Webinar: From AI Vision to AI at Scale: What Leaders Need to Get Right

Amazing. Hello, everybody. Good morning. I'm glad you are here to this workshop on scaling AI solutions from prototypes and POCs to something at at production that actually brings the value. My name is Dominic Mattula. I am head of AI at Profinite and Amdocs' company. And there is my cohost, Tomas Rezek, our head of center of excellence. So we have prepared content for, like, minutes, and then there would be plenty of time to to put you to answer your questions. So if there is anything you would like to ask during the session, just hit the Q and A button. You shall have in the Teams, and we will do our best to answer your questions either here or after the session is over. And with that said, let's jump in. I'm so I'm also sorry. I got some sore throat, so hopefully it won't be trouble for you. But fortunately, this is an online session, so don't worry. You won't catch it anyway. And with that, I'm jumping to the first slide. And I think we shall start with something bold, with some bold statements, something like we are living in the age of agentic AI. But, actually, we are living this these days. So this is not bold statement anymore. We can see that across our our customers that almost everybody is either doing AI or planning to start with AI this this very year. And this is not only my observation. This is also something major major consultancies report. For example, McKinsey, I can see over here, in the recent report, they stated that, like, almost everybody, like eighty eight percent of companies, are actually involved AI in their business. And if you take a look on the chart, the numbers are rising. So I I believe this year, they will report something like ninety two, ninety five percent of companies are actually involving AI into their day to day business. But it's about doing AI and doing AI. It's not the same. If you dig a bit deeper in this paper, there are many more interesting numbers, and I just pick one of them, which is only a third of companies in the survey, is capable to scale it from just POCs, testing, playground to something that is in production. And the same and maybe even more interesting numbers provides another another report by another company over here, BCG, Boston Consulting Group, in a recent paper. They stated that only five percent five percent among all companies investing in AI, So only these five percent achieve some measurable merit from AI, while others just investing money, a lot of money, but still didn't receive any added value from AI, like sixty percent of of companies, which is very, very interesting to me. And therefore, the topic of this of this workshop, how to do AI properly, how to scale from POCs to something in production that brings the value, how to join this VIP party. Of course, this is not something that happens overnight. This needs some effort. This needs some journey, and we will share with you our our approaches over here, our experience because we delivered a lot of AI solutions already. So we try to help you find the the path and to share our knowledge with you. But before I'll cover the the road map and the journey, I'll devote one slide to our alma mater, our our company. We are part of the Amdocs family. Amdocs is a global company delivering their services and systems to a lot of people, like three billion plus people, serving them all all all around the world. But these numbers are pretty pretty much known. So I just like to mention that we are also investing a lot into AI. We are delivering AI solutions. And among other things, we develop we developed agentic OS, our solution for telco and other verticals to this platform for individual agents to work together and to bring the value. I will touch this topic later on. With that said, let's get back to the roadmap. How to scale from something that works locally or something small to something that is actually running in in production and and brings the value to the company. Among other things, there are four we have identified four key concepts, three or four steps on this journey that needs to be taken properly to scale up. The first one is strategy, how to involve AI, our large language models and stuff like that into your day to day business. Second one is the fuel, the data, the fuel of AI agents. The third one is the execution. How to actually implement these agents. And the last but not least is the trust. Trust in your AI powered solutions. And actually, this is not just a road map for the journey, but also is it the road map for this very workshop. So we will devote a section for each of these steps and try to share with you our best practices in this area. Great. So that's for the opening. Now let's jump directly into the first stage, the AI strategy for first step. And I'll start with a with a conversation. As I mentioned that I am head of AI here in profit, an index company. Therefore, I'm responsible for observing the model trends on AI and how to involve them into our projects, how to deliver these advances to our customers and and get our services even better. But this the confession, I'm I'm totally overwhelmed by the speed, how everything is evolving, new AI frameworks, new AI tools, new AI platforms, etcetera. So it's very difficult to keep on track. But still, it pays off to invest some effort to follow the trends that that actually seems to be worth it, that that brings the value for others, that they can help you find a shortcut on the