right?

Okay.

Yeah, the way we designed our persona is based on how individual people kind of retain the knowledge over a period of time and predict the next decision that they would have, right? So the memory associated fundamentally with the model is time-bound or time-based. So what that, what that means is if there is a newer information, the model always looks up for the bias in the recency of that particular, you know, memory to be able to answer. So you're always up to date. Um, so yeah, I would I would say it is almost critical to incorporate into your day-to-day activities. You get this compounding benefit rather than a bot that has been created and then it had novelty sitting on the table, but then it loses the novelty.

Yeah, I get what you're saying. Let me take us a little bit step back. You started saying, when I asked the question, you said, There's a context to it, right? Like some organizations already have structured information, you know, some will not. It's going to be scattered everywhere. Phone calls, Slack messages, some documents, internal pages and whatnot. So if I want to, like if I'm entering this domain for the first time and we're having this conversation and you basically giving what you just gave, like you need to have this, think about as it's a role, you need continuously be learning. What is the tech or processes I can put in to actually capture this information? Are we talking about some type of integrations, you know, API getting and getting fed information, or are we talking about people manually uploading stuff somewhere? Like, what's, what's, what it looks like in practical terms to continuously feed the right information to AI persona?

I would start off with, um, What's your problem? Like, what are you trying to solve? We can role-play it. Let's just say—

Let's do AI per— sorry, let's do a podcast producer. That's literally what I'm building on my free—

We're talking about Mike AI here.

Yeah, here you go. Yeah. So, I need help with that.

Podcast producer. What are you intending this podcast producer persona to achieve for you, Mike?

So I wanted to help me to prepare for episodes, right? So I wanted to be a little bit like a market aware as well, what's trending, you know, what are interesting topics. But once we have a topic in place, I go like, you know, this is what we're going to be talking about. Maybe I'll provide everything I can gather about the guest, my ideas about good questions and stuff like that. And it sort of needs to fill the gap to make it a good, interesting episode to our define listener persona.

Okay. How do you do that right now?

How I do that right now, a lot is manual research, which I believe AI could do that for me. So, you know, I go look at topic, you know, trending topics, YouTube, Twitter, Reddit. I look at the, you know, bigger questions around the topics. I look at what people are saying in the comments to draw some real use cases and stuff like that. And then I, of course, sprinkle my own sort of views on top of that. And that just builds up what we call a structured episode, you know, has a couple of takeaways, but there's a core message or the argument which I introduced in the beginning, which we want to give to our listeners. It's like, that's, you know, that's the core idea, right? And then we have a couple of takeaways and a handful of questions to lead the episode.

Okay. So here is how I would think about it. Think about you have 3 people who are working on the following things. One is specialized in bringing the latest and the greatest what's happening in the market with the AI, right? This could be based on certain publications, certain, you know, specific target tweets that you're talking about. You know, for example, for myself, I have a persona for Ben Sharf from Strategy Chariot. I follow him pretty closely so that way I get, you know, pretty much, you know, most up to date on what's happening with the AIs and trends, right? So that's my like market research persona. Uh, and the interesting thing is, uh, the persona is defined to kind of ensure the topics that is going in are suitable or in context with what I do on my day-to-day basis, right? That's like one— I'm thinking from a lens of the role, but you can also always think about from a persona. So that's the market research persona. And the second thing that you would have would be more, more or less like Mike AI, right? So Mike AI would be how does Mike communicate with the guest on the podcast, you know, while live on the interviews, what kind of questions that he would ask based on what the guest is communicating with them, right? So that, that's more like Mike AI. And the third one would be more along the lines of an expert guest persona. So this would be all the guests that you have had in the past interviews going into the persona, and as well as your target guests that are, uh, that you are, you know, hoping to speak to, they will also go in. Let's just say if you have a new guest that is coming up next week, uh, you know, a LinkedIn profile or any of that information would also go into that persona, right? And when you have these 3 personas on a day-to-day basis, the market researcher is getting up to date, you know, automatically based on, you know, what you are topics and where you want to follow or whom you want to follow. And your MyKai is getting information from basically your, you know, I guess, reverse side MP4 or audio files uploads that are going into the system. And then finally, the third one, it could be a combination of like manual entries because it's extremely difficult to scrape like Instagram and Twitter because they're closing down the doors, right? But you can always take the LinkedIn profiles and then push into guest persona. So when you have the newer guests, you would say, hey, I'm having this newer guest, uh, what are the latest trends in the last, uh, 2 weeks, uh, which you would tag market researcher persona based on the profile and based on the focus of this particular guest, right? So then you would basically narrow it down a highly contextual, highly personalized set of topics for that particular guest that is in alignment with what, what's happening in the market. Uh, and then you would ask Mike AI, it's like, hey, based on these topics, I want to come up with a very, um, uh, a set of like questions, an interview guide if you will, that I can have a very intelligent and a reasonable, entertaining, you know, informative conversation with this particular guest. And then boom, boom, boom, you have this interview.

