AI oriented architecture: The new enterprise paradigm - BRK2291
Morning. Everyone. I'm. Joseph Sirach I'm the CTO of AI in the worldwide commercial, business at Microsoft, and. I'm here to tell you about the. Exciting. Paradigm. Change, that's. Happening, in enterprise, AI and. I, call that the AI oriented. Architecture. For the enterprise, now. Many of you have. Heard of the service-oriented. Architecture, from decades, ago, what. Did the service oriented architecture do, to, IT, anybody. It's. Changed, the way we developed. And deployed applications. We. Created, this concept of, a service-oriented. Architecture, as opposed to an object. Oriented architecture, in programming. And, we. Encapsulated. Functionality. In services, and then, we, had, services, communicate, with each other and collectively. Form, the, architecture. For, the enterprise. So. That was, actually a revolution, in how you scale, up IT. Well. Today, there. Is a new revolution. It's not quite like service-oriented, architecture. But. It in AI oriented, architecture, where. A lot. Of functions. Instead. Of being programmed are. Being. Learned using. AI and, we. Are developing, capabilities. And applications. Some. Of which are programmed. Like in the traditional, sense some of iterators. Data. And, we are making these things work together and there, is a unique architecture. That, comes in play when. You go build, such capability. That's. What I really want to cover in this presentation and as I go through the presentation I'm. Going to actually illustrate, some of this in depth with, two case studies from. Inside, of Microsoft and, show, they are an, architecture, behind it and, hopefully you will get appreciation, for how these things come together there's. Some other things I show may be complex, but you will now find simple instantiations, of it in other cases, so. Let's dive into it so. The future of software, itself is a. Combination. Of traditional programming. Programming. Logic and. Artificial. Intelligence they, come together and. Let. Me explain that with a simple, diagram this. Is traditional, software, it. Takes input data and. Program. Logic now after a program logic is also data right it's set of instructions, in a file you. Know it takes all of that goes to compute, and then you get outputs, and, that's a programming, oriented, architecture. So. What happens, when you have an AI oriented, application, architecture, now, you still have the. Item. That I just showed there's input, there's program logic there's computer's output but there's one more thing one more piece of data and that's, models. Predictive. Models, build. For specific, functions and where, do those models, come from that, come from another workflow, there's, the learning phase so. When. You develop models, what do you have you. Have input, data, expected. Outputs the kind of results, you want right, so, you provide, those two things there. Is an AI algorithm. That then learns from that data so, input samples, and expected outputs learn, from examples, and then. Creates, models. As an output, now. The models, are predictive. And. For. Example let's, take recommendations. You know recommending. New products for you to buy well.
There's Lots of historic, examples. Where, people have. Come to the websites or something and bought lots of products, and they know what products were bought together that's, the input data samples, and their expected outputs are what people actually bought right there which, of the recommendations, is where people bind so, you know that and then from it you learn model and then you have that model, as into going into the, production phase so, this is the oriented. Architecture. Ai oriented, application, architecture, okay. So, let's now compare traditional software development web, ai development, you'll. See immediately some def. In. The traditional, software development you had data schemas, and programs, but in AI development, you have to do a lot of data curation and, labeling. Because you have to setup the teaching, environment for, AI you. Got to learn the model so you got to do a lot of nice curation, of the data and give the teaching signals. So. Then it sort it out happening at functions, and code in the. Equivalent in AI, is machine. Learning models so. Functions, take an input produce an output in the machine learning case if the model takes an input produces a predictive output when. You're debugging traditional. Software, you, attach a debugger to, a process, right and then you go step through the code well. How do you do ji it's not quite like that you, do, experimentation, and you defeated, new input see what output it produces, you, test it in flight so you actually do a/b testing, you've heard run a bunch of experiments, to see if it is producing the kind of output you want you quantify, that and that's how you do it and then, in, traditional, applications, with software, you. Use telemetry. From, the application, especially modern SAS applications, to inform, how the software, is behaving you're. Looking at how metrics, of various types well, in the case of AR applications, you're going to use that telemetry not just to inform but, to actually drive learning, that's the feedback loop from which you can actually learn and you, can continuously improve so, this, is like for example for, things that differ between the two and there, are of course more examples, now. Let's you know it go quickly to. A, couple. Of examples. One. Of the, biggest. Transformations ai, is bringing, today. Is in natural user interfaces, the. Modern-day user, interfaces, voice computing, or handwriting. Recognition or, even. Neural interfaces, of that, are being built now those. Things really wouldn't exist without, ai ai. Is used, to recognize, the signals, in very very precise ways so AI enables, natural, user interfaces the biggest revolution in human-computer, interaction. That's happening today as we speak and in, the next few years is, being driven by AI and. The. Thing is it's very power, in a bill gates recognize that in 2011 he said with natural user interfaces. Computing. Devices will adapt, to our needs, and preferences for. The very first time I mean until now we, had to really learn computers, and how, to type and how to interact with them how to move a mouse card to go click on buoys and all of that now they would start adapting to, us and in, humans will begin, to use technology, in whatever, way is most comfortable. And natural for us and that's enabled, by computers, learning, as opposed, to us humans, learning now. It's very powerful. Ok. So let me show you an example of one of those natural user interfaces one, of the little known things these days is, is the. Power of. Handwriting. Or digital. Inking, so. These are all things that are written by people, right. Try. And type this try, and enter this kind of information, the. Ink. Gives. You expressiveness. You can never achieve today without it on, out with a keyboard. Incredibly. Powerful whether it be drawing on a map or drawing pictures or, expressing.
