Streamline Your Procurement Process with Our Invoice Document

Invoice document for Procurement

for okay so let's get started so I see a lot of excitement about AI everyone's talking about it it's filling my feeds with uh lots of information and Snippets but when speaking to clients people are really struggling to say how do I get from the height to value now I put this question to Ben I said look Ben people are buying handfuls of co-pilot licenses they're having a dabble maybe some people are buying um SharePoint premium but they're really struggling to see the real world use cases so do you think you could sit down down pull together some examples which we can then share with everybody uh and that point Ben said well why don't we just do a webinar on this so here we are so Ben I said when to shap this I said look think about you know the people who are going to be making decisions around this and needing to be aware of what's going on in the AI world so it's going to be people Chief Information officers it directors it managers or it could be people who working you know heavily with documents on a day-to-day basis whether you're maybe a document controller or maybe you're a SharePoint expert administrator and looking to learn a little bit maybe what you can do so that's how I set the scene given all that Ben what are people going to take away from today yeah absolutely thanks ruper so so yeah we're going to look at two things we're going to look at um understand the difference differences between generative Ai and machine learning we're going to look at some real world examples some real live demos and then we're going to look at in particular how we can use machine learning to classify and extract data from documents perfect okay so when I think about AI I think it really comes into two core camps there's a generative AI side and there's a machine learning uh side to it so when I think about generative AI I'm thinking about maybe add digal assistant so some kind of chat bot where maybe trying to resolve an issue that I have or query some information or knowledge that's one side taking that a bit further have something called intelligence search so we all know that you know SharePoint search has got better over the years but it's nothing's quite as good as maybe that chat GPT style search experience so how can I use natural language to find the information which sits within my uh knowledge base or my policy library or within my internet now as I heed up internet there content creation a lot of our clients build out SharePoint internets they have great Ambitions to write content but then they have ideas but not sure how to turn that into an article so content creation that is a key value that people using uh Gena AI for and then on the flip side of that when we got too much information how can we take that information and summarize it so there the key use cases I I think I see around Gentry AI then on the flip side machine learning doesn't get the same amount of PR um as co-pilot and the the Gen side yet it's been around longer but maybe it's the audience maybe it's a bit more complicated but can you give us some use cases so people understand what machine learning can actually achieve yeah absolutely I mean I tend to think of machine learning as a more sort of traditional AI um and if we think of some examples in the Microsoft space it could be things like classifying documents you know maybe classifying a set of documents as to whether it's a contract invoice uh a legal legal document for instance things like field extraction so it could be pulling out the values of contracts or invoices start dates project names Etc key phrase extraction so it's things like identifying specific phrases that most accurately summarize the document um personal data identification so things like Social Security numbers passport numbers or driving licenses and pulling that information out image description so this could be useful in a scenario around accessibility or making images more searchable for instance and then finally this idea around named entity recognition so it could be things like pulling out the names of companies or people or perhaps products um from a uh from a given document and what we're seeing actually is with the Advent of the generative AI uh this is really starting to enhance some of those capabilities within the machine learning space so what we'll see with the field extraction scenario for instance is that concept is enhanced by the the power of generative AI as it starts to Overlay and support the machine learning scenarios okay so look given that we're limited time I said look let's pick out five key scenarios and let's do live demos of these scenarios where we can um and so we can see what these are and maybe some of value that they might bring to people but this is the S to say the business side of it what's running behind the scenes but what technology are we going to be touching on today yeah absolutely so on the generative AI side um we're looking at in particular Microsoft co-pilot which is really really prevalent really key um so that that product that targeted at end users as a product productivity enhancer co-pilot Studio which is much more targeted as Power Platform developers and helps you build things like agents and um and and those types of things chat-based experiences and then underpinning all of that is the open AI uh capability which is effectively what chat GPT is built on but deployed to asure in your environment with your data on the machine learning side we're looking at SharePoint premium which is effectively um an enhancement that natively integrates into SharePoint libraries and much more targeted end users AI Builder which is targeted at Power Platform developers again if you