Why GenAI Fails Without Clean Data

Get featured on the show by leaving us a Voice Mail: https://bit.ly/MIPVM 

🎙️ FULL SHOW NOTES
https://www.microsoftinnovationpodcast.com/705   

What if the biggest barrier to AI success isn’t the model—but your data? In this episode, Sedarius Tekara Perrotta, CEO of Shelf, reveals why unstructured data is sabotaging enterprise AI efforts and how to fix it. From SharePoint chaos to hallucinating copilots, Sedarius shares a practical framework for transforming messy data into clean, actionable fuel for generative AI. If you're a business or tech leader navigating AI adoption, this conversation is your roadmap to clarity, control, and competitive edge.
 
🔑 KEY TAKEAWAYS
Unstructured data is the silent killer of AI accuracy. Most organizations have millions of files riddled with micro-errors that confuse AI models and lead to hallucinations. 
A successful AI strategy starts with a data strategy. Clean, enriched, and refined data is essential before plugging into any AI system. 
Use the “Reduce, Enhance, Refine” framework. Focus on a specific use case, eliminate irrelevant data, enrich with metadata, and refine to a trusted set of documents. 
Automation is key—but not enough. Human oversight and iterative monitoring are still critical to ensure AI outputs remain accurate and relevant. 
AI agents will transform job roles incrementally. Expect gradual automation of tasks, with agents augmenting—not replacing—human expertise.

🧰 RESOURCES MENTIONED
👉 Shelf – AI-powered unstructured data management platform: https://www.shelf.io  
👉 Microsoft Copilot Studio – Customizable AI interface for enterprise use: https://learn.microsoft.com/en-us/power-platform/copilot-studio  
👉 Gartner Research – Identified poor data quality as the #1 barrier to scaling GenAI - Read the full Gartner press release

This year we're adding a new show to our line up - The AI Advantage. We'll discuss the skills you need to thrive in an AI-enabled world.

Accelerate your Microsoft career with the 90 Day Mentoring Challenge 

We’ve helped 1,300+ people across 70+ countries establish successful careers in the Microsoft Power Platform and Dynamics 365 ecosystem.

Benefit from expert guidance, a supportive community, and a clear career roadmap. A lot can change in 90 days, get started today!

Chapters

• 0:00 Welcome & Global Adventures
• 4:43 From Peace Corps to Knowledge Management
• 10:28 The AI Data Dilemma
• 16:53 Building a Real Data Strategy
• 24:33 Automating Trustworthy AI Answers
• 33:13 The Future of AI Agents in 2025

Transcript

Mark Smith
Welcome to the power platform show. Thanks for joining me today. I hope today's guests inspires and educates you on the possibilities of the Microsoft Power platform now. Let's get on with the Show. Today's guest is from New York and the United States. He is a CEO of shelf. He's a three times founder and current chairman of the Board of Artificial Intelligence and Knowledge Management Company Shelf, which serves as an unstructured data management platform for both LM's and people, is previously served as a USP. Corp volunteer and mentor at the Founders Institute. You can find links to his bio and socials in the show notes for this episode. Welcome to the show Sed.

Sedarius Tekara Perrotta
Thanks, mark. Really happy to be here. Thanks. Having me.

Mark Smith 
Good to have you on the show. You got it. And I find that the tech that you're in is incredibly interesting. I'm just on the Nexus of starting a new podcast show, which is actually totally just focused on AI and the the the massive growth we've seen in the last 24 months in this space. But before we get into that discussion, tell me a bit about food, family. And fun. What did it? Mean to you? 

Sedarius Tekara Perrotta 
Well, outside of work, there's really just family. And the more family travel and adventure could line up to be 11 great trip the the happier I am outside of work. So I've traveled all over the world, almost 100 countries clocked. I've done many different types of adventures, love out anything outdoors, mountaineering, scuba diving. Skydiving. Whatever. Boating. Anything you can think. Of I'll do it. I'll try. It I'll love I. Just love adventure so that that's kind of a bit and I'm trying to get my daughter as and and son. As much into it as possible, so that. We could just continue the good times outside. To work. 

Mark Smith 
Amazing. Amazing. I've I'm coming up on 50 countries. So you're double me? That's that's that's incredibly impressive. How do you select a new country?

Sedarius Tekara Perrotta 
Well, I most of my countries were clocked through backpacking, so I would just work as a consultant, which you'll you'll get into a bit. So I work as a consultant and then I would take six months off, 9 months off and I would backpack through a region, South America, Central America, Africa, Southeast Asia, Europe, Australia.