Mathematical, Equations. How. Would you how. Do you change this how would you create, natural. User interfaces, that, allow this, freedom of expression, again. Here's where AI comes in the. Best way to see this is with, the demo it's. Already. We, have already started our journey bringing. That AI for natural user interfaces in PowerPoint, and word and to. Show this to you in action let me invite Robert, Hauser who is actually a principal. SDE in the AI in King team at Microsoft Robert, take it away. I'm. Gonna show you a few demos today and. We're going to start here with a couple of little rollin apps that you all have. Undoubtedly, seen we have word and we have PowerPoint and here, I've have word, running on a PC. And when I want to show you some of the things that Inc allows you to do and we. Really have two different paradigms you can look at in terms of gestures, and you go in terms of recognition and, here we're actually using, gestures. To do natural, things with our documents, that say I have a doc that I've been putting together I want to do a few edits on this thing so, we have this new, ink editor which, click on right here it allows you to do things like strike. Out some text, so. Magic, that goes away I can, do things like select. Just, naturally, make. A selection here I can make this bold and you increase the font all sort of controls available to me here I, can, also do things like hey I need to highlight some text because, this is super important so I can go along here make. A little highlight and. You. Know let's just say I'm really not happy with all, this I'm, just gonna scratch, it all out and start all over again so. That's. A quick example of what we can do in word and I'm. Gonna switch over here real, quick actually. Wait a second, let. Me get this ready for you. Let's. Switch over to an iPad real quick and show, you what we've been up to with PowerPoint. Nope. I hit, the wrong button. So. I'm okay, I could write this then I do. Not want to write this right now so. Here, I have my PowerPoint. Slide deck that I've been working on and we. Have a draw. Tab here and you. Actually notice that my pins actually followed me over here to the iPad and I, want to make a few edits to this so I have my pin available to me right now and I can do things like right.
Here In the in. The, box and I, can select ink to text I can. Select. This and. A. Little. A little, sloppy on the handwriting there oh. My. Goodness, that's. Always a good right, so. Say, I've actually done, a little bit of inking already and I. Want to go ahead and convert this so I can go in here and I can, select. This guy. It'll. Select I would, select with some of the ink there and it. Automatically. Converts that over to text and let's, just say I want to make this a little bit more uniform and. Fit along with the rest of the presentation. And I, can select the, ink, that represents, the rectangles, and it switches all over the rectangles, so, this is a quick a couple quick demos of what, we can do today with ink analysis, and with, that I'll, hand - Joseph. Well. That was a I that's, why he does it it's, not meant to be terrifying, but powerful. Great. So. Again, look, at what it takes to, build this. So. This is handwriting. Samples, and pictures. That the team had, to collect huge. Amounts of data they use vendors, and others then. You see the, boxes. Around them they. Used vendors to actually label, label. The data to extract, the. Key pieces to learn from so, first, step is collect. Vast amounts, the data like this create. The teaching signals so, you can actually apply out of them to learn from it so then. They. Had a number of machine learning models now, well. Is actually rich with choices. And. For different, types of tasks, you need to choose different types of models so for in classification. You would use one type for layout ambulances, you might use something like LST, ms we just got long short-term memory and our hidden Markov model them in sometimes sequence, information is. Very important, because a way you write, you, know there's a certain sequence that you're following and that needs to be captured properly, for handwriting and so you bring in the right kind of models for that for, shape recognition and, then for handwriting recognition you bring in a collection, of different models, now, you, saw. Robert. Actually do a bunch, of inking, and for. Handwriting recognition in, that particular case you know where you saw the wonderful interior, design you. You we had a number of machine, learning approaches, behind, that so, you you bring that in now but the thing is you. Need not, understand. A lot of these things that deeply, yeah at this point you. Know I actually have a saying that you know if you if you can do Indian, cooking you can actually do a machine learn then, the, most important, thing is whether you can follow a recipe or not if. You can follow a recipe if you can collect the data curate, the labels. And.