want to kind of integrate with uh with your your power apps for instance and then Aero AI document intelligence so again the technology that's underpinning all of that um but a bit more U involved to set up and what we're seeing with this my general thumb is that everything on the left is more targeted at users everything on the right of each box is more targeted at developers and that underlying Tech and in terms of the demos we're going to looking at today we'll look at a co-pilot demo we'll look at a scenario where we've created our own chatbot that's talking directly to AER open Ai and then on the machine learning side we've got a set of demos that are using a AI document intelligence which we've then brought into SharePoint I think the reason why that's really interesting is the commercial considerations which go along with that so the the cost of doing these kind of activities I think you'll find that fascinating let's start with co-pilot so here's a co-pilot wheel this has been r a bit is probably already out of date due to the speed that Microsoft are adding new co-pilots um and the area we focus on is what we call Modern work so that's everything you find under effectively office.com so it's things like SharePoint it's apps like word excel PowerPoint um and Microsoft teams so Ben modern work it's generally about end user productivity one of the first things that I wanted to touch on is the challenge around information overload we get so much information these days how can AI help us and you know have you got a scenario where I mean you face information overload yourself yeah absolutely so let's just take a quick look at that uh quick um screenshot um so yeah so this is something it's a screenshot that was taken about three weeks ago that number's gone up so this is my inbox um I have 13,000 unread emails in my inbox um that is a lot it's become so much information that it's impossible to really stay on top of it um so effectively I just kind of ignore it and then flag up the things I need to deal with um and just leave the rest as unread um so that can be quite triggering for some people um but yeah effectively if there was a way of just being able to summarize this and being able to you know really um tap into the core things I need to worry about this is where that summarization concept becomes really useful another area that we're seeing an awful lot of information overload on is in teams meetings we're in teams meetings all the time um so we've got teams meetings all day every day there's so many notes and actions and things that we need to capture uh it's a real struggle to stay on top of it so we're just going to demonstrate this in a video that we recorded of how we can use uh the AI capability in um teams so this is a a meeting that we carried out a few weeks ago preparing for a previous webinar and what we can see here is we've got the recording of the meeting and in the middle we have this AI notes area which is something that's part of Microsoft teams premium which autog generates notes based on the meeting on the right hand side we then have the chat-based experience which is something that comes in as part of Microsoft co-pilot um so if you want to be able to interrogate and ask questions about the meeting so as we can see I'm asking questions about what the key topics that we discussed during that meeting I can go through and do that and ask more questions and and get more detail as I need and what we'll see in a second is um I want a bit more detail on the external sharing scenario for instance and it's just going to expand um that detail in a minute behind the scenes this is what a transcript being created out of that meeting and effectively Microsoft put a sort of chat GPT layer on top of it to pull out this information yeah absolutely absolutely so that's exactly that's exactly it so effectively the meeting gets recorded you get a transcript um and then the AI notes get generated automatically off of that which were in the center of the screen there as I say that's just part of teams premium you don't need co-pilot for it and then if you want to be able to interrogate and ask questions that comes with co-pilot but again the source of that context of that content is the transcript from the meeting yeah and we're seeing quite a lot of clients actually using team premium room as the medium of you know that's actually good enough and I don't need to pay a full work for co-pilot so that's quite an interesting one there um that's meetings the next thing I want to touch on Ben is documents and you know we receiving a lot of information and a lot of documents that I receive have you know maybe marketing material there which maybe isn't that important but sometimes I'm also looking at legal documents as well and for you know if you're in procurement you can probably with through this fairly quickly um if you're not there must be other ways of getting value from this B yeah absolutely so these Master Services agreements they're very much legal focused documents written by legal professionals but if you're a lay person and you just want to understand some of the key points then being able to use summarization can be really really useful without having to read every single line so let's just take a quick look at that within my uh demo environment here so I've got a set of documents available to me I know that this is a master Services agreement I'm accessing this within one drive this functionality has been released to one drive and will be coming to SharePoint fairly soon at at the time of this recording if I click on the uh co-pilot button there I can click on the button that says summarize so as I say this is something that's part of Microsoft 365 co-pilot that you'll need to have a license to be