David Bowman [00:00:13]:
Hello. Hi. Welcome to Fresh Perspectives. I am David Bowman, Product Director for Fresh Intranet.
Jarbas Horst [00:00:20]:
Hi, David. Look, I'm very well prepared for today. I have like a new way how I would like to kick off those recordings now. You know, there was like a Christmas gift. I have been wishing something for Christmas. And there you go. (Jarbas shows a clapboard) It's a new year, new tools, as we talked like in the previous episode.
Jarbas Horst [00:00:39]:
Yes.
Jarbas Horst [00:00:40]:
I'm Jarbas Horst. I work as a Senior Product Manager for Fresh here at Advania. And David, we're not alone today.
David Bowman [00:00:45]:
No. We have a special guest in the form of Manesh Mistry. Introduce yourself, Manesh. Tell us who you are.
Manesh Mistry [00:00:52]:
Special guest. How kind you are. Special to do that. That's so nice. Yes. So I'm Manesh Mistry. I'm one of Microsoft's Copilot Technical Specialists in the UK doing all things Copilot. Well, nearly all things Copilot, a lot of stuff on Teams looking at Microsoft, other Microsoft 365 applications such as Viva.
Manesh Mistry [00:01:12]:
So even though it says Copilot in the title, I do cover a lot of things. Very interesting.
David Bowman [00:01:16]:
Well, and it's good news because the purpose of this episode today is to talk about Copilot in all of its forms, what it is, some of the new stuff, topics, themes, ways that you and we are seeing customers using this thing and getting benefit and value out of it. But, you know, I thought we would start off today with a bit of a primer for anyone that's not familiar with Copilot. I think there's probably a fairly small number of people now, Manesh, would you mind just giving us a quick overview about what Copilot is?
Manesh Mistry [00:01:46]:
Yeah, of course, yeah. So, there's been different ways which we've tackled this question, and one of the ones which our CEO Satya has put forward is Copilot is the UI for AI. So that's how we're looking at this now. And this whole world of what generative AI is capable of doing now. The place where there's been a lot of movement and a lot of work has been done inside of Microsoft, with support from organizations around the world, is around the Microsoft 365 Copilot service. So, this is using the power of generative AI, using the OpenAI models and being able to use that inside the applications that we use on a day-to-day basis. So, for example, having that chatbot capability inside of what was Business chat is now the Microsoft 365 Copilot service and the Copilot Chat and being able to then infuse the Microsoft 365 application. So having it inside of Teams, inside of Word, Excel, PowerPoint inside of Clipchamp, inside of Stream, inside of Forms, inside of SharePoint.
Manesh Mistry [00:02:55]:
So really starting to embed that capability inside of there. But it's all about natural language and it's about being able to talk with your documents, to talk with the application, to create material for you to be able to reference and for Copilot to understand the files and meetings and emails that you have access to just now and make that part of an index which can be used to answer your questions and to create content. So from generating meeting notes, to creating sustainability plans inside of Word, to supporting the better understanding and advanced analysis of of spreadsheets inside of Excel. So lots of different areas. I've not even mentioned PowerPoint and creating PowerPoint presentations.
David Bowman [00:03:43]:
It's important.
Manesh Mistry [00:03:44]:
No, no, look, hey, listen, we've been doing some really good stuff in terms of being able to connect to SharePoint asset libraries, so having images and being able to input those images into the creative process inside of slides. So this is brilliant. And just being understanding that that is a translation of natural language into PowerPoint speak or to Excel speak. So we talk about translation from one one language to a Spanish to English or to French. But this is looking at translation into Excel, translation into a slide. How do you do that? But that's the whole point of this. And this is where Copilot is. And then for every organization there are going to be use cases where they're going to get the most out of it.
David Bowman [00:04:28]:
Yeah. One of the things that I think has been very important for lots of customers and I think one of the reasons that Copilot has been so successful to date is on the security side of things. This is all living inside the customer's tenants. So if you've signed up to Microsoft Terms conditions for, you know, which most organizations these days have done so already, this isn't sending data outside of your network, is it? This is all happening in the place that you are controlling, managing, securing.
Manesh Mistry [00:04:54]:
Yeah, it is. And it's all about understanding that. It's about your tenant, it's about the trust boundaries of where Microsoft applications live. And just thinking about your content, it's your emails, it's your files and documentation which by the way, you only have access, you'll only be able to process documentation which you have access to. So we absolutely rely on the permissions, that base of permissions which is in place already.
David Bowman [00:05:20]:
Yeah. So you know those existing investments in Microsoft 365 security permissions. This isn't something new that's got to be thought about in a different way. This is using all of that existing investment.
Manesh Mistry [00:05:31]:
Look what we've also got on top of that is purview. So the whole information governance side and making sure that what is produced as part of Copilot, prompts or actions which take place, if anything that needs to be flagged, it will be an audited, just like any email message or document. And also that sensitivity label, is it highly confidential, confidential? So all of those capabilities are still part of the Copilot service as well.
Jarbas Horst [00:05:58]:
What is great, I think we have the commercial data protection that also applies when maybe you don't have the Copilot license yet, but then you are using the free version of Copilot. Actually it's becoming much easier now with the introduction of Microsoft translation 5 Copilot chat with web grounding. So you have that commercial data protection even that scenario, isn't it?