Then You. You, spend enough time just learning, how some of these recipes you, can do a pretty good job you don't have to have a PhD and develop, your own algorithms, there's a significant, collection available. Okay. So now let me actually, take. You to one more thing so we, showed you the models, being built and it's now built in a PowerPoint but, what if it could take all of that power of understanding. Ink, and handwriting. And make. That it an API on the clock but, if we could just give it to all of you as software, developers, a powerful. API on the cloud and it. Can take ink data and, do. All of that magic that you just saw in PowerPoint and you can call it from any application any software application, may need their device that is touching able to show you that demo let me again have Robert take, it away. So. Moving, back to our iPad, app here so, we, develop a web app to actually show off that we're actually powered. By a cloud service and so, this is just a simple. Progressive, web app here and what. I can do is I can Inc right, in this. Canvas and. I'm. Gonna send their request off to my service and. You. See how we did a lot of analysis gives you a time and, we. Can do things, with. The response it comes back. Let. Me do this one more time just. To. Make sure I'm showing that way I want to show off okay. So here we have some, information that came back and we learned a lot of detail cuz a lot of numbers and letters there but you, can see things like we recognize there's a triangle here and there's, bounding rectangles, and things like that and it gives you rich information about what. Was inked and how we understood, it using our our AI. And. What you can do with this you can do powerful things like I can beautify, those shapes and your. Your app can do, all kinds of different things with, that shape information that, comes back in. Addition we support things like. Text. Input, as well so I'm gonna load up a little bit text so that means before and so this, is ink as. Well I'll send that off to the service real. Quick I will. Not show you the results, just at it keep, this quick but what can tell us we, actually were able to take that ink and understand, that it actually has words. In there bread butter and jelly in this case as well, as alternates. That the language models also think could be, you know possible, matches for what we found. In. Addition we, also get layout information so we actually have to understand, things like these, are paragraphs, and lines and bulleted, lists and things like that and so we can actually tell. You what the bounding boxes look like and, when you have bounding box and information, you can actually do things like you. Know my list is not really, that great here I could clean this up a little bit and actually I can tell the app to go ahead and align that over, there to the left so you can manipulate those, ink strokes just you. Know because they have some, meaning behind them and we know what they actually are and, what the user intended, and so you can you know increase the size of those strokes all kinds of different things like that and.
