Name: AI at Work: Art of the Possible with Microsoft
Uploaded: 2025-05-20T16:01:10.052Z
Duration: 45 min 4 s
Description: AI at Work: Art of the Possible with Microsoft

Transcript for "AI at Work: Art of the Possible with Microsoft": Hi, everybody. Welcome to this session on AI and the art of the possible with Microsoft. I'm joined with Xochitl from Microsoft. Welcome, Xochitl. Thanks, Jim. Thanks for having me. I'm excited to to have this conversation with you regarding AI. So as we start the session, I'd like to do a quick recap on the core value proposition of Nasuni. So as many of you who are watching this will know, what we do at Nasuni is we eliminate legacy file silos, and we really let companies manage their data from edge to cloud with a single hybrid cloud file platform. Now this hybrid data approach enables global access to file data, so not siloed access. And it allows stakeholders across different locations to access the same information seamlessly. Now we handle the locking semantics in that in that scenario. Many of you who are using Nasuni today will probably have that pattern in action in your architecture. And this not only breaks down, what I would say, geographical barriers, it also improves collaboration because people can actually work on similar datasets in different locations or different part of the datasets. And and another thing, around Nasuni from a core value proposition, in cybersecurity, you may often come across something called the CIA triad. What that stands for, that acronym is, you know, the c is confidentiality. Only people who should access the data can access the data. And, of course, Nasuni ensures that only the people who are authenticated and authorized can access file data because it, you know, handles all of the ACLs. I is for integrity. Make sure the data is accurate and unaltered. And Nasuni stores file data as immutable objects, So files become objects, which means they cannot be changed because they're immutable. And a stands for availability. This make this means making sure the data is available as needed, when it's needed. And Nasuni can move that data to where it's needed on demand through its intelligent edge devices. And again, as you all will know, it protects and can restore it if needed, in minutes and hours, not days and months. Many of you will have seen a lot of the videos that we do around that. That global name space I talked about earlier, it can be accessed through a single edge, a single endpoint. And you can do that as you need to access the global data. And the reason you can do that is because the data is unified. It's not split into separate NAS silos. Hence, the theme of this particular event from Nasuni, Unify. Now let's quickly touch on how Nasuni's core act architecture relates to AI. I talked earlier just now about the global name space. That provides a single source of truth to global data. No messy integration projects to try and get the data in into, you know, one place. It's all done transparently by Nasuni, and it can easily be accessed and made available to AI. Also, the Nasuni edge to cloud architecture means that any fresh data from the edge can, you know, transparently be made available to AI systems in real time. So So just think about that for a second. Think about companies that have twelve, fifteen, 18 legacy NAS servers spread out across fourteen, fifteen, 16 locations around the globe. Getting data that's newly created at those locations through AI services is not easy. And you really need a hybrid cloud architecture to be able to really facilitate that. I talked earlier about that CIA triad triad. That's all, you know, precedent really on Nasuni's built in data protection. That provides AI data resilience. It's, ultimately, it underpins AI, whether that is through rag interactions, which we'll talk about in a second, or whether it's agents in AI driven agenda architectures. Ultimately, you want to make sure that data is protected, available, and restorable. I would say, ultimately, the summary, Nasuni's hybrid cloud approach to data is well suited as the data foundation because it's got a globally distributed data plane, and it's able to unify and secure data by its design. It was designed to solve some of the problems that are precedent in AI in AI today. Now I think it's worth having a a chat about, the use cases from an AI perspective. I see a lot of use cases working with customers. I've probably worked with hundreds of customers over the last twelve months. And, at this point, I'll probably bring in Xochitl as I talk about some of these. The first one, that most people probably know about today is something called RAG or retrieval augmented generation. Now RAG really combats the problem that AI has in terms of what data it has it has access to. Anybody who uses AI or has used a chatbot will have come across that annoyance where they type in a prompt, ask a question, and get told that they don't know about actually the question because of the training data cutoff. So AI is only as good as the data that it has access to. And most AI systems, that you may end up using a closed source systems or closed weight systems, which means that actually the training data, is not made public. Not all, there are open source systems as well, but a lot of