Manesh Mistry [00:06:19]:
Absolutely, you're right. It's so important with Microsoft just making sure that everybody's on that same journey with us and making sure that when it comes to the privacy and the compliance, that still is absolutely at the forefront when it comes to using even the version which is available as part of your enterprise license in there. And look, this is how I started using it before we got really deep into the Microsoft 365 side. But it's again giving people the chance to even start to understand how to talk to a generative AI service. And really just talking to Copilot and making sure they understand what they can say, what they can put inside their prompts and the output which is going to come back. And for me, it's been 18 months I've been on this learning journey and I still continue to be on it. As to how do we talk to Copilot, how do we talk to these AI services? And I think about the realization that the organizations I've talked to, they're thinking this is a new skill. Just like you do Excel, just like you've got PowerPoint and Word, if you want to get the most out of it, you've got to understand how to do that.
Manesh Mistry [00:07:29]:
So yeah, adoption and having sessions which start to introduce, prompting and introduce the functionality has been very, very important.
David Bowman [00:07:37]:
Manesh. I considered myself to be something of a good prompter. And then I watched a demo that you were running using Copilot and I felt like a total amateur. Tell us. You know, I know it's difficult without kind of being sat in front of a prompt, but just, you know, give us some insights into sort of your approach where you're up to on this stuff. Because, you know, you're right. This is a really important skill for people to be able to start learning.
Manesh Mistry [00:08:01]:
David. It is one of these things where you're constantly learning. Even for me, if I was to look at a demo, which I did a year ago, it's not like it is today, but my approach is, it starts off quite basic. If, let's say, for example, you want to create a document, and I'm going to use sustainability as an example here, but to create a document around sustainability, it's so easy to say, hey, Copilot, you don't say that, but hey, Copilot or dear Copilot, you don't say that either. But can you create me a sustainability plan for my organization? And then you might put please at the end. All right, okay.
Jarbas Horst [00:08:38]:
Do you do that? Like, do you put please at the end? Do I Maybe at the beginning. Do you?
Manesh Mistry [00:08:42]:
I did in the first couple of months. And then as I started to get into prompting and as I got closer to throwing my computer out the window, I got to the point where I stopped saying please. And what it meant is that I never got a smiley face when. Yes, but I am. I'm pushing what the service can do. Because what I've realized with the organizations I speak to is we're trying to replicate human complexity. So it's not just about creating a sustainability plan. It's about creating a sustainability plan for my organization, which is going to be a strategy over the next five years based on these specific budgets with these specific outcomes that we're looking for as part of our organization that I'm looking for these sections which need to be created and at the end give me a conclusion and also give me a project plan as to how I should approach the full creation of a sustainability plan.
Manesh Mistry [00:09:42]:
How to get the buy in from stakeholders. And then once I've got buy in from stakeholders, then I want to make sure that we have plans in place over the next five years. Even putting in a schedule itself and looking at creating milestones and dates against it, and then be able to give that project plan to a project manager and then for them to be able to run through and so and create a risk register. David, can you see? It just becomes this. You've got to stop thinking about it, just doing one thing and start thinking about it how it can. What are you really looking for? Why are you looking for this?
David Bowman [00:10:16]:
And be specific, being as specific as possible.
Manesh Mistry [00:10:19]:
See, I should have just said that. Be specific. And I didn't need to go through all of that.
David Bowman [00:10:25]:
We've got 45 minutes to fill. So I'm happy for you to be a bit more elaborate than my throwing out words.
Manesh Mistry [00:10:32]:
Look, Dave. And the thing is, what I've also noticed is that, yes, I've said all of that and I can put this into paragraphs, but maybe the reason that you've looked at my prompts and looked at a demo and go, all right, okay. Is because it's the formatting that I use. I'm not just using a paragraph now and then embedding references to other documentation. This is about putting my references at the top, setting up what the purpose of the prompt is. So if somebody else, let's say Javas, if you were to look at the prompt, I want you to understand what it's doing in the first few lines. And then from there I put in my requirements and then at the end I've got my constraints. Because sometimes we need to ask Copilot what not to do.
Manesh Mistry [00:11:15]:
Because, for example, Copilot loves summaries. Summarization. It will give you a summarization. But sometimes all I'm looking for is a table of data and of information. I don't want summarization. And sometimes you need to tell it to do that. And this is really good for consistency.
David Bowman [00:11:33]:
Yeah.
Manesh Mistry [00:11:33]:
And making sure that every time I run this prompt, it's going to come out with almost the same detail every single time.
David Bowman [00:11:40]:
Yeah. You know, I've seen where I've seen people being critical about generative AI, you know, not just in Copilot, generally, because the prompts tend to be a bit thin. You know, a single sentence or a question and then not being entirely satisfied about the response. And I think, you know, what I saw in the demos that you were running was you are kind of anticipating that you're not going to get what you want out of response. And the more context that you're feeding this thing in the beginning and writing actually quite an elaborate prompt because you're trying to produce quite an elaborate outcome, Right?
Manesh Mistry [00:12:09]:
Yeah.