able to use and we can see that we're getting a summary from this uh from this document it's picked out that it's a master Services agreement who it's between is picking out key points and bullets it's not simply just regurgitating the section headings but it's pulling through those things that it sees as being really important so things like confidentiality and intellectual property it's summarizing what the key status of those uh Clauses is for instance and again if I wanted to go and ask further questions of it I could click on the ask a question button and then have that more chat-based uh dialogue that we saw um that we saw earlier cool okay that's a great demo let's go on to the um next piece so Ben intelligent search so SharePoint search has definitely got better over the years um but it still requires you matching the same keyword that is in the document uh and we're getting more and more used to using a s say a chat GPT or natural language type of approach of getting information out so if we think about how we apply intelligence search into a you know document management world what's that look like what are the some of the benefits of that and um I've often heard the word Vector search being mentioned in internal meetings so maybe you could explain what that is as well okay all right good stuff so let's take a quick look at that scenario I've got two demos I'm going to run through now so the first is using something called a co-pilot agent so effectively a co-pilot agent is something that you can create that's targeted to a SharePoint site or a set of documents or a set of LI or a particular Library within a SharePoint site so you're effectively limiting the scope of the source documents to get more relevant results out of what you need so I've clicked on the co-pilot button within the marketing site and I've got a scenario where I was actually in a customer meeting a little while back and I was looking for the the PowerPoint deck that we did for one of our previous webinars and actually what happened is it had been renamed and actually moved into a different location so it kind of threw me off guard uh in that meeting and actually what I was able to do is use co-pilot to give it a sense of what I'm looking for so the context of what I'm looking for and then get a response so let's take a look at that now so find me the um PowerPoint deck that talks about the best way to structure SharePoint into a set of hubs so while this is running so effectively What's Happening Here is um co-pilot is interpreting what it is I'm asking and then it is searching documents based on their meaning so it's not just matching keywords and phrases and text it's actually looking at understanding the way uh what what is in that document this search concept is called Vector search which is basically where the document is almost turned into a really big kind of number or a vector and my question is also turned into a number and whatever documents closely match that result will come back so you don't have to match keywords it's actually matching the sort of meaning or the sense of the document Which is far more powerful and that can even go as far as the meaning but in different languages for instance which is incredibly powerful so we can see it's found this deck uh what's the best way to structure SharePoint it gives me a quick summary of what that deck talks about it looks like that's probably the right one let's have a look at that reference yeah so this is the deck um it had been renamed so as I say that kind of threw me off guard um but uh I was able to use this to actually go and get that document back without even having to filter through a set results another scenario then is within our policy Hub so we have a policy solution um we have this search experience here but again as people are starting to expect they may want to ask questions about given policies um and have that kind of question and answer experience so we've created an interface that sits on top of a Zer open Ai and what that means is we can choose which documents are going to make up part of this solution so if for instance we're only going to pass through the published versions of documents as opposed to any kind of work in progress and we have control over that and then it means we can go and sort of ask those questions as we need to so I'm going to just ask a question what is the procedure for requesting flexible working so let's so what we're doing here Ben is protecting you know your in this scenario your HR employees from getting the same questions over and over again and then they can deal with the edge cases the unique scenarios um so small investment in it yields a big investment in time saving and the other side of this is that when you raise that question you're getting an instant response you're not waiting for the HR person to come back to their email or maybe if you're a chef worker you're not waiting for them to come back in on Monday so that's pretty cool yeah absolutely and we can see it's giving me quite nice ni summary from this some of the key points from that meeting sorry from that document is giv me a clear representation of the document if I wanted to go and follow up with some further information and then the other thing this this does uh the other thing this does which we're starting to become quite used to with these um chatbot style experiences is come back with a set of follow-up options or questions that I might want to ask so let's maybe ask uh this question how to prepare for the meeting um when we're going to have that uh that meeting about for working yeah I think this is an interesting one because actually one client fedbatch was saying people are finding out more about policies just because they keep keep clicking on the