systems are closed weights. What you really wanna be able to do as a company is you don't really want it working against the data it's been trained on, even though some of that data probably encompasses data that is made available through you on your website. You want it against the domain data you have inside of the enterprise, and that's where RAG comes in. Because, basically, that retrieval augmented generation architecture basically means that when you ask a prompt, there is a a retrieval step to go and retrieve the data from your data source, you know, that then comes back into that AI system and is presented to it along with the prompt. So that it can leverage that to provide you an answer, not using only the foundational models data, but also your own domain data. And that pattern has become well bedded down in the enterprise over the last twelve months. And in many ways, I would say it's been democratized. Democratized in the sense that it's no longer companies having to bolt it together themselves to get access to retrieve log meta generation datasets. But companies like Microsoft have made that, you know, much much easier and built it into their products. And at this point, Xochitl, I'll probably hand it across to you and maybe you can tell us a little bit about what Microsoft has been doing around rack. Yeah. Absolutely. Thanks, Jim. So I guess, yeah, dive a little bit deeper into kind of Microsoft's vision for AI. You know, I feel like I've said this, you know, time and time again. Our goal is really to make enterprises have availability to AI at a scalable pace, but also available and accessible to everyone. So you think from your frontline workers to your developers to your senior executives. And what does that look like? Well, we use various different tools within AI. So you can think about, one of our unified platforms, which is AI Foundry. And within AI Foundry, that is gonna be more geared towards, I would say, our developers. So you've if you've got companies that wanna build more custom LLM models or custom agents, that sort of thing, you're gonna be looking at AI Foundry as kind of that that platform to build everything custom. Now when we're talking about business users, and that's kind of, like, the big, big thing, we have ready made AI via m three sixty five Copilot. And this is where I see Nasuni really, you know, hitting home with connecting to m three sixty five Copilot. That's gonna be your more ready made assistant where you can leverage out of the box right capabilities through there that connects to companies' data. So you can think about being able to leverage, you know, the capabilities within Word, Excel, Outlook, and it's gonna give companies that ability to summarize documents, draft emails for them automatically to give time back to those users. But since that is more of an out of a box, it's gonna be more related to our business users where they don't have to be super technical to be able to build out those those agents. Additionally, kind of the neat thing within m three sixty five Copilot is you have access to the agent store. So within that agent store, we've actually just released a a new feature. It's actually rolling out here within, like, the next over phases is what we're doing with, this agent. It's called Researcher. So that's gonna be kind of, like, more of a if you think about o three minutei, that's what it's built on, our Researcher agent. It's gonna provide that deep research into the web for very specific, you know, analytics that maybe companies might be looking for. But this is very specific to m three sixty five Copilot. So there's no building that you need to do. You can just go and launch that agent within your your company, and you have it there. Now I've I've sort of just been ranting here on m three sixty five Copilot, but we do have all other ready made assistance as well, such as Copilot Studio, which provides more of a DIY, I guess you could say, approach to agents. So if you have a specific use case that isn't available through our m three sixty five, it's user friendly to where you don't have to have the develop and technical skills. You can go in, create your specialized chatbot or agents through Copilot Studio and have it create those unique workflows that are very specific to your company. But now with that, it's going to automatically just explore your company's data. Now we have companies who are gonna be like, well, what about what if I wanna connect to Nasuni? Or what if I wanna connect my data source to Salesforce? We have what's called our graph connectors, and that's where we can connect very to various data sources outside of your company to be able to surface more data into your Copilot studio or your Microsoft 365 studio. So what I really enjoy about that and there there's various other ways you can connect, but the most common one right now is through that Microsoft Graph. And that is, I know something that Nasuni has already done today with your customers, and you've helped, you know, set that up. I guess my question, Jim, to you is, I know you worked with various different customers on this aspect. How have you seen, like, those use cases and those customers? Like, have they provided any insights on how this has helped them achieve and given time