David Bowman [00:12:10]:
And it's making sure that you've got that sort of proportional balance in what you're putting in versus what you're trying to get out of it, you know, if you're trying to get a very basic outcome from this thing, a basic question is probably fine. But if you want a five year sustainability plan, that's quite a complex thing you're asking it to do.
Jarbas Horst [00:12:26]:
It's a complex task for anyone. Right. So, like, for us as well. So if you were to do that, you would, like, be sitting with people on the room thinking about all of that. So you need to be also more specific with generative AI to get kind of the outcome that will help you then bring that for further discussion with your team.
David Bowman [00:12:43]:
I think in your demos, Manesh, that you're running on Copilot conversations with customers, do you find sort of common. I would like to refer to them as ooze and ours moments.
Manesh Mistry [00:12:52]:
Right.
David Bowman [00:12:52]:
You know, Copilot's very impressive product. Generally. Are there specific things that you found that people go, you know, wow, mind blown.
Manesh Mistry [00:13:01]:
Yeah. Look, I'll take it from the feature which over 70 to 80% of people use as soon as they get their M365 Copilot license. They use it for Teams meetings.
David Bowman [00:13:12]:
Yeah. Yes, same.
Manesh Mistry [00:13:14]:
And if I think about the demos that me and my colleagues do, will spend a good 20 minutes on Teams meetings, because that's the one where people go, wow. I've been spending an hour putting meeting notes together, and you're showing me in less than 20 seconds that I've got all the topics that were covered and the action points which are put inside there. Absolutely. That is the big wow moment. But then after that, you then start to have to focus on, well, where's the value for them? So, for example, working with a court organization and prosecutors, they were looking at witness statements, and they were looking to create a narrative based on witness statements. And this process could take them three to four hours to go through multiple, like 10 witness statements and then to be able to produce that narrative which will help them understand the events which took place over a particular period of time. But the wow moment came. It's not just that it produced the narrative in just a few seconds.
Manesh Mistry [00:14:21]:
It's the bit at the bottom that it put inside there, which was. It showed the inconsistencies between the witness statements.
David Bowman [00:14:29]:
Right.
Manesh Mistry [00:14:29]:
And for them, that was the wow mo. Like, I wouldn't even consider that as an option. But it's against that power of generative AI. It's what Copilot can do in terms of being able to reference these statements to produce those inconsistencies and ones which may not have even been thought about before.
David Bowman [00:14:47]:
Yeah. I spend quite a lot of time on reviewing, reading, preparing kind of contracts for, you know, people buying products, buying services, those kind of things. And, you know, I'm not legally qualified. Right. And I should caveat that all of this stuff has to Go through our legal team before it leaves the door.
Jarbas Horst [00:15:03]:
Right.
David Bowman [00:15:03]:
So there's a safety net there. But the thing that's been very useful for me on this is asking Copilot to review the legal documents and, you know, the questions that we're getting back from partners, customers, just getting Copilot to suggest to me, you know, why are they asking these questions? Can you suggest me questions that people may ask on this stuff? Right. You know, having a buddy that's working with me on these things that can help me understand what people are looking at in some of these contracts, that's been really useful for me.
Manesh Mistry [00:15:30]:
Oh, yeah, this is like the bread and butter, isn't it? This is what Copilot can do. But I know that we've made lots of announcements of where it can go. But just to give you just another anecdotal one is one of my colleagues sent me a message and he told me, manish, can you put me a prompt together? Which put alarm bells ringing like flags, like, why are you asking me to put prompts together for you? But he asked, look, it's been a really busy day. I don't know what I've done, what I've achieved, but I just want Copilot come back and tell me. Just tell me what's been going on and just tell me to relax and chill. But, okay, I'll take that action. And I put this prompt together and it doesn't need to be about the really serious witness statements or looking at document comparison, Although that's. I'm not trying to push that away.
Manesh Mistry [00:16:14]:
Say, yeah, that's not important, it absolutely is. But I created a prompt which started off, dear Copilot, I've had a really busy day today and sometimes I just don't know what I've achieved as it's gone so fast. So tell me, based on my emails and my team's chats and the documentation, what have I been up to today? What have I achieved inside of here? But do it in the narrative of me sitting on the beach at sunset, during the sunset.
David Bowman [00:16:42]:
I like that.
Manesh Mistry [00:16:43]:
And it was fantastic. Because we're humans, we love stories. So being able to put our day inside of a story, help me to understand. Yeah, I can picture this in my mind now and I can see all the meetings I've been part of, the messages I've responded to, and it makes sense to me. Again, this has been really important because at the end of that message it says, malice, you've had an amazing day today. So just chill and relax. I Thought, well.
David Bowman [00:17:12]:
Great.
Jarbas Horst [00:17:12]:
Something to try.
Manesh Mistry [00:17:13]:
Yeah, yeah, so please try. Absolutely.
David Bowman [00:17:16]:
Yes. Yeah, we'll do. On the flip side of that, do you think there are kind of common challenges that organizations are finding with Copilot? You know, where are some of the friction points? You know, for anyone listening that's thinking about, you know, I'm about to embark on a Copilot project. We're embarking on a Copilot project. Where are some of the challenges coming in the ones.