Mark Smith 
Wow.

Sedarius Tekara Perrotta 
Yeah. And I would just spend time, just literally bus from station to station checking out, exploring different regions and different museums and different national parks and all sorts of stuff. So now I kind of selected on family friendliness. That was not the case. I usually choose the most dangerous. Because I could possibly find when I was younger. Now it's the safest, most family friendly place and we're going to be heading to the Yucatan of Mexico in in about a month. So yeah, there's a lot of fun things for the for the kiddos.

Mark Smith 
Where? What? What story comes to mind when you think of someone? Says. Tell me something about your travel experience that really stands out, or highlight or or location or something. Just the first thing that pops to your mind.

Sedarius Tekara Perrotta 
Oh my goodness. I could I there wouldn't be 1, so it's it's some people ask me what's your favorite country? And it's. I always answer for. For what? What what type? Of experience are you looking? Is it for like outdoors? Is it for wildlife? Is it for culture? Is it for food? So I'll answer your question from the perspective of food because that was the answer that I didn't give you. And just as a, you know, transparency sake, I am Italian.

Mark Smith
Ah.

Sedarius Tekara Perrotta 
By nature, from from my family and whatnot. So Italy, the food in Italy, I I've been to Italy, I don't know a dozen plus times. My families from Napoli so. It's very hard to get a bad meal outside of the centre central areas in Venice and and and Florence, Rome. But if you go out on the periphery, Oh my goodness. It's so hard to get a bad meal. So yeah, I mean, I guess from the perspective of food go to Italy, go to the malfi coast in summer when the fish is fresh and you can get the food. Mari's going to the going to the Piedmont region and get the barolos and the Brunellos and all that, and you'll be guaranteed to have a really good trip. Yep.

Mark Smith 
So good, so good. It's my wife and I's favorite country as well. I went there last year. I'm going again this year.

Sedarius Tekara Perrotta 
Where do you go? Where you go in this this coming year? 

Mark Smith 
Ohh. Specifically I'm going into Venice because I am then going over to Slovenia to speak at a conference and so yeah, so hence going via Venice.

Sedarius Tekara Perrotta 
That's smart, you know, flying to Verona.

Mark Smith 
Yeah. Yeah. So I actually. No, we're flying directly into Venice.

Sedarius Tekara Perrotta 
OK. That little air? That air? Yeah. Yeah, yeah, yeah, yeah. And and Verona, but yeah.

Mark Smith 
Yeah, yeah, yeah. Yeah. Look, because we're coming from about in the world, we've got one stop over it in Qatar. I think it is in Qatar has a direct flight to. Venice. So it's awesome.

Sedarius Tekara Perrotta 
Have you been to Verona by? Any chance? Yes, I have. Yes, it's awesome city. Most people don't know about it. Incredible city it. It does not get it. It's two credit. Really incredible food as well. In the home of Romeo and Juliet.

Mark Smith 
Yeah. And the Colosseum is, you know, so, you know, still fully intact there, right. And of and of course that's.

Sedarius Tekara Perrotta 
One I've seen anywhere, actually even better than one in Rome, I think because the the one in Rome is, you know, it's been kind of taken apart by different power fractions over the last 2000 years and it's not what it used to be, but the one in Verona kind of looks the way it looked.

Mark Smith 
Yeah, yeah, yeah. Exactly. Exactly. Now, beautiful city is probably, in fact, I think it was a. First town we went into in Italy when we were travelling in 2017, so yeah.

Sedarius Tekara Perrotta 
That's a good intro because it's also small and you could. Walk the whole thing.

Mark Smith 
Totally. Totally. And I mean, you know, it has its its links to so many different things with Shakespeare. And one of the other ones, one that there's a the famous statue of. Of the guy that talks about the seven levels of hell or something, some famous philosopher? Yeah. Dante. Yeah.

Sedarius Tekara Perrotta 
Ohh Dante. Dante.

Mark Smith 
Yeah. Yeah, yeah, yeah. There's a square. There's a Dante square there.

Sedarius Tekara Perrotta 
Everywhere you go in Italy, it's just some type of history, just like, Oh my goodness.

Mark Smith 
Yeah. Amazing so true, right. The history is just phenomenal. Tell me.

Sedarius Tekara Perrotta 
I feel so dumb.