back to them? Yeah. I mean, I I think that's an interesting question. I mean, I think in terms of, in terms of RAG, you know, the the big democratized use cases m three six five as you've just described. Mhmm. Because you're right. You don't need a lot of technical skills. You really need to hook the graph connector, you know, up to Nasuni, and it works on a schedule, and it takes the data that, you know, obviously, Nasuni is feeding it from the edges that I talked about earlier. So you're getting all fresh data in there as well, and then you can you can leverage any of the Microsoft 365 tools. I think that's a, a common use case that, IT departments like because they don't have to control the life cycle of the Copilot. It's built into Microsoft 365, it's part of the Microsoft 365 suite. It's not some extra tool that they have to pro provide training on. It's, you know, it's all there, you know, as soon as, you know, it's enabled. And IT departments, you know, like that for employees because it's not, an extra strain on their support, elements. Where I see, you know, and the the next use case I was gonna talk about was Copilot agents. But where I see the Copilot agents, which is what you described, in terms of Copilot studio, where I see that used and I do see them used in conjunction with Microsoft 365. It's certainly not a one choice or the other. Sometimes it's a both. And where I see Copilot agents being used, sometimes that's, I would say, front facing, you know, like sales chatbots, for example, you know, and replacing the kinda, you know, chatbot wonder all that you get on websites that were very rules based. Lots of companies, you know, using, Copilot Studio to create those, giving them access to datasets, you know, to augment the answers. But mostly what I see is it used internally. Sometimes that's for employee chat. Sometimes, you know, that's for policies, you know, just to be able to go and talk to a Copilot to get some answers on, centralized company policies. Sometimes it's specific, I would say, data for groups that they don't necessarily want to put the data into Microsoft 365 even though it's ACL protected. And that could be things like legal departments, you know, that, or HR departments that can check the contracts and they build a specific copilot and then they, you know, embed that copilot in something like Teams. And then only, you know, people from HR can access the HR team's channel and chat with that Copilot. Only legal can access the legal channel and chat with that Copilot. So I see, you know, Copilot studio, you know, used in that regard. And, of course, it's, it's as you said, it's it's WYSIWYG. What you see is what you get. So it's very easy to be able to use it and set up and create a Copilot. Actually, I often find that the fear of doing it, you know, thinking that it's gonna be really difficult to do, you know, once people start to see it in action and how quickly they can actually build and deploy something, particularly by looking into Nasuni through a graph connector. It's they they don't oh, I didn't think it was gonna be that easy. But it it is a lot. Even with Microsoft 365, it's the same. It doesn't take long to be able to, you know, pull that data in through, you know, the graph connector into Microsoft semantic kernel and then make it available alongside the datasets that they have already embedded inside of that Microsoft 365 environment. So, yeah, it's interesting. I think the the the the talking about copilot agents, you know, kinda leads us on to autonomous agents. Right. Because, obviously, the big thing for the next twelve months if the if the last twelve months has been, you know, all to do with retrieval augmented generation, I think the forthcoming twelve months are likely to be all due with agents. And it's, you know, it's funny because you I see a lot on LinkedIn that people talk about, oh, nobody can define what an agent is. You know, for me at least and I think that's partly because the definitions of talking about a co pilot agent and then on a autonomous agent, they become a little confusing. But but really, for me at least, the co pilot, agent is something you chat to. And an autonomous agent is something that is has multi steps, is task oriented, and can go away on its own and fulfill that task under the guidelines of the prompt that's it's been embedded with. Mhmm. I see I think we'll see a lot of that. And we're starting to see, obviously, some of that rolling out into some of the tooling now. You know, you maybe you can tell us a little bit about what Microsoft done around that piece. Yeah. No. So yeah. And I I agree. It's, there's I feel like there's a lot of various different meanings between behind agents. You got the agent chats, like, you I agree with you. It's kinda like more copilot. You're chatting with your agent. But, regarding, like, agents, various tasks, very various ways you can do it. And autonomous agents is definitely one of them where we're you know, I feel like when we're looking at autonomous agents, we're looking at building more custom via, like, AI Foundry, where you're kind of building on that multistep action. I know right now, you know, as as we're having this conversation, we also have Microsoft Build that is currently going on this week. And, there's