Manesh Mistry [00:17:33]:
Again, as a technical specialist, I'm going to focus on some of the technical elements to it, but it is about the confidence of, of adding a generative AI service to the way the organization works. There is still nervousness about this and there is also nervousness about how or will Copilot learn from the data inside of our tenant. So one of the first things that we have to do is give that confidence in all these different areas in terms of that information governance. We have so many conversations around that. There's one around the data protection impact assessment that needs to take place, so all of these elements need to be there. The other one is around SharePoint and about OneDrive and making sure that there's no oversharing, which is being done. So thinking about the friction points, there needs to be absolute clarity onto these types of data and how much is going to be used when it comes to generating Copilot responses from prompts. So that's the technical side in terms of friction.
Manesh Mistry [00:18:35]:
And that could take a few months or just a few weeks, A few months, depending on the size of the organization to do that. But you know, we've got tools now available, so we have the SharePoint Advanced Management which can support looking at oversharing and producing of reports which help you understand where those oversharing bits are. So we've been able to accommodate that. But the other area is adoption and just being able to get people to use it. And this is where we've tried to be quite adventurous in how we approach this and we use these events called promptathons. Now, I don't like when we put two words together and produce another word, but promptathon is the one.
David Bowman [00:19:17]:
The jam jam. I used to love a yam jam jam.
Manesh Mistry [00:19:22]:
You know, there are some terms which I just want to throw into a bucket, but anyway, look, I'm not going to say promptathon is one of those because it was one of my colleagues which put that together, but so I'm going to put it in the buff bucket because prompted on. But it's the event thanks, David. It's the event itself and being able to take people out of their normal working environment into a space where they're working in groups of like five or six different people and you give them tasks in terms of. Let's identify the top five activities or backlogs and challenges that you have as part of your day to day running. Okay, so we start to produce those and then we start to look at the prompts which can be used to put those together. So in Microsoft we use the framework of the Goal, Context, Source and expectations or GCSE to put these prompts together and to understand the best framework to do that with. So once they've identified the activities and tasks, we then take them through the production of the prompts themselves and depending on the time, we then go onto a laptop and produce the prompt and then present them during that same session. So if it's like a full day event, that's what we do.
Manesh Mistry [00:20:31]:
So and just repeat that for me, man, that was GCSE was Goal, Context, Source and expectations.
David Bowman [00:20:40]:
Really nice little framework there for people to start using.
Manesh Mistry [00:20:43]:
It is, look, I don't want to pitch my own videos, but have a look at LinkedIn at the way I've my recent ones because I actually separate the prompts I've put together into those four different categories. And to give you examples of how that's done, I've seen one of my colleagues and he was scratching his head one day and how do I put this prompt together? And then just instantly for him, right, gcse, what's the goal, what's the content? And he started to put that together. And David, you mentioned about how some people are frustrated with just having one line and not getting what they want from it. If you can start to drive home, look, gcse, think about all these elements when you're putting your prompt together, then yeah, that's what really helps. But again, that's the prompt thon session. They go away after an hour and a half or two hours, a half a day and they go, wow, I can see this now.
David Bowman [00:21:32]:
Yeah.
Manesh Mistry [00:21:33]:
But then after that, the whole adoption piece, you still need to have these regular catch ups. You need to have these ask me anything sessions or open office hours where you get people to come in and just talk about some of the good things, the highlights and the lowlights of how Copilot is running as part of the organization. And then there's the reporting and the analytics at the back end which can show which different departments do I really need to focus my attention On So like I said, technical, it's all about the information governance and then on the user side it's about adoption. That's where we're seeing we need to do the most work to get this to work properly.
Jarbas Horst [00:22:08]:
Yeah, you mentioned some very nice use cases here, Manesh. Like what's kind of your main use case that you use Copilot for?
Manesh Mistry [00:22:16]:
Oh, that's a good one. Right. So there's two parts to this one. The main reason I use Copilot 5050 here is one, I'm using it to support the organizations that I work with and that I'm helping them understand how to put prompts together based on their processes. So part of my time is spent actually producing prompts for other people. That's what I use it for. But for me I. There's so many different areas I've looked at.
Manesh Mistry [00:22:44]:
So let's take creating a post, for example, a LinkedIn post. Then I will create my text and then I'll push that through Copilot to tell me, okay, I want this to be a LinkedIn post. If you need to change the language, do that, but make the post relatable. And that has helped me. I don't really use it for email. It's a funny one, isn't it? I don't use it for the creation of emails for me or to draft, but I do use it to create content. So just as an example for me, I was running a Teams Phone training session, an internal one. What I needed to do was produce all the different areas that Teams Phone integrates with.
Manesh Mistry [00:23:24]:
So thinking about the devices you use, looking at the manufacturers of telephony systems, all of these different elements, I needed support in to create that presentation and to create the documentation to support the presentation. This was a, it was, it was a three hour task which I used with Copilot, but this could have taken me days to do. But because it gave me the draft, it gave me the structure and I had this conversation with it to understand. Right, yeah, you've created this for me, but I want you to make changes here, do another change here, undo that one, make it into table, now create it into a presentation. All these things have been so useful for me as a technical specialist because of the way I support my colleagues around the business. So that's how I use it. How about you tell me.