Okay.

See how like this specific expert persona, and there is zero chances for, you know, that output to be not, you know, like outside the context or outside the boundaries of how you would think. Okay.

So I love that, by the way. I'm gonna build this. I did this yesterday. That's a, that's a, It's an interesting approach that you mentioned, you know, the guest persona. I didn't have that in my mind as sort of like a guest persona, which tries to simulate, you know, the future guest. That's actually very cool. Okay, so we would have those 3 personas. They would be talking with each other. How do I feed that info? I mean, is that API integrations at personal.ai?

Yeah, combination of everything, right? So, okay. Now, you know, you are using Riverside. There must be somewhere where you have all your recordings, right? If they are, let's just say, you know, in a folder in a Google Drive or in a Microsoft Drive, you can automatically integrate that into the memory in the system. So every time you have a new recording, it automatically gets trained, so you don't have to, you know, actually do that. Okay. Um, for the Marker Researcher, like, similar. And if there are specific premium publications that you have that only you have access to. There could be API-level integrations that are needed. Most of them are usually one-time setup with existing native integrations. There are workflows that support to manually input in the system. The fourth one I will say is the platform is designed to be collaborative in nature. It's more It's more like an AI-native Slack experience where you have users, which is people who work at a specific company, and you have personas that are very expert AIs that are internal to a specific company, and they all collaborate with each other. When you have these channels, all the channels are remembered. If you have a project channel, that is remembered as well. When you have a specific employee talking to Mike AI, that specific relationship memory is preserved as well. So everything becomes like highly, you know, personalized experience within the platform. Uh, so you will not lose, you know, any tiny bit of context, uh, that is being exchanged within the setting of the company. But at the same time, it has the discreteness and discrete context on what is happening with each of these, you know, channels and DMs, rather than everything going into like one gigantic model.

Okay, so, so if we're feeding the defined personas with the knowledge they, it should have, so there's no structure with thinking too much about it, right? It's just like, hey, I know you're gonna be my market research persona, and here's my newsletters I follow, here's the people I follow, so here's their YouTube channel, here's their blog post, Here's their newsletter and through APIs or other means, we just put everything into that persona. We don't think about structuring that information, right? We just—

Yeah, I mean, I wouldn't entertain customer to start with structuring the data, right? I would entertain customer to start thinking about the problem they're trying to solve and what is needed for that specific problem to solve, right?

And when we say what is needed, we're saying what data is needed or what—

Yeah, what data is needed. —especially in your case, like you are a business user, right? As a business user, you know how you make your decisions. And I asked you that question, like, how do you make those decisions? How you make your decisions today is informing what data goes into the persona.

Okay. How you make decisions. Yep. Okay. That makes perfect sense. So what do you— what wins do you get then if you go a step further and you start structuring that information.

One more time, like, can you rephrase the question?

So what benefits do I get if I invest time to structuring that information?

Well, the benefits would be, again, goes back into you wanted to have a specific interview guide for your guests that are coming in so that way you go in prepared. One, it is saving time. Two, it is increasing your levels of cognitive capacity to a larger degree and awareness around your guests and preparing you for having a very fruitful, productive conversation.

Somada, I think you might have misunderstood. So if we're feeding information into an AI agent and we're not structuring that information at all, we're just feeding into it. If I go a step above and I say that I'm going to structure that information somehow, I'm going to maybe add some metadata around, you know, transcript from YouTube video. If I'm going to take that step further to just structure that, is there any benefits I get from that performance of AI Persona if it gets a little bit more help with the, you know, information it's receiving from structure?