going to be quite a few announcements that I suggest everyone goes goes back and looks at the book of news, regarding agents and Copilot agents as well, because there there are going to be quite a I mean, at Microsoft, we're working on constantly staying up to date with all of AI and ensuring that, you know, we're giving easy access to AI, responsible AI, and making sure that the agents that people are creating are going to be following on all of those guidelines, but also providing good intelligent feedback and, you know, responses are accurate. Just continuing on with these use cases. So another big one is, search. In this case, the Nasuni search. Mhmm. I say it's a big one because if you think about the core challenge that a lot of customers are are trying to solve with their own structured datasets, it normally comes down to, you know, visibility. Visibility of data and being able to interrogate data. And and actually, there's there's a couple of ways you can do that. You know, one is you can do the more traditional keyword search. You know, where you put in a keyword and then you get some data back in a list, and then you put another keyword and you kinda narrow it down until you get what you want. And in many cases, you can start to add, other filters into that, like date, for example, or some metadata labels that get you a pretty accurate, you know, narrowing down. And then eventually, you'll find the document you're looking for and you can access it. And then you've got the other way of actually, interacting with data. And that's these days, it's to go to a chat window and ask a question and then get an answer. And I would say, actually, that what most of the customers that I speak with are really looking for is hybrid. You know, the ability to use keyword search where it makes sense, and where AI is not helpful. And then the ability to chat with data where they just wanna get a quick answer to a question, from a particular dataset. And I can probably give some examples of, you know, at least on the keyword search to make it very clear why keyword search is still needed. If you want to go and find, for example, Xochitl if I wanted to go and find all of the, documents from Xochitl in the last week, that's traditionally not what AI would be good at doing, to your list. It's not good at those types of interactions. It's getting better for sure. But keyword search would do a very good job on that. However, if I went and asked a question and said something like, you know, what did, Xochitl tell me about research agents and how it's gonna come to, come to, m three six five. Obviously, keyword search would do a very bad job on that. And actually, I'd probably get a very good answer out of that on on on a chatbot. But the ability to get both from a central place, for me at least, is key. And, obviously, that's some of the some of what Azure AI says provides, the that ability, you know, to be able to do keyword when you need to and be able to do semantic when you need to. I think that's fair. I mean, I I would say that, there's pros and cons to either approach. Right. If you use an m three six five, of course, we talked about it being democratized earlier. Democratization means there's less, you know, bells and levers you can pull on it Okay. Since June. So you kinda get what you get, and and and the engine itself is doing all the work. Whereas, obviously, with something like Azure AI search, there's a lot of things you can tune. You know, the the semantic ranker, you know, the, the chunk size of the datasets, how, how your vector DB is set up and what type, you know, the embedded models. And all of that stuff can make the quality of what you get out the other end, you know, much better. Yep. So so when you have got, a particular use case and a particular dataset and you really need some control, Azure AI search is a really good fit for that. Yeah. Absolutely. And I sorry. Don't mean to cut you off. Maybe you're gonna complete your thought here. Well, after you. Go for it. I was just gonna say, you know, for those what I see is, like, for the very basic use cases of, like, summarizations and just, you know, looking up for those specific documents, you know, that's where the Graph Connector, I would say, is really, you know, the key because that's that's the most common use case that I see. And then when as you're getting to more of those in-depth use cases where you need that search capability, that a % makes sense. And it's I I think the the getting to that point because it it require you can't just do it in, like, a a short time frame. It requires a lot of, like, development, testing, making sure it works before it launches. And I think those resources are what's hard to come by at the time across various companies. So, anyways, carry on. Okay. And then there's we got two use cases left at the end just to to touch upon. One is, the vision use case. Yes. Now, obviously, in the last twelve months, you know, multimodality inside of models has gone a long way. And now, you know, a lot of the models, have the ability to, generate images, to be trained on images. And and in a way that previously was more difficult. Because, you know, previously things like DALLE, for example, were, you know, separate neural nets compared to, what