follow on questions and that's actually creating a little bit of interest and curiosity about um what's in those documents okay perfect okay that's an interesting use case as well but we've actually got two references there and then what you've got in your bullet points is a reference to which document it's pointing to so very much in policies it's all about detail you can then go say well actually the policy I need to go look at is number one I'm going to click on that it's going to pop up in the screen that's pretty cool okay so taking that b the other side of you know let's say internet is very content creation so content creation we do as I said top of a call we do a lot of internets but people get stuck with content they have ideas and I think that really AI broken down a lot of barriers about what do I write about it how do I write about it you know what level of detail can I turn this into points or kind of basis off some previous content so Ben have you got a couple examples of how we can use AI to help with content creation yeah absolutely so let's take a look I think what we're going to do is just jump back to our policy scenario here so I've just sort of set up a teed up this question about remote working um again we've got our working from home policy and let's say for instance I wanted to promote this policy as a news article so you know to summarize some of the key points which I could then publish onto an internet so the previous examples were very much about finding some information and answering some questions from it and now what we're looking at doing is creating new content off the back of existing content so creating a news article from a document or creating a set of FAQs for instance so I'm just going to paste in a prompt that I uh captured earlier so we'll have a look at that so sumarize this response with an internet news article that describes some details of the benefits uh and provides three detailed FAQ so let's just have a look at that and I think you know depending on the kind of news you want you can also set the tone of voice is this good news is this more serious um and so you can get that right as well yeah absolutely so again it's referenced a couple of different documents here I've got followup questions I might want to ask but what I'm going to do is just take that content and actually I'm just going to copy it and um paste it into um a new news article I'm going to go and publish so if I just go new News Post within the policy Hub I can create a SharePoint news page and uh just going to go and paste that content just take that out put it in the title um we do a bit of tidy up I'm not going to worry about that right now so we TIY this up a little bit we may even want to or probably should um reference the actual policy itself using the file and media web part for instance put a nice image on it Etc and then once we're done we can post and publish that to our internet homepage so another scenario if you come back to the um webinar deck that we were talking about so another scenario here then is um um so another so another scenario here then is um taking this SharePoint architecture deck I want to be able to create a LinkedIn post um promoting some of the things that we talked about in our webinar um so created this content or maybe this is something we'd be doing in preparation for a webinar where we've got our deck and we want to generate some interest in the webinar before people attend so I've got that deck and I want to go and create a LinkedIn post that can help me promote it so summarize this uh PowerPoint deck into a LinkedIn article that uh has uh that describes the things to do things to stop doing has a controversial title because we do like a controversial title to capture the eye in LinkedIn and a set of next steps let's see what this comes back with I've even spelled title wrong so hopefully uh it was able to figure out what I meant even though I kind of spelled it slightly incorrectly so I do find with this uh with co-pilot you get some funny artifacts as it's scrolling but once it's finished it's it's pretty good so again we we've got our title why your SharePoint structure is failing and how to fix it so that's quite catchy um it's got some uh some good summary content in there the things to do so it's recognized those instructions and pulled out some of those key things and I can go and take that and paste it into LinkedIn and then sort of promote it so you know again we're not you know some of us aren't necessarily quite as comfortable as writing really strong copy and for something like this you don't necessarily need to go and get a dedicated copywriter for it so this has giving me a pretty good starting point that I'll then want to just tweak and put my own tone of voice and make sure it's using the kind of right right language Etc but as a good starting point yeah this is this is pretty good I'm happy with that yeah and I think the way a lot of people describe is treat uh co-pilot or chity like your intern the more instructions you get the more value you're going to get at the other end but you will need to do a review you can't you know just send it out the door um and that only that's where we are today anyway so Ben I think that wraps up the gen AI piece I want to move over to now to the machine learning and particularly the document processing side so what I want to talk about is classification documents and then extraction of information from the documents that we classify so those are the scenarios I mean you know a lot of people we don't like doing metadata you know it seems tedious and painful but we know there's value in doing it uh we can use that metadata to drive um policies around documents information how we going we going to protect those