Mark Smith 
A bit about your. Career in tech how to how to come about? How how did you get to where you are and you know you've you've got a few companies under your belt to give us a the snapshot of that.

Sedarius Tekara Perrotta 
Well, in your intro, you kind of teed this up for me. It all started when I. Was actually in the Peace Corps. I kind of worked as a knowledge manager because I was my service was in a think tank, so I was working on the think think we were doing research on ICT trends for the government and I was given the tasks that nobody else wanted, which was essentially organizing all the research very you know. Difficult spot to be for any knowledge manager out there. I'm sure you could appreciate it how how unfun it is organizing other people's mess all the time. So anyway did that for a period of time and got pretty good at. And then when I and this also ties to your first question and then when I got out of the car, I started doing knowledge management consulting and I I did consulting. Gigs for MIT and. Harvard and Stanford University. It's societal organizational learning and Boston College and World Bank, and just did a lot of really incredible projects. We're just trying to. Organize trying to get our heads around a particular domain. So particular problem in the organization that they wanted to have greater transparency, greater understanding of the key decisions made, best practices, lessons learned, key strategies. And extracting that out and and providing a way for people to access it in real time and preventing the their most important and valuable learning to to essentially disappear into the vapor of time. So I got I built my my domain expertise in that and. Then Tech was forced. Upon me, because when you start doing this in the age when everything's digitalizing and everything's going to the cloud. Where you're kind of forced. To adapt and that's when I I got into the the kind of the software side of things that I I taught myself to code and learned how to build platforms and learn how to customize SharePoint and that's was our that was the beginning of my really long standing relationship with Microsoft where learn to customize. The SharePoint instance for different knowledge management initiatives in in large organizations and. That was kind of. The second company I started, which was a services company around knowledge management. And for like Fortune 500 companies, enabling them to have a a knowledge strategy on SharePoint. So SharePoint is just a technology, you still need people and processes and that's what I help provide. So that's how I got into tech. So I hopefully I answered your question there.

Mark Smith 
So obviously you know, as things have progressed, you've got into the space of. Artificial intelligence and its effect on knowledge management, which I've I've found almost stunning over the last two years, is. At, you know, 7-8 years ago I would be presenting at events talking about the four mega trends that we're going to change the world from a tech perspective. And one of them was data and and our production of data would increase to such a phenomenal rate. And I feel that for the last 5-10 years, companies have got really good. At collecting data. And no, no, no idea around how to orchestrate that extract value from it, derisk it and and then along comes generative AI. We run that over it and all of a. And patterns in that data start to appear BI information that has been on the networks for years is now all exposed and companies are I heard an interesting story recently of a company that is a Fortune 500 company considering AI and their concern from IT. Was that they said we know how bad it's bad. And what a is going to do is expose how bad our data is and. So you know. I I consider that security by obscurity, right? Nobody knows how to find it. It's either secure, but because nobody knows how to find it, we take some comfort in that which is, which is crazy. Really. Right, so, you know, especially from a. A cyber crime perspective and this concept of a walled garden and that as long as we people can't get in, we. Safe, and of course it's not the case. So what are? You finding in this whole area of you know SharePoint is often the repository of unstructured data inside an organization. What are you finding that challenges that companies are now coming across as they go, you know what we want to step into this area of AI, we know we're going to get a. A lot. There's a lot of fear of missing out right of of our competitors are gonna get the jump on us and and whatever technology area they're oh, sorry, business area. Karen, how do you have those conversations with these with with these type of organizations to bring about clarity and then move forward with their their data story and then ultimately their AI story off the back of it?