Yeah, it's definitely going to make things faster and more efficient. The reason I was emphasizing less unstructured data is that's not where typically you would start. That's where I was trying to lean into thinking through what you want to solve first before start structuring the data. The platform, when you give, let's just say, lots of unstructured data, has mechanisms we call the memory transformers to take that unstructured data and structure it for you. Right? Okay. So that one step of structuring the data is kind of eliminated specifically for AI accuracy, for AI precision. Hmm. Right? Mm-hmm. If like, let's just say the system such as large language models, when people normally use and build, you know, RAG components to a specific problem, right? Mm-hmm. There is a lot of work that goes into either manually structuring the data in advance or trying, or like writing custom transformers to be able to structure the data to go into RAG and then gather the, collect, you know, right context at that moment in time and then make a call to a large language model to get the inference callback. Um, uh, and in, in, in case of like personal AI, those, the structuring of the data is part of the platform, part of the persona. Okay.

So unless you're doing that something bespoke, you probably don't need to worry about that. And if you would, the benefits you should be expecting is the speed, accuracy, and precision of that AI agent, right? So does that mean our answers will be just better? Less vague or less hallucinations, if I might say, just going to be more on point. Is that what you get if you structure your data?

Yes. Structuring is almost like a prerequisite for the models. And right now we are at a point where the technology has advanced to a point where the platform or the system structure the data for you. Right. And that is the reason how we are able to push like almost zero hallucinations and accuracy levels, you know, that is accurate by the memory that goes into the system. So absolutely.

Okay, so let me ask this. Why do you like, when you compare general assistants, and when I say general assistant, to give example, let's just say ChatGPT, right, which everyone using, just, you know, the one you just log in and you can use it. Like, that's a general assistant, right? It just does everything. And then you have these role-specific personas, right? How better the results are to the same question? Like, is there any quantitative data which allows for people listening to realize how much better dedicated AI personas are?

Good question. From a purely like a value standpoint, there are 3 key values for dedicated personas. Um, one is programmability, two is precision, and three is privacy. Programmability in the sense that when you do have highly, uh, specific needs, uh, that has the depth of the context that, that is beyond a context window limitation it is extremely difficult for you to ground a large language model into large amounts of data. So you will like, you know, tip over like north of 30,000 tokens into kind of hallucination factor. Um, and the idea of the programmability is you can control that expert persona to the degree that you would want, so that it's highly customizable, that otherwise you will not have with a typical GPT. Over ChatGPT or any other LLM. The second thing is the precision. When you do have this context window limitations, the precision comes from the depth of how deep that you can go into your datasets. The traditional ways of RAG over LLM, you would have to put all that context within the context window to be able to construct a response. And sometimes One, you may not be putting the right context into that context window, and two, you may not put enough context into that context window to actually make the decision meaningful and reasonable. You'll get highly precise responses. The third one is the privacy, which is more centered around, well, where is this data going? How is this being used? Is data into one specific bot or GPT used by somebody, someone else. Is there enough discreteness to each of these expert AIs that we guarantee? There is a lot of privacy component to the personal language models that we built. Those are the value side. On a purely metric side, currently we are looking at 100% accuracy results. That's kind of our promise to our customers. And anytime when there are accuracy issues, we go deeper into what exactly is happening. And in most cases, it has to do with either a bug in the structuring of the data or the fundamental memory itself is incorrectly, you know, uploaded or pushed to the model.

Okay. I was actually very interested into that, so I did some digging, and the only Study I found around this was last year, which was done by Cornell University. So they published the paper and effectively they were saying that AI roles or AI personas, if you define it, and they worked with recommender analyzing and devil's advocate personas improved diagnostic accuracy in a multi-agent setup by 80%. Effectively highlighting the a much better approach than using generic ones. Of course, it is end user or not an end user. It's a power user type of segment. I would not expect every living person to go into the notion of AI personas, but it's very clearly where the market is moving, right? Like we're having so much success and everyone is trying to squeeze that out. And another one which I found, which was quite intriguing is that McKenzie, which is, if you know McKenzie Company, they do like a huge research, it's a huge enterprise, right? So their consultants, which is basically 45,000 employees, they basically build AI agents on all their information, which is like 100 years of firm's knowledge. And it's being used 70% in any given day by 45,000 employees. That's insane. And their time saving and efficiency just through the roof. It's such an interesting paper as well to read. So the question follow-up I have, how do you know if AI agent is hallucinating a little bit of underperforming? It's like, do you have like a red flag? I want to say it's like, you know, how do you like when you work with the AI persona and you think you did everything right, you know, the data is there and you know, your prompts are great and stuff like that, but you get an answer and you have that sense is like, I'm not sure about that, right?