journey. And one of them we follow and we appreciate is this. This is the approach of having a fleet of AI agents to resolve your troubles or your your business business needs. Will describe it just briefly. On the left hand side, this is rather a monolithic approach. This is something you usually start with if you are about to start with AI. You have some some limited use case, and you open some some framework, implement or prompt AI model, which are very powerful nowadays, implement it to resolve those specific task. And it works like charm. It's perfect for demo use demo purposes. But if you would like to scale this, single model solution into production, you need to handle a lot of things. You need to take care of different corner cases because the bird around is quite heterogeneous. It's not just an example, it's a wild bird over there. Apart from the corner cases, need to add some guardrails. You need to prevent from injection. You need to be future proof, etcetera, etcetera. So from the small agent down there, it becomes a monolith bulk solution that is difficult to develop and even more difficult to maintain production. There is a parallel between having this agent, like super agent, and having some kind of employee that is capable to handle your business process from end to end. It's possible to find such a unicorn, but it's difficult. And therefore, the society developed another approach to decompose the the business process into some kind of workflow and assign specialists to different tasks and let them work together to deliver the value. This very same principle is applicable also in the in the world of AI. And it's quite easy. It maybe sounds a bit difficult, but, actually, it's not that difficult. And from my perspective, this is a great opportunity to jump into this AI era because you have three simple phases that you can follow to to get there. So if you would like to start with this agentic approach, you start with mapping your workflows. Just pick a workflow, put it on a canvas, and describe all the steps in the workflow, including back loops, including all handshakes, and and other other necessary middle steps that needs to be taken to finish the workflow. Once you have something like that, you proceed with the second phase. Identify different places where AI can be helpful, where AI can be plugged in. And these steps, move to another picture, these steps, let's say, I will take over this one. You need to isolate. You need to identify what possible inputs are there and what expected outputs are there. This way, you mark the playground for the developer for who is responsible for developing the agent, and he can focus on developing this agent. And once you have identified this thing, you can develop the agent and plug it plug it in this workflow, which is quite great. You that it's not necessary to bother yourself if other other other agents in the workflow. Just pick one one place at a time and deliver the agent, block it in, test it, supervise it, and also gain the value from the agent from the very beginning. Then you can move on to the next one, and the next one, and the next one. So that's quite straightforward approach. But one pro tip before I conclude this section is do not rely on existing workflows only. Take the identification as an opportunity to move from old existing workflows to something that more to suit better to agentic era. This way, you will save substantial effort in implementing implementing agents, and, also, it will be more straightforward to observe them, to supervise them, to monitor them, and stuff like that. And we will touch agents and how to develop them in the third section, but right now, I would like to pass to Tomasz to talk about data readiness. Thank you, Dominik. So, Dominik already mentioned that forgetting about foundations when building AI solutions will cost you. And data is, of course, the very important cornerstones when using AI to create value. So, let's have a look what we can do in this domain. Let's be honest. Having really AI read data is not the reality in which most of the companies are. And that is not only the question of Central Europe. A lot of companies are still struggling to complete cloudification projects, stabilize their data lake houses, or even initiate a proper data driven culture with data governance. But that is okay. That is the reality and it will not get better on its own. But the CDOs and other sponsors and decision makers must know what is the reality and must be aware of the limitations the current status brings. On the other hand, in our experience, it is very difficult to get sufficient budget to improve some of the AI ready data aspects when compared to receiving support and budget to build fancy AI agents that will solve all the issues we can find. Fortunately for us, the data people, the situation has been changing recently. It has become obvious that the best AI solutions will fail if the data is not adequate. That is why only so few AI projects make the cut in the end, as we have seen in the beginning of the presentation. So yes, let's cheer for the fancy AI agents, but let's get data ready too. We have said already that AI related data is important, or data in the cocktail is a very important ingredient. But there are also other parts of the equation, so we must not forget the other. Remember how difficult it was to become a data driven organization? Now, with AI at scale, it requires at least the same effort or even more to be an AI driven