Sedarius Tekara Perrotta 
It's interesting because there is a element of education, mark, you seem to be very educated on the problem. A lot of organizations are not essentially what we find is that it's a crawl, walk, run approach to to to any type of transformation, organizational transformation. If you try to skip steps you you usually. End up in a bit of trouble. So what we saw in 2024 was it was a year of POC, POV's, proof of concept, proof of value. Organizations were just trying to get something stood up trying to play around with the technology, try to poke at it, find out where its weaknesses were. And one of the main things that came out of that period. That, that, that experimental period was. I think an awareness, at least from Gartner's perspective awareness that the the data is a real issue, Gartner identified it as the number one issue preventing Jenny I at scale. So first it's it's an acknowledging that that there is a data issue and I'd like to just quickly take a step back on the data issue because it's easy to get. Caught up in these big words. Or short words, if it's data, but it usually doesn't mean anything, so I'd like to just ground it on what it actually means from the standpoint of knowledge from unstructured data and. Other word that. Usually doesn't mean much to people and trying to ground it on something that's practical and interactable, and that is what is unstructured data. So we as humans, every day we are. And in our professional lives, we are producing materials. We are turning thoughts into things we are, we are, we are typing out reports and and creating material documents and we're creating presentations and forecasts and reports and that that is the output of our day, day in, day out, week in, week out, so forth and so on. And. Because we are human and we do, we do always make mistakes. It's just part of who we are, as we're we're constantly learning and constantly growing, constantly getting better. But through that process we're making mistakes and those mistakes are in those documents. There's multiple versions of a document that shouldn't have them. They're they're they're saved in the wrong folder. A. A section of a document wasn't updated. A presentation has an old slide. A price wasn't updated. Whatever. Anywhere. Any department, any use case people have are just by their very nature making these very small, tiny mistakes and those tiny mistakes are getting saved in one drive and SharePoint, Google Drive and everywhere they're they're getting saved wherever people work and now. What the issue is? So I think we can all agree that at the current state, AI is not sentient. It is not artificial general intelligence. It isn't able to govern countries and and and and and run businesses and so forth. It's not there. So since it can't think like a human, how can it possibly identify? Human mistakes. So with these mistakes in the data the the data is the fuel of AI. So it's like asking the car to analyze the gas. It's a different technology. It has a it has a different source, it runs through different processes and you can't combine them and just assume it's the same. They're not. They're completely distinct things. One is fuel and the other is the engine. And journey I is the engine and what it needs is clean fuel and that fuel that it's being given from the. The the perspective. Of just shared is that unstructured data which? Is all those the. Summary of all the documents that an organization has created and having all those micro errors in that. In those documents and then asking the machine to just know which ones are right, which is the right version, what is the right section that had something wasn't updated? What is the correct price? That is something that a human needs to fix, and if you don't have the human fix it, then you have this issue that we've seen in the news about hallucinations and about issues of accuracy and so forth. But a lot of it is sourced around the data. So my advice is answering your question, it is, you know, if you want the engine. To run efficiently and you want it to go far, you need to. You definitely need to have a data strategy, not just an AI strategy. They're two different strategies and they need to be run by different people, but very well coordinated and organized. And we're seeing that organizations that are succeeding and getting their, their getting projects off the ground are doing exactly that.

Mark Smith 
So. So let's talk about that data strategy. Do you see any patterns of good data strategy surface?

Sedarius Tekara Perrotta 
Oh for sure. Thankfully there they they have emerged. So first of all, the biggest mistake I'll start with the mistake. So best practice as an as a background and knowledge management, best practices, lessons learned, key strategies. Let's start with lessons learned. Trying to do everything all at once is a really bad. Media so there was a survey by virtual business intelligence that Jenny a 2025 outlook. They had essentially said that the vast majority of data organizations, they're the vast majority of data is is unstructured and that the average organization has 10 million files, 10. Nearly an average, so I imagine that's the average size company that's a 500. Thousand person organization. Imagine what a 10,100 thousand person organization has 10 million files. So trying to boil the ocean, so to speak, never works. So you need to focus on. So the first thing that you need to do is is is not do that not boil the ocean. Focus on a specific use case and around that specific use case. They will be specific. There are specific documents and files, unstructured data that will be relevant and. And and that's your first task. It's it elimination task. So you want to reduce the total number of data that you're looking at and identify what is the right pools of data that you want to start fixing and proving. Managing. So. So that would be the first step. Then the next step is you need to enrich it. So it's an enhancement layer. So now that you've reduced 1,000,000 files down to, let's say 5000. Miles, the next thing you would want to do is you wanna you wanna add context around it. You wanna add metadata, you wanna enable all that unstructured data, all those documents to be compared apples to apples. Because if they don't, if they aren't all filled out with the similar types of metadata around an ontological structure, you're going to be comparing apples to oranges and then you can't. You'll find it, which is the most important stage, which is the last one. So you want to go through a process of enrichment, you want to enrich it as much as possible, and there's tools for that. And then once you've you, once you've got your 5000 files, you've reduced it to 5000 from from 10 million now you've enriched it. Now it's the refinement stage and now you can take that 5000 probably bring it down to like 1500 files that really matter that are accurate up to date and trusted and then connect it to your Gen. AI. And then that final stage of it is to obviously monitor and continuously improve, cause you're not gonna get it right out of the gate. You're gonna get it right by iterating by, constantly focusing on it. And if you do that, you will succeed. Eventually you will succeed and you will have a successful use case that you could then start thinking about scaling.