Yeah, it's like, like the two ways that we do that. One is something called a personal score. So for every inference call, we generate something called a personal score that will dictate the accuracy levels as well as the, um, you know, how contextual it is to the user as well as to the memory. So it's basically 0 to 100%, um, and you can double-click on that particular score to see if the accuracy is 100% or if there is you know, and elements of doubt. And when there's elements of doubt, the users will have an option to look at kind of the under-the-hood source kind of data to verify what's happening. And if, you know, if the source information is incorrect, uh, or something else is going on, right? So there is a, uh, tooling, if you will, which is the personal score. That's very nice. That will guide the users to, you know, give, you know, benefit of doubt or not benefit of doubt to kind of double-click and they have transparency into the memory layers, which is one of the other cool functionalities, which is you get to see the data that actually goes into the model and structured data. Having that access is pretty useful for our customers. It's definitely useful. Yeah. The second one simply comes down to user acceptance testing. When you have the model, You have specific sets of criteria and tests that you would define for user to say, yes, you know, this is on par, accurate. Yeah. And almost like have a user acceptance, you know, test criteria to sign off on this one. So it's a combination of both human supervision as well as, you know, fundamentally baked in, you know, personal score models that are built into the platform. By the way, I also sent you a paper, which is a paper from NVIDIA around these small language models that also validates some of the architectural choices that we built and pushing out to the customers, which is more around the economic efficiency as well as increase in accuracy and scalability of how the smaller models and highly specialized models could be very beneficial operationally and economically for customers. That's excellent.

We'll add it in the show notes. So question now, can I get my business culture, tone, and some other, I want to say, like global attributes to all the personas? So let me give you maybe an example. Let's say I have two personas. One is customer support and another is a marketing agent or something like that, like WritesBlogs, right? So they have different functions. I'm going to be feeding them with different information so they can perform well, but I want them both to have my business, you know, brand or my business tone. I don't want to train them separately on the same. Yeah, good question.

We could solve it in a couple of different ways. One is, let's just say a simple straightforward would be a brand persona. A brand persona would have your brand guidelines, your existing writings, and maybe past content articles that defines the tone and the style of it. Similar to custom ChatGPTs, there are additional additional custom directives that you can extract and bias the model more towards a specific tone and specific voice. The interesting thing is, given it's a memory-based model, the context will still be very rich when you have the brand guidelines and as well as some of the content examples as part of the persona. It becomes super rich along with the directive. It's Directive Plus. The memory of the brand. So let's just say you do have that, you know, brand AI, right? You know, brand expert AI. Then you have a marketing persona or a customer persona. When there is an inference that is made for a specific, like, a marketing strategy, right? You are giving it to a specific client. Oh, the strategy sounds good, but it doesn't sound like me. You can then ask, hey, this is my marketing strategy. I'm, I'm good with it. Give it, ask your Brand AI saying that, hey, can you please rewrite this brand strategy, you know, so that way then it basically incorporates all the elements of tone and voice. So that's like very simplistic way and also very solid way to get, you know, really good results.

Yeah, I love that. That's excellent. Yeah.

And however you are asking about, okay, how can I get every persona the same thing? You could incorporate some of that same memory into your, uh, like marketing and the customer support personas as well as the same levels of directives as well. Uh, and it may become a little bit advanced use cases on how you would prompt the persona to ensure, uh, that not only give me the marketing strategy, but please update according to my brand guidelines, right? So if you're able to instruct the persona to be very specific in that direction, then you will still be able to achieve the same results. Okay.

So I think the easiest probably, or the path of least resistance is to have a dedicated brand persona, your brand quality assurance persona, or your brand police persona, you know, whatever you want to call it. It would, it would just, it would just concentrate on literally making sure that everything that goes out.

It's like your PR team, you know? Yeah.

And that's the reason why they exist. Yeah. That's a much better name. It's your PR persona. I was trying to come up with a good name. That's like your PR. And then everything you get before you put it out, you would go like, hey, be my devil's advocate. You know, just tell me what's not with the brand or rewrite it to be the brand. Okay. That makes perfect sense. I love that. How can you protect I don't know if that's possible, by the way. So very interested to understand if it's from technical perspective, if it's possible to protect persona to basically drift into irrelevant answers. And I want to give an example. It's like when I deal with ChatGPT and there's a couple of use cases, especially when I needed to reference a source, right? I would push it, I would push it, I would push it. I would go like, no, not it. You know, I need the Source, give me that. And it would try to give me what I need multiple times. And the more it, you know, goes into the same sort of circle, it would just drift to like absolutely like just nightmares, like not even relevant anymore, right? It just drips away from the main point. And I feel like that's— it's just trying to serve you. So it's just trying to expand its sort of like reach to Source more, and it just becomes very, um, you know not the reference score, it just decreases, if I might put it like that. Is that something you can protect yourself against when you're doing AI personas?