documents for so there's some of the scenarios Ben do you want me to take through what is a document model and how do we create one yeah absolutely so let's just have a look at the document model process and and just talk about what we mean so realistically what I'm talking about here is an AI model AI machine learning model that is able to carry out tasks on documents and those tasks could be classifying them into this category or that category extracting information from them or some of those other scenarios that we talked about earlier and the process that we have to go through when we're doing this typically is as follows we start off with Gathering some documents so going and finding good examples of documents that we already have or we've already worked on to be able to do what we need to do we're then going to label them so if it's a classification model we're going to label them with the appropriate categories uh that are relevant to that document um if it's a field extraction model we're going to effectively label the set of fields that we want and show the model where to get them from those documents once we've got those examples we need to train the model so this is where the AI model learns based on the examples that you've given it now fortunately we only have to train it with a relatively small number of documents compared to if we were going to do this ourselves you know not using some of the Microsoft tech because Microsoft have kind of pre-trained a whole load of these models so you can actually get going really quickly on this once the model is trained we need to test it so typically you'll set aside 30% of your documents for testing 70% for training so we're going to take our test documents apply it to the model and see what results come back with and make sure we're happy with it and then finally we're going to deploy that model into the real world to start predicting new documents moving forward so we can start um getting business value um out of this model so the first four stages are much more about creating and tuning the model and then the final stag is deploy it and making it available um to start using it in the real world okay so that's the theory Ben can we see it in practice yeah absolutely so let's take a look so I've got a couple of models here that we're going to run through um so the first we're going to focus on a classification model um I'm doing this within as your AI document Builder um or document intelligence Studio I should say um I've uploaded a whole load of existing documents here and if I click on them I can see I can see the document within this pane here and I can see I've got some categories that I can assign to it if I wanted to I could go and create a new category or just simply pick and choose from one the ones I've already created if I just collapse uh these categories on the right hand side you can see within those I've got four different C categories and I've got five documents of each category so that's quite key you need at least five documents when you're building these models um five documents for each category but the other thing to be aware of is you really need to balance how many documents you have there's no point or there's no you really don't want to do you know have a hundred statements of work but only five msas because the model is not going to understand how to interpret that appropriately so you need to balance the number of categories you've had so in this case we've got five of each um once I've uploaded those documents and tagged them the next thing that I would do is click on the train button I've done that already it takes a few minutes so we're not going to do that live so you upload them you train it and then we need to test it so I'm going to get some new documents that I've set aside for testing upload them and see what we come back with so I've got some statements of work here let's just go and drag and drop them um I'm going to drag a couple of playback documents that we use these are all fictitious company names so nothing to worry about um I'll grab a couple of uh msas okay so I've uploaded my test documents and then I'm going to run analysis on these so let's just kick that off I'm going to kick it off all of them you can see the first one's finished already and let's see what we've come up with so this document this is a master Services agreement and that has been correct IR ly identified as an MSA over on the right hand side with a confidence score of 80% so the confidence score is effectively How likely it is that this document matches the category that's been assigned to it nothing is ever going to be 100% you almost don't want it to be 100% because if it is it's a very good chance that your model has become overfit and isn't going to generalize particularly well to other examples of these types of documents so I'm pretty happy with 80% there again Master Services greement again 80% really happy with that this is a slightly different type of document it's a playback deck 64% in this or 65% I would say in this case um but it has still correctly classified it I'm not too worried about 65% it's really when it gets to 50% or lower that I'm a bit more concerned this is also a playback deck but actually much higher um in confidence so 82% so this perhaps represent as far as the model is concerned a better example of a playback deck this is the statement of work yeah that's 83% really happy with that 88% for that statement of work happy with that too 86% for that statement happy with that too so yeah I'm pretty happy with that model it's um it's it's classifying it well nothing to be too alarmed uh about in terms of those confidence scores so the next thing is say given I have classified my document and I've identified statements of work