culture. So, having said that, we must have a look on what is actually AI ready data. We can take this definition, which is used very often by Gartner as an example, where they have five domains with specific characteristics. Surprisingly, the metadata rich domain is about metadata, but it covers not only the operational metadata, but also descriptive and business metadata, which is a little bit overlapped to the context because contextual and relevant data, as defined, should be aligned with business goals and this data is described with business domain context. And there should be also the information for what the data is actually used for from the business side. The governance and security that looks for the compliance with legal obligations, like auditability and access management and all this stuff, which we typically have in all the data solutions. Another challenging domain is the accessibility and integrated data. These data should be easily found via catalogs, APIs, and should be combined across different environments: on prem, cloud, different storage accounts, whatever. The most important characteristic is that according to the definition, they should be at least in a near real time process than make available. And that's all the important, I would say, because we know that these definitions tend to be very theoretic and does not really reflect the real state of mind. On the other hand, having the quality and structure is the same as the governance secure. That's the features we often already have in the data, and it's just about completing some parts which are missing to declare that we have AI ready data. But is it so important? We know that the data makes part of the equation how efficient the usage of AI will be. But it does not mean that we should stop doing everything we do and start transforming the solutions to comply with this definition. Because apart from many better reasons, there is this famous hype cycle used by Gartner. We will have to wait a couple of weeks to get a fresh version. Nevertheless, I don't believe that there will be a big shift in the position of AI Ready Data. And as you can see on this picture, this version is from June twenty twenty five. AI Data concept, or the idea, is at the peak of inflated expectations. What does it mean? Well, simply that we think too much about it and it will not bring as much benefit as it looks. So, it's good to be aware of what AI ready data means, but it does not mean that we should stop doing everything and just focus on this. That's also the reason why the AI ready data must not become another carrot in front of us. AI ready data is a tool. It's not the final product. It should help us to achieve something with AI, but it's not the final goal. AI ready data will not bring any value to the business. It's using the AI ready data for AI use cases that will generate the value and will help us and the business to find new ways and new opportunities and new information assets within the organization. But unfortunately, let's be frank, deploying AI solutions and agentic AI, that's just another carrot. So we have to think about the other features and not stop doing what we are doing already, which means we need to continue to provide value to the business as data providers. Yes, we have to improve the data to be more AI ready, but we must not forget about the other parts of the solutions we have already. Another important point of view is that getting the AI ready data is rather a journey. It's not a point when you will say, Okay, at this moment I have AI ready data. This is just a sample plan provided by Gartner, which is saying that for selected use cases, they estimate a project around nine months, or at least nine months, to actually get AI ready data. So, you can see that achieving these definitions can be a rather lengthy and costly business and therefore, we have to be aware of the current situation we have within the data and really pick the parts or the features which make difference to the use case we plan to deploy. This is something we will see later during today's session. However, we go for the data itself. Let's use the technology to improve the data to become AI ready. Especially metadata solutions and context are perfect task for AI. So this approach will allow progressive establishment of AI resources or even the center of excellence. It will comply also with the recommendation to start with simple tasks for AI implementation, changing simple projects, building trust, which is very important in the usage of AI. And yes, getting AI with the data using AI will not bring immediate value to the business, like new sales or more efficient marketing campaign, but it will improve the existing data solution and it will bring more value in the near future when there is the real usage of AI and agents and assistance to support the business processes and generate value. So far, I've talked about data, but since AI ready data will be quickly achieved, at least a little bit, there is another challenge waiting for us, and that is the context. This is a topic we have been focusing on with my colleagues for some time, and I'm looking forward to have an opportunity to present you our approach at the AI hotspot this fall. But now, let's go back to the AI use cases and back to Dominik. Thank you, Tomasz. Over here. So, we have covered strategies, some great approaches to involve AI into your business through business challenges. And Thomas have covered data as a fuel for the AI agents, at