The. Last thing I'm going to show you is. What's. Perhaps most powerful for the enterprise is. Because. We do understand, what the ink is we. Can. We. Can do things like actually. Search, this text right so here I'm going to search for the word ink and, I'm gonna go ahead and hit find and. We're. Actually able to find that, the, set of strokes that we represented, that word so, behind. The scenes you can not only store the ink as ink, you can actually keep the the, textual, meaning as well and so you can search and index that at a later time enabling. All kinds of powerful. Capabilities. For like field workers and such that might be using pin today or, in the future so. I'm. Gonna switch back here. Thank. You so. With, that in mind I want to touch a little bit on what Joseph, was talking about its Rai oriented, architecture, and Inc. Has been around for a long time actually been around for almost 20 years. Tablet, PC days and we've come a long way and it's been a very organic process for us to to, build the machine learning that you just, saw witnessed here, today so, this is a rough, idea of what we do a. Lot. Of this has roots in pre cloud era, so we actually still do some of this online or, an offline, now, but. Everything, starts with you know gathering, up samples it is very very difficult to gather good, quality, data but, it's something we work hard at and we're able to get over. Time and we do that through a 32 vendors and crowd sourcing things like that, that. Data all has to be cleaned up and labeled and we. Gather lots of statistics, on it because we, need to know what's what's working what isn't working. And. Then we can move that into a pipeline right, so we do training, with the GPU, clusters, and different. Technologies, you, know we run through a continuous loop of trying to determine whether our machine language. Machine. Learning is actually doing a good job and in. Tossed out models and reiterate. Again make, slight tweaks and make sure our accuracy, and performance is all good and we. Do that on a regular basis since we do ship regularly, and opera. I'm. Not gonna even try to see that word right now but basically moving us out into deployment, we. Move out into the cloud which is basically the API oh I just showed you enter, traditional, windows api's. So. The. In recap, our, mission really is to give you guys the tools that. You need to create expressive. And intuitive, and now, intelligent, ink apps. We're, looking to build faster, and more frequent updates which is why we're moving things into the cloud to. Give you guys the most consistent, experience, and bring, in bigger and better, language models, that. Can produce more powerful, results and also, give, that to, all your different devices, so. But that we want to hand it back over to to Joseph and, while he's coming up put if you want to come see more of Thursday we have a quick talk or, we have a breakout session. To. 299 is position number up so great, give it a big hand to Robert. Thank. You very much. So. Here's, one, other takeaway, I want you to have this. Is a journey, that, we. Are taking, you all on from, Microsoft, witches we. Build these sophisticated, capabilities, incredibly. Sophisticated capabilities, for inking for example, in Windows, and it's incorporating, the PowerPoint, and then, we take, that next step of. Encapsulating. These as cloud API is in Azure and bringing. Them out to all the developers, so that all that, effort the incredible, science that went in over a decade and over a decade plus, the machine learning models plus the work they did with vendors, to build, that data, labels. And teach this and test it and all of that all, of that now. Comes to you as an API, and you, can now start calling, it from like a web app on iPad which is what Robert just showed and you, can as a developer, start using it in your software application, to make your software application, intelligent, that's. The journey we are taking on so in South East keynote in the morning it's a speech you saw language, and translate, and so on we're, bringing all of these sophisticated AI, capabilities, but, you're not just algorithms, anymore that it is decades, of work by experts, but, it's now coming to you as a simple API. So. Every, enterprise of the future will, need an AI oriented, architecture, and you're. Going to build an AI or into architecture, and, what. Are you going to use it for you're going to use it to empower employees you're. Going to engage customers, and, incredibly, smart ways you're, going to optimize, operations. In important, ways and then, transform, the nature of products, itself the, product itself is going to be very different because, it has an inking capable in a naturally user interface or other, smarts around it so.
Let's, Take some examples for. Example you. Know you'll see tray a demo, of call Cora as sales assistant, that we have deployed internally, to empower our very. Large sales, force with, AI so, they can be far more effective. ASOS. Is one of our customers, is a fashion. Retailer in the UK they, have built. 13, they, have builder recommendations. So they deliver 13,000,000. Personalized, experiences, with 33, orders per second that's recommendations. And personalization. Imagine, a fashion, retailer doing, the personalization, then there's, Kensei, a startup, that is, doing an AI oriented, architecture, for health care to, optimize, operations, of hospitals, so that operations. In hospitals, can be as efficient, as possible meaning. We'll. Be staffing management, or bed management, or emergency, rooms or all of that or even predicting, chronic diseases and then, quarter, spot they've, really. Transformed, their product, with machine, learning and AI so, they've got 85%. Savings in risk calculation, and underwriting, costs using AI and machine learning models and 50%. Lower loan. Defaults, because, they have used the power of machine, learning in AI so. Let's see. Some of the components, that go into it so, on Azure now we're providing an incredibly, rich, variety, of components. For, you to build AI. Architectures. With now. You don't have to use every, one of these of course but, you have a great selection available because many architectures. Are not. The same they are different so, you've got to choose the right ones so there are soft educated, pre train models for computer, vision for, speech for, language, for search these, are all EAP is there. A popular, frameworks, that are available and open source now that you can run extremely efficiently, in the cloud whether it be PI towards tensorflow Kara's or onyx there. Are lots of productive. Options. To. Develop these machine learning models you can use actual data bricks which is spark as a service, as your, machine learning machine, learning BMS, bot service and. Behind, all of that is