David Bowman [00:24:11]:
Yeah, it's interesting you say that about email and I think that's probably my, outside of the Teams meeting thing, that probably my main use case is email. I've had feedback Historically that some of my emails come off as being a bit blunt, somewhat demanding, sometimes written in haste and you know, actually having something that I can sound this off in front of, you know, how will Persona A respond, react to the email? Am I going to get the outcome? You know, this is the outcome I'm looking for from this email. Am I going to get it from this type of person or this type of person or. So you joked about this when we were talking earlier about.
Manesh Mistry [00:24:48]:
Yeah, yes.
Jarbas Horst [00:24:53]:
Yes. For me the use case that I'm using Copilot. So first of all like meetings, right? So in the summaries around meetings, all of that. So that's definitely really powerful and I'm starting to use that more and more definite like first use case. Then the second thing is like finding information. You know, you go to the Internet, to the place where maybe documents should be, you go there, you don't find it. Then tendency is like that, you go to the colleague that you know that person has worked with the content and you might ask that person, right? It's very convenient. Now I have been using Copilot for that.
Jarbas Horst [00:25:25]:
So David is getting less questions nowadays about kind of hey David, where's like that document? Because we collaborate a lot.
David Bowman [00:25:31]:
So now I'm getting less unpleasant emails in response.
Jarbas Horst [00:25:35]:
It's a win win for both and Copilot is helping us there. So it's a nice thing. But I think having Copilot is like this companion throughout the day. Right. So you need maybe support with an email, you need for support for creating a LinkedIn post. Very common also then the aspects of finding information, that's the part that I like. Right. So there are many scenarios where you can apply it too.
David Bowman [00:25:55]:
But I think the point about adoption and helping people to contextualize this, right. You know, it doesn't need to be sort of force feeding people. These are the prompts that you should use every day. It's providing enough guidance and examples that people are able to start putting that into their own context, right? Yeah, you know, having some framework for people to refer back to, you know, helping them to find, you know, here's the kind of killer use cases that you're going to get most benefit out of that sort of the adoption challenges as we see it.
Manesh Mistry [00:26:19]:
So David, one of the things that one of my colleagues has created is something called Prompt Buddy. So this is available on GitHub and many of the organizations we work with, they've adopted this, they've added it into their tenant. This becomes this community place where you Store your prompts. So it's not really controlled in terms of moderation, but it's about getting your prompts there, the ones which you found the most useful and adding tags such as I've used it for this department or for this use case or for this subject matter and being able to collate all of that in one place and then to start voting up, voting some of these prompts as to how useful they are. I mean, inside of Microsoft there's one prompt which is called one prompt to rule them all. That's the title of the prompt. But this is the one inside of Microsoft which has got the most votes against it on what it can do. But that's again, there's an example PromptsBody is one way, which is crowdsourcing for community, but there's also the prompt gallery, which is more of an official place where inside of your tenant, where you can officially store prompts which are going to be useful around the organization.
Manesh Mistry [00:27:23]:
And inside of Copilot you've got the little stars where you can find out more about prompts and get ideas for prompts. This is where it would appear. Yeah.
Jarbas Horst [00:27:31]:
Look, as people learn more and more about the technology, like they also start to use it more, they learn more about limitations and then the wishes come. Right. So that can be a feature request, which sometimes can take time to get implemented. It's Copilot black box. Can that be extended?
Manesh Mistry [00:27:46]:
Absolutely. And hopefully when you say extending, we can start to bring in other applications into the way that Copilot works. Because one of the questions we were asked quite early on is it's great that we're able to access our emails and chats and 365 based files and documentation, but we have case notes inside of a CRM system and we're looking at HR data in another system. What would be really useful is to be able to add that into the way that Copilot works. So yeah, absolutely. When it comes to extensibility and being able to gather knowledge from other applications and other data sources, yes, that's what we've really been working quite hard on just now in extending that. And by the way, it's not just about the third party applications. Sometimes when you're looking for data as part of a Copilot chatbot, what you want to do is not look at everything.
Manesh Mistry [00:28:45]:
You just want to have a look at one document or one folder or one SharePoint site. So these are all the capabilities that we're introducing now and I guess we're.
David Bowman [00:28:54]:
Kind of getting into the topic of agents here. Right. Which is kind of a relatively new concept in terms of Copilot, been kicking around in the world of generative AI for a little bit longer. And it's an area that we've been interested in. Because, you know, what this is providing people with is the ability to maybe have something that's a bit more focused, targeted on specific scenarios. You know, we've been thinking about the intranet as an example of one of those scenarios. Can you tell us a little bit more about agents?
Manesh Mistry [00:29:21]:
Yes. What are they? What are agents? Right, okay. And I sort of explained it already. But what we're trying to do here is we're trying to bring in the data and the knowledge which organizations have, which are not inside or which could be inside of M365, but also outside of it. And that agency is that portal that way into. We talk about again, Copilot is the UI for AI. Well, behind that UI is going to be agents, and it is that interface into all of these different areas. But to make this into a natural language, a conversational process, to be able to bring this in to talk with a user.