the end of the day, AI agents are still deep learning models that needs to be handled carefully. Right now, I will be focusing on execution, how to get these agents. But since this workshop is rather a general one and there there are many, many frameworks and options how to implement them, I will focus on principles that are applicable in general. If you are interested in some more detailed discussion considering your specific tool set, contacts, and the data you have at your disposal, reach us after after it's sent or we can have a have more detailed discussion. But let's start with this section. First of all, it's pretty straightforward, but I still want to mention that any AI solution is still a piece of software. A piece of software that needs to be handled and needs to be engineered the way to to reach something that is that's maintainable, that is sustainable, and it helps you in the long term. And as a piece of software, it has potential for many troubles to arise during the development and also after the solution is ready. Some of them have been already covered, like tight coupling and one off solutions. These are the troubles with the something small that you designed and you would like to put it into production, handling all the stuff, just inserting new new capabilities, inserting new request about handling some corner cases can destroy the stuff, destroy the functionality because you are handling with a probabilistic system, which is very, very powerful nowadays, but still it's yielding yelling the true answer or right answer only with some very high probability. But still, you need to be careful with that. Another trouble we can observe among our customers is lack of ownership. Usually, there is some expert or team of experts like center of excellence in the company responsible for agent development. They develop the agent, put it into production. Wow. Wow. It's there. But then they focus on another agent to be developed, and the old one is just living there without any guidance, without any supervision. And sooner or later, it will cause some substantial troubles. Another example of problems or flaws with AI in production is I I will start with this, like, no no no monitoring because it's the right no ownership. Still, this is the probabilistic system. You need to be very careful of using it in production. Of course, it you can you can prompt it. You can handle it. It will work for you, but still keep in mind that if there is something substantial happening about the model, you should know, and you should be able to step in. And regarding testing, as mentioned, it's a piece of software. So things like unit testing, integration testing, small testing, and then any other approaches that work to do do while developing a piece of software, it's still worth it over here. But apart from that, you need to handle the the probabilistic nature of the AI in its core. You need to design carefully the the experiments, let's say experiments, and cover all all the possibilities. Just to be sure that the AI agent behaves behaves properly, that it doesn't cause any any harm. And last but not least is misuse of AI. Of course, AI or large language models are very, very powerful these days, But still, they have some some weak points and it's and they are, of course, quite costly. So it doesn't make sense to use it everywhere just because you can. So use the AI with some wisdom. And let's get back to the to the original original strategy to break some business process into workflow and insert agents or implement agents and insert them in this workflow. If you do this for the first time and for the second time and maybe for the third time, you realize that there are many agents at your your disposal. And what's great about this approach is that you can reuse them. This is great principle of software engineering. Do not repeat yourself, and this is the same can be applied over here. So if you have a new use case, new workflow that needs to be handled, you can reuse already existing agent. These agents are well tested, approved by by by by production that they actually works. But now you need to handle the same the same activity, but in the different contract, in the different context. So you just teach the agent the new inputs and expected outputs, re prompt it a bit, but do not you do not touch the implementation anymore. So it helps you to speed up the development and implementation of AI into production. But, of course, there will be some some places, some agents that needs to be developed for this specific use case, but later on, you can take it and put it into your fleet, which will help you in the future. So far, I've been covering the the the topic in a theoretical point of view. But right now, I'd like to demonstrate this approach on on some use cases, and I picked this one, Underwriting AI, which we developed some time ago, and it demonstrates this approach as well as the reusability of agents. So this is our solution for automatic handling of any document heavy processing. Let's imagine a mortgage process. You are a client of a bank and you would like to have a mortgage. You need to send there or bring to the to the branch a lot of documents, including your ID card, the property information, the agreement, contract with the counterparty, may maybe other other documents. And then there is some pay a person in the bank responsible for handling these documents, checking all of the all the documents, check that there are all necessary documents, there's