extremely, powerful infrastructure. There, will be CPUs, or GPUs, or AI specific. Infrastructure, like FPGAs. Which, accelerate. AI models, dramatically, and give, you incredible price performance so, that's, a paper and then, deployment, options we, allow you to actually even take models out as containers, and deploy democracies. Or on, an IOT edge or of, course in cloud we'ums in close to your application, so all of these choices are available you and you will see increasing, amount of choice as you go forward. And, Microsoft. Now is taking all of this capably listen in putting that into every. One of its software, every. Software add Microsoft. Has a will, have an AI oriented, architecture. So. Let's see, more examples, an AI oriented, architecture, to empower employees. So. Here at three, for. Example there. Is a. Telefónica. And now waterfall they, have voice. Enable, BOTS that. Can help customers, find the products they want, in. The middle is flow. Progressive. Insurance is bought in. November. Of last year it. Saw the first auto insurance, policy entirely, over, a bot on facebook, messenger. On. Right is publicist, in Microsoft. Announcing. A bot, called Marcel, for. Use by the, 80,000. Employees or so inside, a publicist, where. They have 200 subsidiaries and, their employees are using this very powerful bot, to come together collaborate. And please work. And. Now let me actually talk about our, own sales, assistant, what we have deployed inside Microsoft, so. This. Is Cora how, many of you were in South esq note here. There's a brief mention of Cora, and interacting. With quarter-over-quarter. 'no but what, it allows you as sales assistant, is to take away the heavy lifting. The, drudgery, that a sales assistant has to go through to find information, he. Needs or she needs and to, interact, with other. Peers. And, follow. Needs and so on this enormous amount, of work involved the. Assistant. Makes that simpler, so. Here are examples are the kind of problems you're solving. Typical. Seller says hey, today. We, can do much to teach thinking, because. Our jobs are so fast-paced, we. Spend a lot of time on operational issues, finding. People assembling. Data and, what. I want to focus is on selling, strategy, and customer satisfaction. How. Many of you feel that way I feel, that way very often because, a lot of my time is really spent collecting, and curating, data. At least half of the time I work as a data analyst, and the other half I tried to be a sales professional. What. If I bought, an intelligent, board could, simplify, all of that here's, a day in the life of a seller right analyzed. Gap decoder attainment, look. For news review. Customer account, plan but identify, the right people and if I write internal, external contacts, prep for a customer, meeting update, pipeline, prep for manager meeting and in the process, imagine, how many tools, that sales person is going to he's sitting in front of so many screens and context switching all the time and searching, and, that's.
Actually Very difficult, so. So. The thing is how, can a, bot, now come together get. The right recommendations. At the right place with. The right content, do, you have the right conversations. With. The right people, that's, ambition, that's, what we are putting in front of a eye to solve, as a challenge, so. Here's. What we are building. And we, have built a lot of it and there's lots more to come so. Guided. Engagement, I mean from left. Bottom, going anti-clockwise. Guided. Engagement, guided. Selling I have personalized, user experience, looking. At customer, success Pro activity competitive, intelligence referential. Intelligence, relationship intelligence. All of those are being empowered by box and. By, the way we. Have in we have a lot of machine learning models internally, which we are integrating. Behind this bar so, for example quarter setting let me tell you about this one, look. We, have tens of thousands of sales folks they. Get, a quota, set. For, sales every. Year at the beginning of the year typically. In a vast majority of organizations, houses. Code has set it's. Set by people, managing data on Excel spreadsheets and it's. Highly variable there, are thousands, of managers, in Microsoft, Salesforce. Managers, who are actually setting the quota highly variable not, always objective, well. How do you solve their problem we, took data from multiple years we looked at all the context, of the data and we built a machine learning model to, predict the quota attainment actual quarter attainment, of a, seller using. All of the data available, to us and, we found that it could actually be really, predictive, could, be in, 808, percent error rate and then, we put that in production for. A year and tested it and we, gave managers, the right to override it and it, turned out a lot of managers, didn't override, it and so we developed confidence that the model is actually setting things predicting, things well and then this year when we rolled it out at, the very beginning, of the year every, salesperson could, get a quarter set by the machine learning system that is data driven that, was objective. Compared. To you know it took three months in previous, years like the first quarter of the year was gone because there was really no quota set for the seller now the Court has said at the big of the big transformation. This, is a kind of transformation that you do with AI and then, there are models for lead scoring and churn prediction and, automated. And customized, offers so, we, are increasingly, building, models for every, activity. In a seller's lifecycle, it's. A complex diagram and I'm not going to go through all of these but. We you know these are examples, of the things we're adding in lead, scoring and routing a pure, recommender. For pure sales, content. Graph that, helps, you bring, in all the information, and, research a, slip deal predictor, so you can, advance. Gate early warning about these things, upsell. And cross-sell model. So you know what else to sell to the customer, churn. Prediction and. Forecasting. Of what might actually be the quota attainment, code. A setting, and. So on so, and. By the way this is a journey the. Thing I want, you to take away is when you put an AI oriented, architecture, in place you are taking. Yourself on a journey of continuous improvement, learning. From data continuously. Launching, new capabilities, and improving. Year after year in a very scientific way with metrics so. Now to show you the demo of the intelligent, sales bot let me invite Wanda. Thanks. Yourself, hello, everyone. This. Is something which we are calling it as Korra as, Joseph. Introduced, we are trying to get to a sales assistant where, we can get the, right context. And the right information in, a recommendation to, a stellar at the right time most.