Mark Smith 
So in that scenario, do you need human curators that are, let's say, you've gone through those 3 stages that then maintain that like you know, let's give you an example that I've had in the last year or so, a a bank, they have a sample where somebody comes in to the bank branch and they and the person says. What's the latest term deposit rate and when they search for that, let's say on their SharePoint system, there's 400 documents of the latest term rate because each you know each time it's updated your documents produce. And you talked about that metadata, the context of, yeah. But what's today's date? What's the most recent one? Blah blah blah. So that bank spends 3,000,000 a year on an internal contact centre so that person can pick up a phone and say, ah, what's the actual the term rate for today? And of course, they're wanting AI to solve that. Pull them and have a person at that teller machine to be able to. Get the full. Metadata context of location. You know I'm in a bank branch. I got a customer in front of me. It's today's date, and I'm talking about the five year term deposit. Does does it mean that someone a human needs? If you've distilled down to that 500 odd documents as as as you said needs to then go. Yeah, but that one's old, that one. 'S new or. Or are you looking at tech to do that?

Sedarius Tekara Perrotta 
Yeah, I mean, this is exactly what we do. So you couldn't do it manually. It's very painful. You're not going to get it right because there's the volumes is too great it it's not something that a human brain can process. Looking at 400 documents and going through and figuring out exactly what area of it is right and and not right, it's not. Efficient. So that goes into that enhanced and refined step. So again it's it's reduced. So you so you get a perfect example. It's been already reduced to 400. So this particular use case has been reduced to 400, it's 400 documents. Now you're enhancing it. So exactly what you said automatically adding the metadata, you do not want humans adding metadata. To that many files you you will waste a lot of time and money. You could do that automatically also using a. So it automatically generates and standardizes the data. The unstructured data on a metadata layer, and then it's a refinement algorithm. So now you have a whole bunch of algorithms looking at different heuristics of measurement and comparison, conflicting sections, similar sections. Inaccurate out of date. Other things that you could have heuristics around are contacts. The topics that that is relevant, the location that is relevant, and you could essentially have a dashboard on top of those 400 documents and just saying here are the five that are most likely correct for this specific use case. Do you want to filter out the other 395? Yes, boom. OK, now look at 5 instead of 395 and identify the exact right one for this specific use case. Attribute a specific topic, a specific geography, exactly what you were saying. Timestamp what what? Not when you want AI to pull this particular document. Cause usually when you go from 400 to 5, usually those 5 or 10 or 15 are all correct. They're all correct under different use cases in different scenarios for different locations and you don't want. To filter them out what you wanna do is you wanna pervade again. You wanna provide that metadata layer that is providing context for the Gen. AI in order to properly retrieve it. So when the the retrieval augmented generation technology gets kicked in, you have this layer, this intermediary layer that's sitting between. Your structure, your unstructured data and the LLM and that's the the key thing that we provide is we provide an automated view of your unstructured data that enables you to seamlessly manage that that very, very manual difficult process. If if we didn't exist.

Mark Smith 
So just to jump into a bit of the tech detail there, are we talking about just SharePoint data sets and then how does that feed into the Microsoft graph or are you talking about dedicated? And You know copilots, if you're to use Microsoft term that you are building out and therefore has its own interface independent of, you know, M365, copilot. How does I take it this what you do with shelf allows for the to a large degree in automation of this area.

Sedarius Tekara Perrotta 
Yeah. So it's fun. I'm going to use these terms again. It's crawl, walk, run. So Microsoft has out-of-the-box copilot for 365. So, OK, great. You have an out-of-the-box. Copilot get started there. You're going to run into its limitations. Then they have the AI studio AI copilot studio. So then that gives you a little bit more. Customization a little bit more power to control your outputs and it gets you so far, but it also you'll also run into limitation. And then you get down to the the deep layer which is when you really want to scale the AI. And that's AI foundry and that's where you have full control and that's where we plug into all three of those, right. So we can help clean the data before it gets into the, the, the, the Microsoft 365 use cases, teams, whatever. But it you're not going to be able to provide that. Context layer that I shared. With you because we can provide the context, but there is no customization on the LM side on the receiving side of it to to interpret it. So we'll be able to eliminate that that going that process, reduce enhanced refine, you'll be able to reduce a whole bunch and get better samples of data or better data sets, but you're not going to be able to refine it and get it really dialed in until you start. I mean the the. The copilot studio will get better. But then you really have full customization and control at that AI foundry level where you can really receive you can you can customize how the data is received and how that context is interpreted and then control the outputs of your your copilots.