nothing nothing missing, that they are relevant to the use case, etcetera, etcetera. And this can be handled by AI. We have implemented, for this use case, multiple agents. For example, over here, this is the document, classification agent. The agent, responsible for identifying what documents did the client sent us because this is this can be anything by by mistake or or there can be a picture of a kitty or something like that among the documents. It it needs to be robust enough to handle this situation. Or there can be something missing, so the age the the agent should should shall prompt, shall ask the the the client to supply another document which is missing. Having this information, another agent attribute extraction over here is responsible for retrieving the the right pieces of information from given document. For example, let's say you have an ID card, and from ID card, we would like to retrieve information about name and surname, maybe address. If this is a property information, we would like to identify that this is the property mentioned in another document. So this piece of AI is responsible for retrieving the information online. You have an agent responsible for retrieving information from third party for, like, registry or from databases and stuff like that. Also, agent responsible for checking whether all places where signature is expected are actually actually signed. And the agent that is capable to overtake a signature from the document and compare it to to some reference signature. And, of course, there is business rules implement there are business rules implemented to be able to yield suggestion whether to proceed or not with this use case and provide some kind of report just to summarize the the reasoning behind all these agents' decisions. And this is not just an example of implementation of this agentic approach. It's also example of reusability of this single purpose, but well tested, well implemented, or engineered agents. We have tested this approach in two different use cases, one for corporate corporate lending process, where there are other types of documents, other requirements to to be checked in the registry, etcetera. But still, the only thing what which was needed in this case was to reprint or introduce a new contract to these agents, and, the rest was about reusing the whole pipeline, the whole workflow for this use case. And the same applies for the benefit handling, another completely different use case. They have different type of documents, different nature of documents, but still the principle was the same. There were some random documents incoming from client side. You need to be sure that all the documents that are requested are there. There's not there's nothing nothing missing, and you need to retrieve the information from that. Tomasz, I will pass to you regarding data. Thanks. This is a great example, right, because we have seen an existing process and we have replaced some of the activities with agents And yes, there is the next step to try to completely redesign the process and change how it works. But at this first step, we can see that all we have talked about relating to the AI related data is not fully applicable. The data about the clients is already there because it's a known client. We already have the list of values, the types of document, and even if we don't, the application using the agent is capable of dynamically adjusting the list of the values. So, what do we really need? It's not about AI read data, but it's about the foundation. We need to know how to use the structured data that will come out of the application as a result of the process, how to store them, how to process them, and how to use them for the relevant processes that are following, reporting probably triggering some other actions and events within the existing processes. So, again, focus on what brings the value and do not follow just the definitions. But it's used by a use case by use case. It cannot be a general solution. I'm muted. Great. Thank you, Tomasz. We are on track with time, so I'm able to showcase the underwriting AI, how it actually works. I'll just re reshare my screen. Just a sec. This is our repository of existing AI agents. It's can oh, maybe that can bring value and that are well implemented and ready to be inserted in some workflow. For example, here this is the document classification agent responsible for identifying what documents are there on the input side. You can insert. So, I will just use example documents for document classification over here. So, let's say this is some kind of passport from Slovakia. I'll run run the check. It will process it. It will of course, this is this is image. This is a scan. So you need to do some kind of preprocessing. This is the the thing. You can see this is the the passport specimen in this case for the demonstration purposes, and it has been identified. This is the passport. Of course, you can calibrate it. You can introduce new categories. There are many, many, many more over here. It can handle unknown types of documents. Let's say over here, this is the picture of KTI premise that will be tested, that it's capable to handle. Of course, it says it's other it does not does not fall into any category and provide some kind of reasoning. But this is just a single purpose agent. We can build on this information mined from the document and reuse it with, let's say, extractor over here. This is another agent I've been talking about. Agent responsible