Yeah, I mean, fundamentally, it's designed to just not do that because there is no drift because everything is based on your own memory and your own context. So it's just simple as that. Yeah, that's the whole value proposition, which is The expert personas are fundamentally grounded within the context that you provide within the memory.

Okay. So they're not being trained on any pre-trained general big models like Llama or— Okay.

So the way the architecture, the way to think about it is you have— I don't know how technical you want to go into this in the episode. You go. Let me say it.

Yeah. Say it again. Yeah, just go as much as you think you need to give to convey the point.

Yeah, I will. I'll say I'll try to keep it simple, but in most use cases, you have the context and then you have the RAG components and then you have the LLM, right? Mm-hmm. Essentially, the context is shared between that RAG layer and the LLM layer. With Personal AI, the way to think about it is you have the end user, you have a small discrete model, which is PLM, that may have RAG and they have multiple different components within the PLM, and then you have the LLM. When there is an inference request, majority of the job is offloaded for token generation to the PLM, right? And then LLM supports the foundational language, the general task, the transformations, the summarizations, the formatting— like all those things LLMs are really good at. So PLMs fill the purpose of highly discrete, highly specialized expert inference, and then LLM supports the rest.

It's a simple answer to that. There's no need to even think about that. Okay. So let's talk about the parallel workflows. So I looked at understanding what's the difference between, you know, like you have one agent which does pretty much everything. And as well, you have multiple agents which are responsible sort of for the same outcome. And, you know, the speed win is absolutely there, but it raises many questions about overlaps, multiple agents doing the same and not talking with each other and sort of a lot of these challenges. So from technical perspective, how do you actually make sure that multi-agent setup, which is responsible for the same outcome, is actually talking with each other well and don't overlap information?

Yeah, good question. Currently, there are a good number of standards that are, you know, being built. And we are working with, for example, like protocols on how do we make the interpersonal communications much more standard, very similar to persona to an external system communication. Because at the end of the day, If the communication protocols are standard, it's going to be better. But from a simply agent orchestration, the one thing that we always encourage our customers is to think through the lens of their organization. Rather than specifically inventing or designing newer roles or newer agents that could be like many, and you have no proper ownership for a particular agent. So that means the data stewardship for the memory of a persona is not clearly defined. And that problem from our perspective is no different than an employee who doesn't have a manager or is not provided proper guidance or doesn't have any guardrails or rules around it, right? So one thing that we push with our customers is to keep these personas alongside existing team members and then have an accountable manager for that particular persona, very similar to a team. So now a director who is running a corporate strategy team has 4 people and then 6 personas to manage. They are the ones who are approving and making the decisions around what goes in it and what doesn't go in it. Then the technical answer about agent orchestration, which is, hey, I have this problem, team, and let's think about it as a channel with 4 people and 5 personas. And you have a question or you are trying to make a decision, you normally come into the channel and you have a conversation about it. And usually the team members, people, humans will jump in. And now, very similar to that, the AIs will jump in as well, right? So if you start thinking about, rather than inventing like new sets of workflows, new sets of, um, uh, what you call processes, but rather think about placing these personas alongside, next to your employees where the processes and workflows already exist, that is the least path to resistance to create, you know, instead of creating like a confusion of everything is everywhere and there is no access controls around it and no accountability, more importantly.

So if you put multiple AI personas in the one channel, will they talk with each other and find an answer?

Yeah, yeah, it's a, uh, there are some prompting and techniques involved on how they would work with each other. You can, um, currently in the platform you can pose a question and then you can configure the AI personas to raise their hand and then come up with their best answer if they have one. That also is defined by, you know, the personal scores that each of the AIs kind of self-determine if they have a good answer or not. And only the ones that, you know, we think is going to answer, then kind of shows up, right? So yes, there is like configurability involved to create that kind of a panel within a channel.

Yeah. Okay, so, so effectively they, they just need to be aware of each other and have a conversation with each other to, to make sure they don't duplicate the work. And when I say duplicate the work, I mean, you know, from my perspective, it's like If I launch a question to two personas and they will be doing their own calculation and coming up with an answer. So if I keep them in the same room, they effectively will not be doing this. They will just self-elect whoever will run with the task. Is that the right assumption?