Mark Smith 
But, but you're only talking about the second part of that equation, right? Because you know what you started with is. That, and let's say copilot studio as an example. I can point it in a data set and go there. God-given it the data and I've given it those 10 million documents, right? And go, hey, I've given it the data and wonder why it hallucinates, as you say, all those errors and micro errors have been entered in over time. So does your tool only work on that second layer? Or do you also look closely at, let's back up a bit and say what's the data source you're pointing this your AI at. We gotta do a whole bunch of work there first before you can actually start using it in your AI model.

Sedarius Tekara Perrotta 
Are you talking about the structured data mark?

Mark Smith 
As in getting your data into a format that you're down till you like the 5th doc, the five documents rather than the 10 million, you know that you that you start with because what what I see is I see consultants, they go hey I'm I I know copilot studio. I've worked with its predecessor power virtual agents and. To do a. Rag process. All I need to do is point it at a data. That's it. And they don't do the critical thinking around. Hang on a second. Is this data set the the data set that we should, you know, am I pointing it to the ocean or I'm pointing it to a really refined set of data that's gonna give me amazingly crisp results and everyone's gonna go wow, this is magic not realizing that. Well, the magic was actually refining the data that we were looking at to start with. Are you on the data side? I'm suppose I'm asking as well as the the integration to the you know, the Microsoft ecosystem from an AI perspective.

Sedarius Tekara Perrotta 
So we're definitely on the data side and we are also an integration into the Microsoft ecosystem. And I guess one other thing that will be helpful here is we're also able to process the outputs of the copilot. So the answers that are coming out and we're able to also look at those, I think that's that's critical. That's a continuously improving and iterating. Understand what? So you've you've cleaned the data you're like. OK, great. We think that this is gonna that, that we have everything handled and then outcomes bad answers outcomes. You. You went from 74% accuracy maybe you went up to 82 and you're like wow, that's that's awesome. We're at 82 pretty happy with. Yeah, but you could keep going. You just, you just need to know what are the answers being produced and then being able to source which ones of those answers are actually due to bad data due to the wrong documents or having the wrong context around those documents so that the LM's aren't able to properly interpret them and produce the. Correct response. So we do the data full on. So we'll help. But but we we don't store anything like we're not storing the data, we're just a processing layer on the data and then we provide that interface to enable a constant, monitoring it and then again that that interface to connect it to either. Microsoft 365 Copilot Studio or AI Foundry and again it's like a pyramid at the bottom of the pier. It's the most powerful stuff for the organization to get into, but they probably want to start at the top or. Or maybe they've already played around with the 365 and they're ready to go copilot. But yeah, we integrate with. All three of those.

Mark Smith 
I Like it? We've got links, of course, in the show notes to your tooling here. I think it's critically important. Tell me about 2025. We're recording this in January 2025. We've seen the hype around AI agents, and so we're going to get to much more rather than just a question, answer type scenario, conversational AI, we're going into an action or a a generic AI. There's a big difference between the marketing and the reality of these things and and what is, you know, so much of what I'm seeing is a gender AI as automation, right. It's all traditional automation. It's a rule based if then type type scenario. I'm I'm not seeing the the AI taking its own initiative around memory, self healing, correcting itself, learning from patterns of behavior, which is what I see as a utopia of agenda. Gay I, you know, give it an instruction and it will work out. What? It knows what it doesn't know what it needs more information on. Etcetera. That's one area big prediction for 2025. I don't know if we'll achieve it in 2025 to the level that that the marketing is telling us that we will and this can be from Google. This could be from open AI. This could be from anthropic and Microsoft, any of them in in the space llama. From letter. But tell me. What do you what do you in in the knowledge that you've collected over over the time in the space? What do you think is going to be a defining thing of 2025 In an AI's perspective.

Sedarius Tekara Perrotta 
It it's very interesting because the exact same problems that we just went over with more of a rag based use case are going to be the same for an agentic use case. If the data is bad, if you're pulling, if if you haven't cleaned and you haven't quality controlled what is going in.