for mining the information from the documents given the knowledge about the document. For example, let's say I have identified that one document is an invoice. Over here, I downloaded this invoice from Internet. There are for the technology the technology demo over here, we have, like, three different type of documents. I will start with the invoice, and it's related to these parameters that are set to be mined from the document. And we will see this is something that's happening right now, but an hour ago or something like that, I've just processed this document, so I will use this one to speed things up a bit. So you can see this is the document provided. This is the invoice downloaded from the Internet provided to the agent, and you can see this is the list of of information mined from the document. You can see some kind of reasoning as well as you can see where it's located in the document. And, course, you can adjust the the value if there is something rotten, maybe there is some troubles in the scan, some misprint or something like that. So you can adjust it If you are about to use it in this interactive way and download it, store it into our database, doesn't really matter. And the same applies for the agent for the signature check, etcetera. Signature check. I'll just because we are short on time right now, so I'll just use this this approach. For example, let's say this one an example scan of of document and there was identified this this guy that is that's assigned this stuff, and it's more similar more similar to the some signature in the in the database of reference signatures. Again, this is just tech demo. This is not the final product, but we are just demonstrating this agent works, and this helps us to understand how the individual agents works and how to interconnect them together. But I promised I will demo the the underwriting AI. So let's jump in. Underwriting AI. Overwriting AI over here. You can see, again, I will I will put in a list of documents. I have, the example documents for both of the use cases mentioned in the presentation, so I will use this one. There are multiple documents. I'll just name it, work workshop and my name just to be sure and I you would like to use the prompting relevant to this loan approval process. Press run. It will happen under the hood. This one is checking. Okay? And this is something I've just checked just before the workshop, the same version of documents, etcetera, so I will use this one. So you have these documents in the inserted menu documents. Sorry. They are in check, but this is the beauty of the large language models that can they can work or for with any any input language. They are they are just this robust. And you can see it has been identified different categories. If I go down, you can see, for example, that we mined information about a tax ID, and it is the same across all the documents. And, for example, addresses are the same. This is something you would like to know see how where is the document that there is some not now document added by by chance or or by some some fraudster, and you would like to check that everything is okay. And down down downstream, there is a stated success. It's success. You can download the report. You can or maybe view the report just to to see that this is something what the the the underwriting AI yields as a as a summary. Okay? And this is just a tech demonstrator for the sake of this of this session. It can be in integrated in another another solution, another another application. So this is what I've already covered. An AI hub, this is how we call the repository of agents. You have seen just a few of them. Right now, we have this better version, but believe I do not have time to because we have run out of the forty minutes, I promised. So just keep in mind, great to have a repository of assets, of AI assets that can help you to integrate and build new workflows. And another topic I just mentioned right now, this is what we are up to. In Amdocs, we are building agentic OS, a platform where individual agents can live, can introduce your own agents into this platform or any other platforms, know, alternatives on the market. But just it's great to have the ability to not only introduce the agents over there, but also to build a workflow, interconnecting the agents, describing the contracts, and make the make the workflow happen. Actually, we have also an agent like supervisor agent or assistant in this in this platform that is capable to understand your your needs in some kind of prompt or or or input text input, and it will build the workflow proposal for you based on knowledge about what agents are there in the platform, what are their skills, and stuff like that. And I would like to also cover the trust, the last step I've I've described in the in the journey plan on the road map. This is a big topic because it's difficult to make a great or good first impression for the second time. So you need to be very sure that the agent you just built, the workflow, agentic workflow you just introduced, it it really works because users won't trust your AI solutions or your clients won't trust your solution. So the AI adoption can cause a little harm. And how to be sure that it works? There are, like, usual approaches in the world of AI. Of course, there is a supervision. Before you are pretty sure that it really works as expected, you shall take an eye on on on the the agent and its behavior. But to do so, you shall be you shall have an opportunity to