Right. And the guardrails can be set by a human manager for the particular channel. Okay. Yeah. Let's just say, let's take a real example, right? Let's take, you know, you have a Slack channel with 200 people, right? And a CEO of the company, or like, you know, some exec comes in and poses like generic question to all the 200 people, right? Then you'll have like tons of noise on, you know, what every person, you know, would think. Yeah. Or if they have, you know, potential contributions to make or not. And sometimes you may have noise, right? But however, depending on that particular context, you have a channel that is very highly focused on to that particular topic or a decision, and you have 6 humans in there, and then you pose the same question to that 6 humans because you're already kind of narrowing it down to where, you know, to those specific people, then you will have, you know, highly curated sets of results, more thoughtful responses rather than creating like lots of noise with hundreds of people. And it's no different than how these AIs will behave in these channels.

Okay. That's excellent. I mean, it gets my mind racing. So you mentioned the managers, which effectively own the AI personas, they responsible of seeing what goes in, what goes out, and then configuration. So are we saying that there's now a requirement for managers who manage teams to actually be able to manage AI personas, and that's a new skillset? Oh, 100%.

We actually have. How persona manager roles inside the company. Okay. How—

oh yeah, it's a full-time job. Like, how big is this? Something like a spin-off, but if you have actual positions, that sounds like a full-time job.

Uh, yes, in some scenarios, uh, it could be combination of the job that you would do otherwise and then at the same time be accountable, uh, or own a particular persona. Those things are evolving. And in fact, one of the customers recently did end up hiring a full-time persona manager, or I believe converted the existing role inside the company to a full-time, like, a persona manager. So that becomes like subject matter expert, if you will, for taking all the requests from across the department side. Oh, um, price competitive. Oh, well, you know, yeah, this basically like offer there is these existing personas. We may want to build this, you know, this particular problem in a particular way, right? So they not only become the persona managers, but they also become, you know, architects, if you will, of a particular use case.

I love that. I mean, that's so easy to see becoming a full-time job, right? Because you just become like a manager, but not of people, but of AI personas. That's so cool. Right. So let's talk a little bit about governance of AI and how you scale that in enterprises, right? So I personally work for regulated markets, hundreds and thousands of employees. So if I, let's say, want to bring in AI personas alongside to work alongside the employees and we're saying we're going to roll out 500 AI personas. We regulate that market, we have audits, you know, to regulating bodies, we need to provide documents, there's a lot of securities, a lot of structures of who sees what and handovers and things like that. How applicable AI personas are for markets like that? Are they being used? And if so, what are the rule sets, guardrails or approaches to get them safely used?

Yeah, I mean, in fact, that's kind of our market. That's the market that we serve in, which is highly regulated industries that are, you know, subject to sometimes not just processes internally but also externally, like such as federal laws as well. And it goes back into kind of the architecture of the system that we were talking about earlier. Which is the rules and the practices and the processes that you have at the company for the controls that you have, you would still keep it the same way. Instead of bringing the governance to the AI, you bring AI to the places where the governance already exists. It's a reverse way of thinking about it. Don't change the governance, don't apply the governance to the AI, but get the AI to sit inside the existing frameworks of principles and controls that you already have. That concept is no different than, you know, which organization owns what kind of personas and who is responsible for what and under what circumstances. If there is a human in the loop involved for an approval of a decision that AI makes versus not, right? We serve banks, for example. Within the bank setting, you know, you have a few personas that are making decisions around like loan processing and underwriting procedures, as well as the, you know, validity of the applications that are coming in, right? There are few expert personas in there, but there are people who basically govern those personas and oversee the decisions that they are making, you know, based on certain rules. If the AI is highly confident with our personal score, which fundamentally means that it is heavily predicted, you know, the response, just like a human would, because it's fundamentally based on their own internal data, then you may not need an AI— sorry, human approving it, right? So I think those things will evolve. But the interesting thing is putting the frameworks and tools in place for the enterprise companies or customers, whoever needs to ensure governance matures over a period of time.

So let me read this from, from Europe Union AI Act. High-risk AI systems shall technically allow for the automatic recording of events over the lifetime of a system, and technical documentation shall be drawn up before placing on the market and kept up to date. So that's not a mandate, right? It's just the AI Act. And some other similar documents sort of exist, which implies that AI agents are being treated as a piece of code, right? So which means if we're making a change to configuration, all of these changes are getting tracked. It's the same as, you know, changing, you know, your code. Everything AI is returning when we're having conversation with that, it's getting logged. So we have potentially a trace back to see how it made decisions and things like that. So effectively then it means that AI personas in regulated markets, you're treating that as you would be treating a piece of code. Same release processes, same change processes. Is that correct assumption? It is current assumption.