Mark Smith 
Mm-hmm.

Sedarius Tekara Perrotta 
Then you're going to have garbage coming out. So I think. If we went rolled back the clock. To January 2020. For I think a lot of people, pundits would have been saying in in January 24 that we probably would have been further along collectively by January 25 than we actually were. And it's it's a normal learning curve. Everything's a bell curve, right. So there's there's this, this, this experimentation, there's this.

Mark Smith 
Yeah, yeah.

Sedarius Tekara Perrotta 
Adaptation. It's just figuring out what mistakes you've made and then correcting it and then trying again and trying again until you get it right. I I think out again I'll take a quick step back and just have a a larger trend and then I'll and then I'll answer the specific trend on agents cause I think agents are critical. The larger trend is we're going through one of the great transformations in, in our civilizations history with with AI. This is not a an Internet bubble. This isn't going away. There may be several years where things are not progressing as quickly where promises are not being made because they were a bit too optimistic in the very beginning, but the the trend of automating more and more. Their work and and and using artificial intelligence to do so is here to stay, and it is truly. Societally transformational in in every regard. So if that's the case, agents is really the means for the the vast majority of the transformation cause agents at the end of the day are doing I I like to think of agents as job descriptions. So if you just look at any job. Description Inside of your organization. Maybe not the one that you would post online, but the real job description of everything a person is doing for a specific role and you've listed 100 things. So the agents maybe in 2025 will be able to do three of those 100 things it it can do the the simple things, crawl, walk, run again. But every year that goes by, it's going to be able to take a greater and greater percentage of that job description and be able to automate more and more. You can see that with knowledge management as a as a function, there's already we have an agentic framework. We have over 100 agents right now we have 500 agents. Agent algorithms. By the end of this year and thousands after that. I can already see how it's augmenting the knowledge manager. It's augmenting the human work and and reducing kind of the manual failure based work that's extremely databased, which you just can't see. What you can't see. And then obviously for for the knowledge management industry, that is it's, it's. It's an accelerator to their function now. In 2025, is it going to be noticeable? That's the question I don't know, and I don't think anyone really knows how noticeable it's going to be. I think generally the trend is, is, is it going to be 3%, two percent, 1% of a particular job function, which job functions, obviously some job functions like a. Like a call centre agent is a higher percentage of the tasks can be automated, so maybe that's like 15. Percent. Maybe you can have Agentic framework enabling all of the write ups of the tickets and the call summaries and the tagging and the organization and the next steps like maybe it does a lot of that and it's very impactful in that particular use case. And then there's other use cases like I said like. Marketing, product marketing, maybe it's like 3% for product marketing, maybe it's 1% for SDR or 2% for an SDR. We don't know, but the the overall trend is it will be a greater and greater percentage for more and more job roles every year that goes by where it is at at at a particular day and time. I don't think anyone knows. And if they say they. Know maybe you should. You know, weight weigh that perspective a bit, but the general trend of where? It's going is, is. Kind of defined.

Mark Smith 
Yeah, totally agreed. Side's been awesome having you on the show before I let you go. Is there anything you would like to close with?

Sedarius Tekara Perrotta 
I would just say that we've talked about the issues with data. We've talked about the issues of hallucinations. I've. Used that crawl walk. One methodology a bunch of times here I I would just say regardless of whether you're using our technology or not, shelf really focusing. On your use case on what you want to. Solve in 2020. Five and rank ordering those use cases and saying here's 1520 use cases that we want to solve, but this is the one that we can solve. This is the one where the the the data isn't so great. It's only 1000 files or whatnot and not 10,000, but identifying that use case and getting that right. 1. Beginning is the greatest probability of success and then using a technology like shelf to automate as much of that process of the clean up enhancement, elimination of the bad data and the monitoring really sets up any organization for a higher success, more automation, more momentum and. Competitiveness in 2025, so I would just say focus on the data you get the data right, you'll get the answers right, you get the answers right. You'll be able to automate more and more and and differentiate and be more successful in the market.

Mark Smith 
Hey, thanks for listening. I'm your host business application MVP, Mark Smith, otherwise known as the NZ 365 Guy. If there's a guest you would like to see on the show. Please message me. On LinkedIn, if you want to be a supporter of the show, please check. Out buy me a. Coffee.com/NZ 3/6. Five guy. Stay safe out there and shoot for the stars.