dive in its communication, its outputs, and maybe some reasoning. And this, helps us also, to, achieve by cry trail guardrails, maybe some monitoring, and and stuff like that. But these are some activities that needs to be implemented, by you into the process. If you follow the advice to identify through a workflow and isolating the agent's buys through contracts, it helps you to build the trust into agents. Why or how? Because there is the contract. You have an expectation about the inputs and outputs. And if there is something happens, it's more it's most likely that you will you will receive wrong inputs or wrong outputs. And, of course, you can have checks, manual checks or automatic checks, of course, in this case, that will reveal that something is rotten about the current evaluation of the pipeline. And last slide on this is how to build the trust by design. Of course, there are multiple multiple degrees of freedom. Multiple agents are there that needs to be tested thoroughly. We are reusing them. Therefore, the effort will pay off multiple times. But still, you need to engineer them carefully and test them and monitor them. This is not the only thing. You need to also do the testing on the application level to be sure, to know what's going on, what are the outputs, how it behaves with different inputs, and also to be able to explain understand and explain what's really going on. And again, this is something the agentic workflow helps you with because it's not only the black box that yields the right answer or hopefully right answer. This is a complex system that is designed, and you can step in and observe what individual parts of the solution are how individual parts of the solution are built. And this helps you to understand what's happening and build the trust on your side, also on the client side, because you are able to explain how the results have been achieved. And with that said, we are reaching the end of this content, of this workshop. So let me just summarize the road map, the key steps on the road map to achieve this VIP group of companies who are doing the I AI the way that it actually brings the value. So we covered the strategy. You need to carefully think about involving AI into your business processes. Do not worry. It works. It really works, but it needs to be done in some reasonable way to not cause any harm. You shall have a fuel for your AI agents because, of course, this is just deep learning model, garbage in garbage out, as everybody knows. The execution, you need to think about it as a a piece of software, so with all its flaws and and benefits. Apart from that, there is AI core that needs a special handling, and you need to build the trust from the very beginning because it can if if there is no trust in your AI solutions, the customers and employees won't use it won't won't use it anymore. So and I I have reached the the last slide. Thank you for your attention, and now it's time to answer your questions, any. Well, I see there a question regarding the AI with the data. Well, whether we should think about AI with the data independently or only when building AI agents. Well, I believe that for the AI with the data, there are the features which makes sense even if you don't want to have agents or AI solutions. Right? Metadata, you know, regarding the usage of the data itself, the context, the meaning, all this does make sense even within the standard solution. On the other hand, for me, for instance, the accessibility through APIs or near real time processing, these are features which can help or can provide better services, but it needs to have a real, I would say, added value in the process. If you just want to have a real time data just to have them and there is no real time process or activity using them, there's no point. So, I would suggest just think about the features that are within the definition of the AI ready data and see what will bring you the value even if you do not decide to use these data for immediate build of AI agents. On the other hand, when you build the AI agents, then I would say it makes sense and somehow it's an obligatory path to check what data you have and whether these data can be used with the agent or the solution you are building in a reasonable way or whether there is a significant gap which would prevent the success of the agent as Dominic was showing in the very beginning. There is a vast majority of agents and AI solutions that simply fail for various reasons. Among them, one is definitely data. On the question regarding hammer on a nail. So just to read the the question, are you seeing organizations become more mature in saying no to certain a use cases? Well, not not really. We are still on the hype hype way, and I believe more organization are up to just using AI whether it's possible. Although it doesn't make sense, real sense, just to just to use it just to use it just to prove that we are future ready. But, unfortunately, this is not the way how it really brings the value. And what tends to separate those creating measurable value from those simply experimenting at scale? Well, I believe it needs to be identified the workflow and evaluate, maybe designed experiment before you jump into the AI identification or identification of the workflow. Just to be sure you can bring the value, not only it to prove that you are future ready. Meaning, if you if you decompose the the workflow if you decompose the business business feature in the in the workflow that need that you