I would go a step deeper, which is what's most important for the auditable— auditability of an AI persona. Well, one is the transactions that the AI is taking. Are they visible to the right group of people that need to, you know, oversee or govern? Those, those responses. And the second thing is, whenever there is a change to the memory, whenever there is a change to, let's say, directives, are those pieces of information auditable and visible? So that way you have the, you know, proper association to the changes that you are making, very similar to our change management, you know, within a highly regulated setting. Yeah, right. Um, and, you know, fundamentally, the, the idea of the programmability and the privacy comes down to this core idea of how can you give control and put that control in the hands of the customer for people to get the benefit of the AI without, uh, disrupting heavily their existing, uh, promises and existing procedures responsibilities that they already have as a business, right? Yeah. So it goes back into the tooling and pretty much everything, you know, within the Personal AI ecosystem is traceable and every transaction AI is making is visible to the right group of people according to the configurations that you do.

And currently there's no certification body which would be looking at your systems and saying like, yes, you have that.

Right, like, uh, so from the certification standpoint, we are SOC 2 compliant, GDPR compliant, HIPAA compliant.

So is there, is there equivalent to SOC 2 for AI? There's not, isn't it?

SOC 2 for AI equivalent, uh, depends on how you want to define it. Uh, rather, the SOC 2 certification body would probably enforce or push like additional controls for SOC to AI. Currently we are fulfilling the SOC to general compliance rules and controls as part of our overall architecture.

Yeah, I'm just interested if there will be at some point guidelines you have to adhere to, like GDPR is a good example, but for AI, which effectively says if you have that, that means I'm capturing the logs, I'm doing everything I can be audited and I will pass whatever the rule set is, you know, for a company who provides, you know, such services, you know, for AI. So, effectively, you know, the question is like, do you know if such certification exists? Because I couldn't find anything.

I do not know if such certification exists, but I do know, for example, there is one organization that I'm part of called GLIAnet, G-L-I-A-N-E-T. They focus a lot on the AI compliance-related topics, specifically proposing standards and upgrades to the existing certifications, if you will. So that's like one of the change agencies, if you will, that is in the market that I am aware of, you know, pushing towards those kinds of standards.

It's easy to see that appearing at some point. Oh yeah. Yeah, it's just, it's a little bit of while, while, while.

It's a maturation of the market. Yeah.

Right now we have this disease which is called AI sprawl. I don't know if you ever heard that term. I did hear that. You did, right? Because it's just like everybody wants to use something and then bigger companies, I mean, you don't need to go very big. You go medium or above enterprises, and you'll have hundreds of people. And out of these hundreds of people, there will be a big majority who want to use AI to get efficiency or effectiveness within their roles, but the business don't have processes to onboard AI tools. So they say no, but they still use it, you know, in another tab or another window. It's usually like ChatGPT or Gemini stuff or other tools or other tools or other tools. And there's AI sprawl within the organization. It's just chaos.

Yeah. One of the challenges— it's a good point, right? One of the challenges that we hear from our customers is they currently, with the number of AI tools out there in the market, there is a fatigue on the compliance and the regulations of which data is going into what AI tool. So the whole advantage of plugging in Personal AI in here is this core idea that it all is in your control, it all belongs in your own platform, and you can solve specific AI personas for strategy needs and marketing needs or HR needs or specific supply chain needs or employee-specific personas, like all in one controlled platform which basically belongs to you. So in a way, the IT or the compliance and the security people have control around how this AI is plugged into their existing internal mechanics without going crazy and trying to figure out data privacy policies for multiple different vendors.

I love that. Excellent. Loving every bit of this, Suman. Let's go into rapid questions round. Quick questions, 30-60 seconds for an answer. How do you handle a poorly formatted or inconsistent tribal knowledge when feeding data into AI workflows?

You don't. You don't.

You just allow it to go in and then you see how bad your answers are. How do you safely integrate AI personas into no-code tools like N8N, Make.com, or Zapier? Effectively, how can I take AI persona from Personal.AI and utilize somewhere else? Is that possible?