Of The time they are trying to find the information they, remember, there was an email that came the notification, that came but, they may not know exactly where, it was so, this is our attempt to get it to the right person at the right time and save them some time and make them that much more efficient, and effective. Cora. Is also enabled on my phone so I can actually take it wherever I need to take it to and I can it is voice enabled so I can try and talk to it most of the times and it. Gives me notifications at the right time and they are proactive, notification so. Let's take a look I'm still, person, I really care for my Kota I sit, in global sales and marketing organization. It has given me a, notification. That, my deal is about to slip, so. If they can get the boil opportunity, with keen Tulsa doing seven days has seventy-four percent chance to slip in the current quarter click. On in-class experts, to maximize the impact of this deal so. As Joseph, admission, this is not just a notification that has come it's coming as a recommendation, and, to a seller who is not usually, a data analyst, or a data science genius we. Do not want them to be dumped with a lot of information, to go in process, that so, what we are trying to give it to them it's in a format where, they can use it right away in a single pane of glass and, when. It's giving me a particular, recommendation. It's not just saying this particular opportunity, is about to slip which, is by the way the slip deal predictor, model which is based on deep neural networks, that we have built and you, have tested it over multiple, quarters, and it's pretty accurate in predicting accurately. Across all the geographies, each, of our opportunities whether, it will close in a particular quarter or not it's also telling me how I can possibly go and close it because it's about to have a risk. So. I press on in-class experts. I. Found. Three in-class, experts, for you Northland Insurance, Building, you click. On schedule to book a meeting. So. Working. As an assistant which, has access to my Outlook so it can see as much information about, my calendar, and the other person's calendar, as the, other person as I would have so.
I See okay go ahead schedule. All. Right I found, these upcoming meetings which, one would you like to schedule. Great. I've. Scheduled, discuss north wing insurance, building deal for September, 26, 3:30, p.m.. And. Now I need some material. Here's. Top 10 ranked resources, I found for you on dynamics, ditch jacks so. This is the content graph that Joseph was referring to this is going through 30 different sources in the company that we have completed the tagging for these. Are different internal and external Microsoft. Resources, that we expect our sellers to have access to so, it is giving me at a particular, point in time a direct access, to a particular content, that I should be using in that conversation and it, is available to me as a PowerPoint, which I can pick edit, update and use and send it to a customer directly, or if, I need to I can project it on the screen or at, least I have it right there were suspending, 3 hours trying to find what I should be using, and. Many times where it helps in productivity is also, because a lot of sellers across different, geographies are, trying to find the similar things so, we are trying to identify a, pattern here to see what kind of material, do people use when they're trying to find something, what, is helpful and what is not one, more thing is helping us do is make us more efficient, there, are different studies which say that about 70%, of the content, generated in typical businesses, never gets used for its intended purposes, here, we are trying to keep a tab on what's getting used what's not getting used so, that then we can take keep a measure on what we should not be generating, in future, and. Where we should be investing. More to improve it further, so. Now Cory knows that I have attended the meeting because it has access to my calendar and it, sends me a notification. Greeting. You pleaded your customer, meeting with contoso, would. You like to update forecast, comments, for this deal in MSX, CRM, so. It is giving me an interface, to talk to my CRM ms fixes our deployment, of Dynamics 365.