Manesh Mistry [00:30:00]:
But we're also looking at agents not just to understand that interface to people and the data which they have access to, but it's also to understand business process. Because we're going to be seeing this year, this is the year, isn't it, that we're going to be looking at having business processes in the background, which would be deterministic, that we'll have some level of what it should be doing. There'll be a flow, a process flow. But how about if that flow could be dynamic? What if we could use that generative AI capability to apply reasoning to that process and that the only time that we need to be aware of it is when something doesn't compute, something's not right, and then it will contact us and tell us. Right, you need to help me out here so I can continue finishing off this process. So this is the backend, the agentic process in the background, which also forms part of what agents is. So, yeah, this is so interesting as a way we're looking at it inside of Microsoft.
Jarbas Horst [00:30:55]:
Yeah, like your sake. I was saying that in the podcast the other day. So that as you mentioned, Right. So the agents as the interface and like the third part, solutions, for example, as the foundation providing, then like the agents with the data providing the API so that the things can be connected. And this shows a bit, a bit of a disruption in the way maybe how we might experience work as things move forward, but it becomes more, I think, productive and efficient by having that interface, the agent performing not just the retrieval information, but also performing the tasks on your behalf.
Manesh Mistry [00:31:28]:
Yes.
Jarbas Horst [00:31:29]:
I'm interested to see how things will be evolving in this space over time.
David Bowman [00:31:33]:
And from an end user perspective. These agents are primarily interacted with in the Copilot interface. So I can summon an agent that's been created in Copilot and have a conversation about a specific task or scenario.
Manesh Mistry [00:31:46]:
Summon, you make that all magical, don't you? Summon is my try.
David Bowman [00:31:50]:
I'm trying.
Manesh Mistry [00:31:52]:
Well, look, look, that's a great word, is summon. I think that's such a good one. And if I think about one of the examples I'm working with inside of agents is right. So when we first started it was about, hey cool, let's focus on this document library of policies. So when I ask a question like a bot, I can ask a question about an HR policy to do with travel and expenses. Great, I'll get an answer from there. I thought that's good. And look, this is where it's like first step into the process of using these agents.
Manesh Mistry [00:32:22]:
But I don't want to go back to your summon thing because that's really good. And one of the demos that I'm working through at the moment is a triage risk assessment. So that's to either use some case notes or a transcript with a conversation with somebody and then be able to put scoring and to understand the risk. So in case of social workers understand the risk that this person has based on this set of criteria and then create this rubric, this scoring mechanism to be able to understand the risk for that particular user on what the social worker should be looking at first and what should be highlighted. I thought this is an amazing, amazing use case for looking at, prompting. But then I thought to myself, well, this prompt is huge and I think there's going to be a lot of people who will struggle with being able to copy and paste that, make the changes that they need to make. So how do we make that simpler? Because this now for me is to understand the why, why do we have agents in place? So now what I've been able to do with this agent is I've hidden the complexity of my prompt inside of the agent. So when we interact with that agent, it already knows what the prompt is in the backend.
Manesh Mistry [00:33:34]:
It's already got it inside its memory. All we need to do is feed it the transcript or feed it the reference documentation for it to do its own Scoring and assessment. Risk assessment. So what used to be a good, what let's say 2000, more than 2000, over 2000 characters of a prompt. What I do now is inside my little Copilot box prompt box, I reference the transcript or a doc or a PDF and then I just mention the agent. I press the send button and it produces my risk assessment. Yeah, this to me really shows the power of what this is capable of doing is hiding that human complexity inside of an agent. And then once that's run, we can continue to ask questions about the transcript in the context of the scoring and the risk assessment we've already put together.
David Bowman [00:34:30]:
Yeah, it really as you say, kind of cutting down on the amount of training time onboarding that people need to these certain scenarios. You know, some of these prompts need to be pretty complicated. Right. You know, we're talking about that complicated outcome needs a complicated prompt of being able to box that complicated prompt, wrap it in an agent and that allow people to provide minimal information with the agent actually doing quite a lot of the legwork before it hits.
Manesh Mistry [00:34:55]:
Yeah.
David Bowman [00:34:56]:
For its Copilot.
Manesh Mistry [00:34:57]:
So just thinking about going into another example around the employee self service. So and again, just going back to that HR1, yes, we can ask a question about a policy and to do with travel and expenses, but how about if we were to integrate a travel and expenses application into that same agent. So now not only can we ask general policy questions, we can now start asking personal ones about the data which belongs to us. So tell me about the approvals that I need to do for travel and expenses. Who are they? And to use that, maybe use adaptive cards to be able to bring back that data and then be able to approve expenses based on the information which is inside of this chatbot, this interface. But at the back end it's triggering the right applications and inside of our API to make the changes to the approval system inside of our expenses platform.
David Bowman [00:35:54]:
Yeah, you know, I guess so it's more than just producing text or image information that you've got kind of user interface elements turning up in this prompt as well. And I guess, you know, this is some of the areas that we've been thinking about. This is part of intranet product. Right. That you've got elements of what you would have been looking at on a homepage actually turning up in responses to prompts through things like, you know, an Internet agent, a news agent, something like that.
Manesh Mistry [00:36:19]:
How cool is that? Yeah. Being able to inside of an intranet to give you information which is related to you inside Directly related to you. I love that. I mean, we're trying. So as part of Viva Connections, we do have the ability to integrate the. I was going to say yammer, but engage. I just said, yeah, but. So to have that in there, but yeah, yeah, yeah, absolutely.