Yes, Innate and Zapier are interesting. I think the tooling and the platform has to come closer to each other to continue fulfilling the privacy desires and the fundamental design architecture of AI personas.

Okay, but it is possible, let's say, Open like a pipeline through Zapier to emails or something like that. Okay. What's one quick test you can run today to catch AI hallucinations before they reach users?

Look at the accuracy scores.

Okay. How do you ensure your AI personas don't reinforce biases? Reinforce biases.

Also, accuracy scores, personal scores we call them.

So how is that score calculated behind the back? Is that like a black box? You just get it like a lot of what LLM information, or is that, uh, do you know?

So there's a TinyRanker model, okay, that basically, uh, is a specific model that looks at, um, the memory that the system has and the context that the user has and the historical uploads of the information to calculate the score. Is it a black box? Yeah, but how is it getting calculated is fundamentally based on the memory that goes into the system.

Okay. In regulated industries, what's one governance guardrail you wouldn't launch without?

Today? Based on Personal AI mission and vision, I want to say GDPR. That is the closest that I can get to, as our motivation is putting the control in the hands of individuals and business customers and empower them. So I would say GDPR. Okay. At least this time. Yeah.

What's your go-to metric to measure whether AI workflows are truly saving time or just shifting effort?

I would look at how fast your organization needs additional headcount to be added in the team. Hmm.

Okay, that's nice. How do you design your workflow so AI agents and humans' role and human roles don't overlap or step on each other twice?

Humans are known for their brilliance. So if AI is stepping into humans' role, that means it's a good thing. So that way, humans can adapt to newer and better problems.

So effectively, you get an agent done. If you do it well, you use it and you go like, "I could do something else now." We always build on top of our tech that is given to us. Okay, what's your strategy for updating AI personas when team knowledge or processes change frequently?

Uh, bring AI to your workflows, not the other way around.

Bring AI to your workflows and not the other way around. Okay, so effectively, if we're using AI in our workflows, it will learn from the usage and it will just update itself.

Right. That should be the mindset.

Okay. How do you prevent AI sprawl in your organization, keeping agent workflows coherent across teams?

Go with the persona architecture, not build-a-rag architecture. It's extremely complex to maintain.

Yep. I do agree with that 100%. If you had to explain AI automation to a skeptical executive in one sentence, What would you say?

They do a human job 10x better and 10x faster.

10x better and 10x faster. That's a big promise. Where the 10x comes from, is that just like an arbitrary number or is it actually 10 times better?

Turns into the core idea of like, you know, how many hours you are saving versus how many decisions you are making. Yeah. How many people's jobs you are doing. 10x is an arbitrary number. It's a north star metric for, you know, in the future, the CEOs of public companies are going to go to the earnings call and say, hey, 40% of my workforce is man, 40% is woman. 10% is people of different abilities and 10% is AI. So, the idea of the 10x is it's almost like a base metric or benchmark that any tooling has to achieve to be able to significantly create economies of scale for a particular business.

I love that. I remember listening to the NVIDIA CEO talk about exactly that. And as well saying about the the job roles which we mentioned before, where people just oversee and build AI personas. It's a very interesting future we're working towards. All right, well, that was excellent, Shimon. So to our listeners, what you just had today was information or know-how capture of your team. How can you do that automatically through APIs? Maybe some manual work initially when you set it up. Once you get your personas, hopefully you understood that the role-specific personas are so, so much more better than the generic ones. So if you have an ability to go and experience that versus just using a general ChatGPT, definitely do that. There's some very clear systematic efficiencies when you use, uh, uh, parallel workflows that might not be applicable to everyone, but that's something to keep in mind that you can bring AI personas in the room and allow them to work out your problems. And the governance guardrails is a very real challenge for big enterprises. So thanks for this, Shimon. If people want to learn more about Personal.AI or in general, maybe reach out to you, where they go, what they do?

I mean, you said it, Personal.AI.

I said it a couple of times. Is it just imprinted in people's minds? Simple as that. Personal.ai. We'll have your LinkedIn in the show notes as well. And any other things like your personal newsletter or things like that? Do you talk with the market on what you guys do at Personal AI and with AI in general?

My most active is LinkedIn. So that's kind of where I offer, you know, any insights. Okay. Any information that I would like to push it out to the market. Excellent. Here you go, guys.

Thanks everyone. If you have a friend in your network who's tinkering with AI, not sure what to do, send him this episode. Hopefully that puts him on the right path. And as always, guys, until the next time, stay safe.

See you online.