What. It is doing is the moment I step out from a customer meeting giving me a nudge that would you like to log the meeting minutes those. Are called the golden five minutes when you step out from a meeting typically. What a seller does is they would go at the, Friday afternoon. Or a particular chosen time for their performance, they, go in and update all the notes at that, point you lose the utterance of what needs to be captured in a particular note, and in. Machine. Learning terms, we really need to have the, nuance of the language, that gets captured the words that were used in a particular meeting because that defines how, good the meeting was and what the next step should be by. Just giving me an easy access to where I can update the meeting minutes it is telling me to do it right now and it's going to mobile device but I can type, out. What. Comment what you like to add. Okay. Great. I've updated your comment, click, here to leave your comments, and CRM, so. If I want I can go and check it because at times the voice transcription. May not work that very well but, for most part it is done and it is updated, in the CRM right there so, most of the tellers or most of the people are on mobiles, so when they are walking out of the meeting they can practically, dictate, it even. Before they ditched a card sPRM, is updated, instead of waiting for a month, to update. It before the manager meetings, I just say thank you for ah. So. This is how we are getting to a point where we are trying to enable our sellers, in an AI, infused. Guided, selling so that they can be as productive, as a top 20% of the sellers in terms of efficiency, and effectiveness with. That Factory, Joseph thank you. Great. So. Now, let's talk about era, oriented, architectures, to optimize, operations I'm, going, to take a healthcare example, you know we are all going to be affected. In a positive way with the power of AI to optimize, healthcare we, are all going to go into hospitals, that will have AI behind, the scenes to give you, the very very best care let's, see some example to that Microsoft.
Has A big, investment, now in health care and using, data and AI, to, empower health care operations so. For example unlocking, the potential of health care data with the cloud so that the data is portable, using. What now is called fhir. Fire, standard, for health care data interchange. Empowering. People on the front lines of, health. Care with, the right information and, the right tools for them to be efficient, for then. Transforming. Healthcare with, precision, medicine whether it be AI or, genomics. Or immuno, Genetics, we, have amazing, capabilities. That are being built, let. Me take an example. From. A start-up. A, very. Fast. Boring, partner. Of ours called cancel, their. Mission, at cancer. Is to buy death with data science, is. A huge opportunity and. Here. Is an example what it is it's a predictive, system. Of intelligence, built. For health care so. It covers everything so patients. Come in there's. Machine, learning that is data that's being used for proactive wellness. Plan and engagement, to keep you healthy. Real-time. Care management. For. On-demand care, management services, including. Interfacing, with uber and instacart. Utilization. Driven contract, management and digital, command, centers, for. Hospital, operations now. I'll show you what those things mean here's. An example in a hospital, just take a one, portion of hospital operations you. Know patients, arrive and you've. Got to look, at for example emergency. Department, arrivals, acuity, and you pretty you have predictions, there's. An interesting, model there at the top called LW, BS prediction, let, me tell you what, that'll give you an example what this is lwb. S stands for left without, being seen. Lots. Of people, with serious conditions, get, tired of sitting in a waiting room in, the, hospital, and they, go they leave without being seen, now. If only you can predict, that somebody would leave without being seen the. Hospital, can actually prioritize those, people because if. We don't take care of them they might come back with a very serious emergency and it might cause a hospital, and the whole system a lot more so, they have the build models that are actually very predictive, so you can identify people, who might leave, without being seen and then, bring them in and take care of them and so on so for the ability, to put AI models.
In Every. Part. Of the flow workflow, in a hospital, and optimize things. Again. Lots. Of examples, here, whether, it be for variation. Analysis or, care, management, like disease progression there, are models now that can predict, the risk of chronic heart failure with, high accuracy there's, a metric, called the, area, under the curve AUC, metric which. Now certain, machine learning models can achieve about 85% AUC, it's very powerful, then. You can take you. Know think, about acute, patient for like length of stay very, important, for hospitals, to predict how, long in machine learning models being used to build that models and machine learning models being used to build that model and then machine learning models being used to build that model and then machine, learning models being used to build that model and then machine, learning models being used to build that model and then machine, learning models being used to build that model and then machine learning models being used to build that model and then machine learning models being used to build that model and then machine, learning models being used to build that model and then machine learning models being used to build that model and then machine learning models being used to build that model and then machine, learning models being used to build that model and then machine, learning models being used to build that model and then machine, learning models being used to build that model and then machine, learning models being used to build that model and then machine learning models being used to build that model and then machine, learning models being used to build that model and then machine learning models being used to build model and then machine learning models being used to build that model and then machine learning models being used to build that model and then machine learning models being used to build that model and then machine, learning models being used to build that model and then, they deployed they. Deploy, it in a flow like this this is an AI oriented, architecture, for length of stay prediction, data. Comes from EMR. Systems, electronic, medical record they ingest, the data using, services, like data breaks they, train they have what's called a feature store for computing, various features of the data their, model services, which you can build models they. Then deploy it using. Azure container, services, and then, they integrate, that back into an EMR, system or into a dashboard a power, bi dashboard, in which you can see how the how. The systems. Are performing. So. I, wanted. To now. Wrap. Start, wrapping up, look. You've now seen AI, oriented. Architectures, for. A few use, cases. You're. Now going to see. Such. Architectures. Become extremely common and, be. Built with, the cloud as a back-end, cloud. And AI coming, together is incredibly powerful, because, it puts in place in architecture, for continuous, improvement learning. And micro. Services based deployments. That allow you to do continuous updates, of machine learning models which is actually very important, how do you actually when you build a new machine learning model what.