contracts um I want to then pull out information from those statements of work so effectively what I want to do is get things like project names customer names amounts dates Etc so the first thing we have to do when we're creating a field extraction model in this case is to Define that schema as we can see here so what types of fields are we trying to get and what are their you know what their sort of data types as such we then need to label it so again we need a minimum of five examples to go through and do that so as I go through each of these examples I'll zoom in a bit um I can see over on the right hand side each of those fields is telling me where it's pulling that information from so effectively you go and drag and drop the boxes around the text that you want for each field to then tell it what it needs to be so that's the project name um that's customer name the amount can see where the amounts coming from yeah perfect okay so we go through that we drag those boxes around each of those fields on the right page that you want and apply it to each uh each field there and then we go through and just do that for each of those documents so once that's done I'm ready to [Music] um so once that's done I'm ready to um train my model so I'm going to go to build models and I'll click on it and I'll build it or I'll train the one that I've already done okay that's completed this is the model that we're working with today and then and again I'm not going to do this because this takes a little while this takes about 20 minutes or so roughly speaking for a field extraction model U and then I'm going to test it thank you Microsoft um so I'm gonna upload some uh test files there and I'm going to kick these off so man it's a similar process as the classification but this time we're pulling out the fields or the en exactly yeah yeah so it's the same document model process um it's just slightly different information so that'll take a few seconds to run and we're looking at this your AI docum intelligence approach for this there are a couple other approaches we'll get to touch on those in a minute aren't we yeah exactly so let's see what this has come back with so again I'm just going to zoom in to make this a little bit easier to see custom name it's pulled that through I'm pretty happy with that and for the customer name it's pulled out a confidence of just under 80% so I'm happy with that too the amount so I can see the amount has come through on page three here and again it's given it a confidence score of 78% yeah pretty happy with that now interestingly the effort it hasn't been found so it hasn't found an effort in this particular case and it's 67% confident that it hasn't found that e that effort so that means that effort doesn't exist in that document as far as it's concerned it's that's correct 10% sure yeah okay it's 67% sure that it couldn't find it and it doesn't exist that's exactly it um the end date is two weeks now interestingly that's not a date that's a duration so again it's a bit less confident at 70% but it's still reasonably happy that it thinks it's got something for me got project name 82% now interestingly it's pulled it back for some reason from page 11 in header which appears on every page I'm not too worried because it's got the right value so I'm not going to worry about it too much and the confidence is still pretty high and then the start date 73% so what's interesting here Ben is the confidence level is at a field level whereas obviously for a document we're doing or when we're classifying the document type we're doing it at document level that's right yeah exactly that so in this case is each one of those individual values um has its own confidence and almost needs to be processed individually um which partly the reason why it takes a bit longer to run as well um if we have a look at these other documents here now we're starting to see some quite low confidence scores actually so on this particular field which is effort zero hours per month and it's 46% and it's even flagged that up in red um so it's really not sure about this and same thing for the end date which is even lower at 40% or 39% now what this is really highlighting to me is that even though this is a statement of work it's actually quite a different statement of work from say the documents I trained it on so this is actually a statement of work for a service contract as opposed to a project contract which is quite different so it's not actually been able to pull out the right type of fields so the other scenario here is this document which is pulling out the client name in this particular case um but as take a look at that that client name confidence score is quite low it's 51% so although it's not red it's still pretty low I want to take notice of this and that's quite interesting because in this case it's actually got a customer what I really wanted is the customer name and not the client name but because there's both in this particular case um it's had a stab pulled out the customer name but not the sorry the client name but not the customer name so really what this is highlighting and again um the previous example is highlighting I didn't TR it with enough data to really make it relevant so five documents was okay but the first point was I really should have separated out and had a separate model for My Success contract pull through and even taking that back to my um classification model I should have separated out success SS from Project SS and then in this scenario where I had the customer name and the client name I realized that my training data didn't actually have that it I didn't use a scenario that just had or that had both scenarios both customer and client in my training data so it hasn't actually correctly interpreted