Manesh Mistry [00:36:43]:
I'm having that inside of an intranet, making it personal. But also what I've noticed as well, it's not just about that desktop, it's also about making it available on mobile, especially for the organizations I work with, it's frontline workers, it's people who don't work at desks and they're going to be using just a mobile device. How can we make it more relevant to them as well?
David Bowman [00:37:04]:
Yeah, and you know, I really like that kind of progression that we've seen in tech here of, you know, firstline workers primarily relying on kind of human interaction, organic conversations, getting information, being able to go to policy libraries in SharePoint, run searches there, download documents and read them. To now of being able to open a Teams mobile app and have a conversation with the library, getting a question, you know, much, much quicker than you would have done if you'd had to search for it and then go, you know, read the document, looking for an answer yourself. Yeah, but, you know, that kind of, you know, really nice progression of efficiency.
Manesh Mistry [00:37:36]:
And on top of that, We've also have SharePoint agents, which to me and to others in Microsoft, the way that we're thinking about it, these are documents, you treat them as documents which you share and that you share a SharePoint agent. And this agent has maybe one task. So maybe I'm making this complicated for myself, but one of the things I've done with SharePoint agents is to create a waiting system. And when I say waiting, not waiting for a train, but waiting as in wait and starting to think about if I had again some case notes or transcript and I had specific criteria, I could put more weight on some criteria against another. So I can start to prioritize where I should be looking first inside of a document or if I was looking at some data, where should I be looking in that data? First by putting that weighting on the criteria and then putting that criteria against the data, but being able to use SharePoint agents to be able to do that. So yeah, it's great. But I think it's these use cases. In Microsoft, as part of our adoption.Microsoft.com site for Copilot, there is a use case library where we do put simple prompts.
Manesh Mistry [00:38:46]:
But I think it's important to start to understand the use cases generally what available. So what's in finance, in procurement, inside of it. But I think as part of the adoption process, each organization should have their own and start to think about the Personas, think about their use cases and then sort of collectively start to put the prompts together which is going to support every area of the business.
Jarbas Horst [00:39:09]:
We have like this scenario. We're talking about frontline workers right now. And the other day we talked about health care. I think you had us like a nice scenario that maybe you can tell us a bit more about that.
Manesh Mistry [00:39:19]:
So the healthcare one, I've got a couple here which I've been looking at. So one of them, both of them are actually, both of them are from my other colleagues. So this is not my work. But again, it's just amazing the speed of understanding and innovation which is happening in our Teams. But there's one around for healthcare, around standard operating procedures and being able to make amendments to those SOPs, and then using Copilot to produce a newer version of those standard operating procedures for ways which maybe nurses or doctors or clinicians, how they approach particular procedures, the updates that they need to make. So using Copilot to support the update of those procedures is one which I've seen happening. The other one in healthcare is around an HR agent sort of agent. Bottom.
Manesh Mistry [00:40:13]:
That's like saying agent as an agent. Agent.
David Bowman [00:40:16]:
So there's a lot of terminology here.
Jarbas Horst [00:40:19]:
Oh, yeah.
Manesh Mistry [00:40:20]:
I'm trying my best, David. Honestly, I am trying. If you know what, you do this for 18 months and you just, you do this and you think, yeah, everybody else should know about this, but you don't. But it's not that like the terminologies.
Jarbas Horst [00:40:32]:
Have been like the same throughout kind of those 18 months. Right. I don't want to touch that space. It might be a kind of a sensitive topic. Yeah, let's skip it.
Manesh Mistry [00:40:42]:
No, no, look, you know, it's fine. Yeah, we've been through different changes, but I think this reflects the speed of how the services and Copilot have come on board. It's. It's like, it's like small. You know, there were big movements before in terms of branding, but I think we're very, very. It's all small, little, small nudges now. But we've noticed that these little nudges maybe from going from Bing Chat Enterprise to Microsoft Copilot or now to Microsoft 365 Copilot chat, small changes.
David Bowman [00:41:10]:
Hindsight is one of the most powerful tools in the product manager's kit bag.
Manesh Mistry [00:41:13]:
It is. I Wouldn't want to be part of that branding thing. But sometimes we wonder how they do it because, you know, it's a minefield in terms of it's tough the different directions. And as like Microsoft, how do you not step on somebody else's product's toes because you've got the same acronym as another one. So like I said for sip. So SIP in Teams terms and communications means one thing, but it could also mean something about looking at procurement in another. So it's.
David Bowman [00:41:40]:
We were talking about this on a previous episode that it's easy to look at the decisions and outputs and outcomes and naming changes and go, oh, you know, why don't they get this right the first time? But when you think about just the sheer engineering challenge that Microsoft got with, you know, Microsoft 365 as a total platform, it's huge. It covers so many things and it is involving technologies that are moving so quickly now. So it's tough to get this stuff right first time.
Jarbas Horst [00:42:03]:
And we have a very large organization also. Right. They need to align all of the Teams like you know, to bring Copilot across all of the different Microsoft Transit services and Azure and so on.
David Bowman [00:42:13]:
It's hard enough doing that in an organization of a thousand people, let alone an organization size Microsoft, my goodness, what.
Manesh Mistry [00:42:20]:
Did it mean to just use that word Copilot? And for it to become synonymous in every application and place that Microsoft lives is mind blowing. Where did we get to? Sorry, I missed the.