you were getting from, you know, core chat GPT. But now but now because of the way that it's been designed, it can leverage the prompt directly in the same neural net for the image, which improves the image quality tremendously with things like what we've seen with, you know, chat GPT four o, etcetera. And, I'm interested to understand, you know, from your perspective, whether you see customers starting to leverage some of those vision use cases inside of the enterprise. Yeah. Actually, it's definitely becoming more, so I feel like the vision and the image, generations are not frequently as talked about, but I think it's because it's fairly I wanna say it's fairly new, but the capabilities have gotten much better. So now I I'm seeing a bit more, like, chatter about it. You know, how how can companies leverage this? You know, being able to just pull an image and get the basic information that they need from it is, you know, kind of one use case. Other one is, text to image. I even seen that a lot more as text to image and, you know, having it bring out all of those, you know, capabilities together. And it's actually come together quite well. I've seen various things across, like, stable diffusion, that sort of thing, where they're even coming together with, like, videos, from text to video even. So, it's not quite as common at this point in in the specific industries that I work across, but, there are definitely use cases for, like, the gaming industry. I would see a lot of things for that. And even when you're wanting to put for, like, marketing, various images together to bring something, you know, unique together, that's also, starting to become a common use case as well. Yeah. I see the same. Me media and brand companies in particular, you know, in the you know, I've seen some of those leveraging those heavy, and investing in that space because, obviously, it makes sense for some of that, creative workflows that that's that's kind of their their core IP. No. It's not perfect, but it does bring something together as a good starting point. It does. Indeed. It's the ideation phase. I agree. The ideation phase allows you to have a lot of, potentially different prototypes that you can leverage with, you know, clients and customers. And then, you know, once that they've picked the prototype they want, then it goes back to the more traditional process at least today. But you can imagine in the future that as as some of those, image generations become even better, that even that will eventually come across to AI. The last one on my list is really, around security and threat detection. Obviously, that's, you know, I touched upon that earlier, but that's a big part of what the Sony provides. It's intrinsic into the architecture. And, of course, one of the things that we, we did, last year was that we integrated, you know, with Microsoft Sentinel. To make it even better, you know, for customers who who solidify, you know, the dashboards through Sentinel from a threat detection point of view. Perhaps you can talk a little bit around what's going on, you know, in Microsoft around that area. Yeah. So with Sentinel, I mean, we're we're constantly continuing to build with Sentinel, you know, detect threat detection, that sort of thing. But with agents now, we're adding agents within those to be able to do task retrieval, send, you know, up you're gonna be sending an agent for a specific task. So it's a lot of that, you can think, autonomous agent within Sentinel and security, to drive some of those motions and kind of, you know, just make things more autonomous, and report those alerts back. So within Sentinel, we do have the capabilities to launch agents, but we also do have playbooks available that can leverage AI, to make make those integrations more seamless. So I'd I'd suggest checking out. And, again, lots of announcements this week at Build. So there's various things across AI that are being announced and developed. So take a look at that. Well, thanks for that, Xochitl. So let's, just move on and do, I guess, a kind of a a summary around how you can move forward, you know, with AI and how you can do that with Nasuni and how you can think about some of those incumbent things that might stop you actually moving forward. Because we do bump into companies that sometimes the hardest thing is making the first step. I think the first thing to understand is, you know, as I said in the first set of slides, unstructured data silos make it very difficult to have a good data foundation for AI. Because you end up having to do integration projects to be able to get the data together, or you have to start to think about how to get the data across to AI. So the first thing I would say is that one of the things you're gonna need to do, whether you do that with Nasuni or without Nasuni, is to centralize the data, Unify the data. Get the data into one place, and manageable and available to to AI. Because AI lives on a single source of truth to provide better context and quality when you interrogate it. Whether that's with RAG or whether that's gonna be with the agentic. The next thing, once you've got the data centralized, is to think about, a data assessment. One of the things you'll wanna do when you do a data assessment amongst all the other things