would like to identify, you pretty sure you are pretty sure you you know. I'm pretty sure you know what value does it brings to your company, and you can you can expect how agent will help you or how can agent speed up the things. Of course, you can calculate, expect number of tokens and stuff like that. So at the end of the day, you are able to, have pretty pretty precise numbers that can reveal whether your attempt will fail or or not in bringing the value. Like, of course, you need to assume that, everything goes well and the agent will be developed and it will work and stuff like that, but still some workflows, some some applications actually does not yield that much value to to defend the the investments in this in this area. But I would say that this is also bringing back, you know, the forgotten art of preparing the business case, right? So, you should have a definition of what the success will mean. And the challenge we are facing and we are seeing with some of our clients is that calculating the TCOs' AI solution is getting more and more difficult because the volatility of the price of the token is changing, right? So, something what made sense yesterday doesn't make sense anymore because simply the costs are too high. So, yes, you have to measure the output, have to compare it to cost, there needs to be a system for that. There is also one, let's say, hidden question regarding the ShadowAI. Well, the Shadow AI, we actually originally had a slide for that. The issue is that we are somehow used to have a shadow reporting. That's what we are seeing still with a lot of our clients, that there is the original or the official and certified Power BI or Tableau reports, whatever. And then there are the Excels, which are spreading across the company, and there is a lot of discussion about who has a better Excel and whether there is the correct data compared to the report or not. Nevertheless, there is a way. We do have a framework to put this under control, it's becoming somehow standard that the users can create their own reports. The same goes with the data, that you and the users have a certain autonomy to access the data, prepare the data products, but the crucial point is to have this under control. And that's what's happening to the AI as well. Some of the companies are introducing platforms like Foundry or other tools that enable the users to create their own agents. And they are facing the same challenges and they have the same issues. They do not have the costs under control. They do not control and manage the quality of the agents, and they do not have an observability of the agents. So, sometimes, or rather often it happens that, okay, there are dozens or hundreds of agents which are very similar, doing the same job, and they have been used only once, and then they are just living there in the limbo waiting to be turned off. So, shadow AI is a challenge. If the companies do not have self-service for data, for reports, it can be regarded as an opportunity to say, okay, let's introduce a complete self-service approach that will cover data, reports and AI, because the principles are the same. It has to be managed and either the companies will come up with a method or or or framework under which it can be done, or they need to find one quickly if they already allowed people and users to create their own own agents. So it's a challenge, especially from the cost perspective and also from the security perspective. And that's what I'm describing is, in my opinion, only the part when we consider that you have your somehow secure environment within your company where you have the option to create to create some agents. Then you have I'm I'm not sure if it's still a shadow or even even dark, like AI in a way that you have your own account on a different computer where you use the agent as as a as a support when doing your work and maybe transferring sensitive information to your, like, home PC and then using the agent, which is even worse than than having agents within your company not under control. Correct. Regarding applying AI to yesterday's work folks rather than redesigning workflow itself, from my perspective, it's not that big deal. It's just a thing that can help you to achieve better results by redesigning the workflows. First step is to map the workflow, and if you you spot some weak points that need can be resolved not by involving AI, by just redesigning the workflow. This is the first benefit of re of thinking about the workflows. And as a first step in this agentic area, I don't think so it's better to redesign the workflows because relying on the old workflow, it can allow you to just introduce one small piece, agentic piece, in the workflow and build on top of that. Once you are comfortable with the with this agentic approach, once you have a fleet of agents, it's much easier to rethink existing workflows and redesign how do they connect individual agents. But as a first step in this area, I I I believe the existing workflows are are fine, usually. Seems to me that we have answered, or at least tried to answer, all the questions I see in post. So, assume that will be all for today, unless there is some no, there's nothing new. Perfect. So, thank you very much for your time. I hope that you enjoyed the morning with us. Now, it's the time some break and lunch, and I hope to hear you or see you in person probably at the AI hotspot this fall or at another amdocs or profinet event. So thank you very much. Thank you. Have a nice day.