Is The best way to put that in the production ideal. Way would be wrap, that up in a container put, that as a web service API, compare. That with existing, process, and do, an a/b testing, and flighting and see, if the new model is actually working well and then, put, that back into production all those, kinds of capabilities are enabled, by the power of the clock so. Here's. Why you should build your AI oriented, architecture, on Azure. First. It comes with the broadest set of pre-built AI capabilities. Whether. It be speech, language. Or ink. And, many. More to come you. Should expect thousands, of api's, on the, cloud for AI, that, developers, can easily integrate. The. Second, thing we have a big commitment to is that these are going to be customizable. Auto, machine, learning if you want to call it Auto ml, speech. Recognition for example you can customize, it to your voice with. Your language models. Translation. API we have you. Can actually even build a whole new translator. If you, bring pairs language. Pairs like you bring data. In say English and Chinese enough. Data you, can actually have the system learn at translator. For those, pair of languages in fact, people, have actually built those there's, some of the machine learning some, of the translators, for rare, languages, like monk. Etc. We're actually built by our customers. By, bringing their own data and they built using. Our API. Then. You'll see that we have the most advanced, conversational, with bot framework the a Shabbat, service that. Lewis language, understanding, intelligence service and so on we, have the components, required for you to set up those, kind of conversational. BOTS like you saw with Cora very. Powerful. Then. We. Also provide. The differentiated. Support for AI at the edge. Meaning. Ronnie, ray I, locally. At edge, devices, that are connected to the cloud, IOT. Edge can, now host containerize, deployments. Of machine learning models and. We also bring unique AI hardware, like FPGAs, now you can use field programmable gate arrays with, Azure machine learning they, have extremely, low latency, mission image models for example. Then. Finally we are the strongest. Enterprise Cloud for, data and AI integration. Into. Your enterprise environment, in that hybrid model, of both on-premises, and the clouds nobody. Else besides. Asher really. Enables, you to do it as easily very.
Powerful, So. Here's, know how to start. They, I journey, well. Look most, people's start, the, journey in the data with. Business. Intelligence BI. Collecting. Curating, the data, creating. Metrics and for. That year, our BI and, you have power apps plus flow you. Have Ashley, data service of various types and sequel server. But. These are available as managed services on the club then. When you step up to the next level you. Go to software, application, science, with AI and there. You have Dynamics. 365. For AI, for sales some, of the capabilities that you saw in the coral bot is now available, with Dynamics 365, you, have Dynamics, 365. AI for market insights AI for customer, service and even, for workplace, analytics, so. That's the next step up in sophistication, then. You go further up to, AI maturity. You, can integrate this API that I talked about whether it be an air API for Inc or vision, custom, vision face recognition language. Translation. There's about 30 api's like that and. Then. There are actual, cognitive, search which, is very powerful for knowledge, management and, then. When, you get more, sophisticated technical. Folks in organization, who have data science capabilities, then, you go to custom. AI where, you use sophisticated. Programming. Environments, for AI like actual data breaks or actual machine learning and open, frameworks, and there's some types of the actual cognitive services and your custom, build your own morals with your own data right that's, the journey of AI, maturity. So. The. Additional, resources for you astrodome. /ai, has a tremendous, number of resources we, have an AI school, dot microsoft.com, and AI, lab dot Microsoft, comm. So. Go. Here start. Your journey into the AI oriented, architecture, today, thank. You.