you want to do, you know, such as, you know, governance, for example, is you'll wanna understand which of the pertinent datasets that, you know, your users will get the most benefit from with AI. And the most pertinent datasets are probably the ones that are used most actively. And you can leverage, you know, Nasuni IQ, that will track all the activity around everything that's being used and provide you a report. You can leverage that to kind of figure out where the high value data sets are that you'll be able to give to AI systems to enable your users to get better interactions with that data. Then you've got to think about, you know, how do I get my new data to AI? And the great thing with, Nasuni is when you're working with your edges, your edge devices, and you're working with, you know, data that's local because the data needs to be local because it has gravity and it needs to be near the applications, needs to be near the users because they used to used to having very low latency access to data when they're on-site. When that data, you know, is in the edge and it's cached, when you stop working on it, it's gonna go immediately back into that global name space. And if you've got AI hooked into it, like, for example, m three six five, which we talked about earlier, if you've got that actually hooked into the global name space through an edge, you know, then the AI system is gonna have access to it, you know, pretty quickly. Next thing you need to think about is how actually do you get your data to AI? That might sound a little obtuse, but it's it's a common problem that I see coming across is how do we actually hook up our data set to AI? And there's a myriad of ways you might wanna do that depending on what the use case is. You know, a lot of AI systems when they're on premise might still use file. Obviously, if you're using an edge, you've got a file interface straight in. Great thing about those, you know, intelligent edges is that they don't just provide a file interface, they also provide an object interface, an s three interface. So if you need to get object access either on premise or in the cloud because the AI system needs to hook into, an object store, you can do that, you know, really quickly. And one of the things we'll be talking about during Unify is Nasuni data services, which provides even more mechanisms for you to get access to your data in the way that AI wants to get access to it. So combination of all those things, including things like the graph connectors that we talked about earlier with Xochitl, make it easy for you to get your data from Nasuni hooked into AI services. And then I guess the last one I talked about RAG being really what's happened in the last twelve months. That's really been the big pattern. And I think in the next twelve months, it'll be around agentic. And a lot of companies will start to look at what their agentic AI readiness is, how they can identify business processes, identify applications that can be hooked up together, what those multi steps are. But at some point, you know, in that multi step, you're gonna need access to data. Now that data might be on cloud if, if it happens that the multi step business process you have identified needs access to data and it's in the cloud. But it could also be at the edge. You can imagine a scenario where, actually you've got, a a local agent that has access to other data that's only available on premise, but still needs to get access to data that's stored on Nasuni. And that can happen from the edge. And that data can still be fresh data, and it doesn't actually mean that the data was actually facilitated at that particular edge. It might have been facilitated in any of the edges that you have and then pull down to that edge because that's where the processing for that AI agent occurs. And then the last thing is from a from an agentic point of view, as these business processes, these agentic business processes get embedded down in the enterprise, what you're gonna find is is that they're gonna be kind of set and forget. There'll probably be hundreds of these agents, and they'll be underpinned, you know, by different types of data from different places. And what you don't wanna happen, particularly from your file data, if it's augmenting some of that data in one of those multi step processes, you don't want it to suddenly get locked or ransomware it out. And if it does, you wanna get it back as soon as you can. And this idea of AI data resilience is kind of the next step on from ransomware resilience. You're gonna need to have true AI data resilience, you know, when Agedentic really takes hold and, you know, Nasuni provides that because it's intrinsic to the architecture. It's not a bolt on. So that's just a, a quick summary of, you know, five steps to think about, you know, to get yourself, you know, in a better place from an AI data foundation point of view. What I would like to do, you know, as we end this, is to, you know, say thank you to Xochitl. I think she provided some great insights, you know, what's going on and what's coming, you know, from a Microsoft perspective. So, Xochitl, thank you very much. Well, thanks for having me, Jim. It was great. It was always a pleasure. And I hope you enjoyed this session. Thanks. Thanks, everyone.