what to do so what this really highlights is as many examples as you're able to do is going to really strengthen the the um capability and really increase those confidence scores to something a bit more a bit more workable so I've created a model and now I want to actually use it in SharePoint so I've created my models I want to use them in SharePoint let's take a look at what we can do here so what you show seven wasn't in SharePoint that's in a separate environment that's part of aure and now we need to bring the two together somehow absolutely so I've created a model in aure and now I need to use it um so what I'm going to do is I've created a power automate flow here that is going to talk to those models in aure Via a rest API that is exposed by those models um you need power automate premium to be able to do that in this particular case but there are other options um so I'm going to Klick off this power automate that says classify document about a minute to run and just do it for all three now you're doing this manually document by document I'm assuming there's ways that we could automate this we're run set scale I don't expect everyone to run for every single doed start of a flow yeah I mean to be honest I'm only doing this manually just so that we can record it more easily in the um in the video um the reality is you might have something perhaps on a schedule that processes any documents that come in um over a certain period of time or you might have your power automates to trigger so anytime a new document is added but because those things can sometimes take a few minutes to trigger it's uh just purely for the purposes of recording we're doing it manually um and actually the other thing that you'll see in a second is we've got two separate flows but in reality you could build that into a single flow uh that does both a classification and an extraction okay so let's just refresh my screen so that's completed which is great so it's correctly identified the document types this is the statement of work another statement of work this is a master Services agreement and again it's pulled through those confidence scores so yeah it's 85% confident uh for the two SS 80% confident for the MSA I'm pretty happy with that now of course what you could do with those confidence scores is tie that into your um your workflow or maybe even create a view in SharePoint that says any confidence scores lower than perhaps 60% you want to take notice of that as a document manager or document controller but if it's above 80% I'm I'm pretty happy okay um so now I've got my statements of work let's go and run the field extraction model against those so I'm going to flip over to my s's view kick off uh the other another Power automate which is the extract Fields okay so at this point we've classified the document so we know we're working with statements of work so therefore we know that well we hope that these fields are within that document uh and we're doing that painful work of populating uh metadata around documents and that could then drive a you know project creation process or a workspace process or something along those lines using those key datas okay exactly exactly I understand this might take a little while to work so should we just go back at the deck look at a few of the tech options and also what that comes out of in from a commercial perspective okay perfect so the three key technologies that we're talking about here in machine learning we've got SharePoint premium on one side which I've described as user friendly that sort of power user type element AI Builder sitting in the middle and then as your AI document intelligence so B you just want give a bit more detail about each of those and and then we can talk into a particular use case where we've actually looked at the commercials against each of these um options yeah absolutely so as I said you know on the left hand side SharePoint premium is much more sort of user friendly it's natively integrated directly into document libraries much more targeted as sort of end users document manager managers document controllers who want to be able to go and experiment and build these models themselves and just have them run straight away so you know really easy to set up and use um AI Builder targeted at Power Platform developers let's say you wanted to create a power app um where you could look at documents and pull out information from them as part of a kind of invoice approval um you know custom application of some sort uh would' be really good for that and then on the right hand side a your AI document intelligence and as I said the key thing here is that the technology is the same um the underlying AI technology is the same it's just how we use it is different so in the AI document intelligence use case you don't really get anything other than a a rest API to talk to um if you want to build those Integrations you've got to do it yourself but with that it means that it's actually much more cost effective to uh run those models even if it takes a little bit more effort to to set it up so generally the rule of thumb is the easier it is to set up the more expensive it's going to be to run um uh compared to uh those slightly harder to configure models okay well you teed me up nicely there so Ben I said look I'm going to come up with an example and then I'd like you to price up using those different options so people get a feel of what those are capable of um so in this example I've got 500 documents there four different types of document in there for example contract um MSA invoice for example example and the average document is 20 pages each so the first thing I want to be able to do is classify it so I would just say look bring out those statements of work so running that classify model I've worked out now got 100 contracts on of average 20 pages each and so now I've got those I want to go and extract that key information we talked about so client project start date effort so that's for scenario now you've gone away and crunch your numbers and you're just going to take us through each of three options and how we got to the cost of each of these um Solutions yeah absolutely so with SharePoint premium as I said targeted at uh SharePoint site owners easy to set up the cost for SharePoint premium to do this is 5 cents per page um and that is for each model that you apply so for that scenario if I wanted to identify the documents within this Library out of those 500 um that's going to cost me 500 cents per model that I apply to process those Source sets of documents so if I've got four different types of models because I want to identify statements of work and do something with that etc etc that's actually going to be four times that so we're talking $2,000 to run that initial set of classification and field extraction so that is a bit of a gotcha you get charge for each model that gets applied to your documents and if you have multiple documents in a library every model always going to be applied to it the next scenario is AI Builder so as I said Power Platform developers bit more complex but more flexible um the key thing here is that there is no document classification model available um at time of uh this video within AI Builder so that means you're going to have to have done your extraction sorry you're going to have to have done your classification beforehand using some other mechanism but once you've got that and you've classified them you can um run that extraction model that's going to cost you $100 again 5 cents per page and then finally we have a zero document intelligence targeted at developers because you don't get that much in terms of integration you to set it up yourself more complex much more flexible the key thing is it does have a classification model as we saw earlier um and also that classification model runs at a different cost of the field extraction to classify documents is going to cost you.3 cents per page instead of 5 cents per page so orders of magnitude lower um and then subsequently for the documents that we do want to extract information on so those 100 statements of work for instance that's then kind of that's then going to cost you three cents per page instead of 5 cents so we can see with that scenario we're going to be charged $30 to classify all the documents and then we're going to be charged $60 to process those documents and do what we need to do what we need to do with them but it comes with that cost of ease of setup we'll need to build our own integration whether it's power automate premium logic apps or maybe some custom code for instance so I think the takeaway from this is if you're doing scale if you are processing a lot of documents with a lot of pages you need to invest in as your do intelligence if it's um you know a smaller use case maybe SharePoint premium AI Builder can help you on on your journey so with that Ben why don't we go and see how your extraction is going in your SharePoint environment yep sounds good let's see how that's uh how that's doing so we can see that that has completed now we've got the amounts we've got the customer details so again I can see that customer name's important actually matches uh the document title as well which is useful um we've got the project names the contract names uh sorry the the project names got the effort start date end date so yeah I'm pretty happy with that that's really useful being able to just pass that document in get that information out so I can then go and do something with it and as say in this case a good example could be to then go and notify a project manager who needs to then go and set this up or maybe even take it as far as automatically setting up a project in another project management system passing those values in for instance perfect okay well look I think that brings that bit to a close just a couple of things then from from my side if you're thinking about AI projects I think the first thing is although technology is awesome you need to think of your scenarios first so what are you trying to solve we showed some examples around retrieving information analyzing information um so look at your scenarios first rather than saying I've got technology where's the problem the next thing really important is data quality so uh in some of the earlier examples we talked about using intelligence search against um authoritative data so authoritative data is maybe something which has been approved and it's something that we're happy for people to search against but also when we're thinking about field extraction well we're going to classify the document first to make sure that we're extracting data from the right kind of document and then the final piece is particularly on the Gen generative side is end user training so because you're not replacing an old system with a new system there's no compulsion for people to use this technology unless you guide them so end user training becomes more important on these kind of solutions and roll outs if you want to get True Value if you'd like a conversation with myself uh and explore a bit further then just jump on our website and if you click book a discovery call that will drop something into my calendar uh if you want to get really technical I'll bring Ben into that call as well so he can give you a bit of insight as well um otherwise that's a wrap for this webinar please do join us at the next one we try and do thesea every month to share a bit of insight and knowledge around document management and SharePoint but thank you for your time