David Bowman [00:42:32]:
We were talking about agents. You know, I think what would be useful for us to talk just a little touch on briefly is, you know, We've talked about SharePoint agents.
Manesh Mistry [00:42:38]:
Yes.
David Bowman [00:42:38]:
We've talked about kind of customizations and extensibility. There are some options, aren't there, for people that want to build agents? We've got something in SharePoint. We've got this thing called the Copilot Studio and there are, you know, obviously kind of pro code approaches to building these things as well. Let's just do a kind of quick summary on what the options are.
Manesh Mistry [00:42:56]:
Yeah, Jack. So. Well, unless you want to do it, I mean, I'm happy for you to go through. I'll try.
David Bowman [00:43:01]:
I am the wrong person to be.
Manesh Mistry [00:43:03]:
Talking about any of this stuff.
Jarbas Horst [00:43:04]:
I can talk a bit about like the pro code version. Right. So it is what we are using quite a lot here in the organization. We have been, of course, like looking how can we extend Copilot, how we can integrate with Copilot. Right. Of course, we believe in the direction where Microsoft is heading to like we've Copilot as the UI for AI and I think the agents then being the interface where people like will be then accessing data from different sources, Internet being like one of those sources as well. And agents provide declarative agents in Copilot. So that's the direction with the pro code that we have been experiencing.
Jarbas Horst [00:43:38]:
And that gives you unlocks like a lot of potential and gives you more flexibility in terms of extending, integrating. With Copilot you have Copilot orchestrating the conversation between kind of the user and the logic that happens behind the scenes. So Copilot is doing all of that magic and then you can create your code solving kind of more complex scenarios that might be about kind of retrieving data, giving that back to the user, maybe using then the adaptive cards like to visualize the information, but then also maybe then having the agents performing tasks on your behalf and you can achieve all of that and via like the pro code version quite nicely because you have kind of the flexibility on how you implement all of that.
David Bowman [00:44:18]:
These are like the most complex scenarios that we're catering for kind of ISV vendor product solutions that people are building for repeatability. Right. Install it in multiple customers.
Manesh Mistry [00:44:28]:
So for me I see that as the Copilot Studio side of it. This is where you have those, the full capability there to do the pro code approach to be able to interface with other applications inside of there. So absolutely that's the Copilot Studio side. But from a end user perspective, if you have that Microsoft 365 Copilot license, gosh, that's a mouthful, isn't it? But having that gives you access to the Agent Builder which is a, it's a sort of a cut down version of the complexity which is inside of Copilot Studio, but brings that into the Copilot interface instead of going into Copilot Studio. But it's the Agent Builder which will guide you through a conversation. So hey, how are you doing? It doesn't say that, but being able to say, well what do you want this agent to do? And you do your description and then say what kind of knowledge sources are you going to be connecting into? So it could be an official organization, third party resource application that you would connect to a data source or it could be a SharePoint site, OneDrive, shared library, even just one document in the cases that I've done. So in this conversation you would say what it does, the information that it's going to connect to and it waves its magic wand. I'll keep the Magic thing going.
Manesh Mistry [00:45:46]:
And what it does is it will produce the agent in that same screen. So you can see I've named this agent this here. Do you agree with it? And it normally comes up with a pretty good title and then it gives you some starter prompts. So what kind of starters can we give to a person who's going to be using this agent and then to test it. And once you've tested it, you create it and then you make amendments and you just keep republishing and then you share that. So this is that front end, no code approach to this. And then when it comes to SharePoint, SharePoint is very similar to that. But yet you then start to go into that low code and pro code approach inside of Copilot Studio.
Manesh Mistry [00:46:26]:
So yeah, those are I think the different ways of getting in loads of.
David Bowman [00:46:30]:
Options ranging from kind of config in Copilot in the browser through to kind of, you know, full on customization involving developers and testing and creating product.
Manesh Mistry [00:46:41]:
Yeah, I'd like to think as well that even though we get developers involved, we don't forget the why we're doing it.
David Bowman [00:46:48]:
Yeah.
Manesh Mistry [00:46:49]:
Okay. That is, and this is where I'm even for my colleagues who are putting these agents together, I'm being the annoying person and just asking the why. What's the business value that are going to be achieved by creating the agent? So it's not just about tech for tech's sake. It's got to have purpose to it.
David Bowman [00:47:07]:
Yeah. And not ending up with a proliferation of agents all over the organizations that are all performing these very sort of niche tasks that actually could be achieved fairly easily with Copilot out of the box.
Manesh Mistry [00:47:18]:
Yeah. So what we've done is we've really, from an administrative perspective is really start to give administrators the flexibility as to who can create these agents inside there as well and who can access them. And then remember, we still have that same information governance and the data protection behind it. So again, we need to make sure that's in place as well.
David Bowman [00:47:39]:
Yeah. Thinking about the future then and you know, not expecting you to make a sort of roadmap commitment or you know, breach NDAs or employment conditions. Manesh, in your personal opinion, what's next? What are you excited about? You know, you're very close to this technology. Where do you think this is going in the next sort of 12 months.
Manesh Mistry [00:47:58]:
12 months. Well, I can tell you in the next few months what I'm really looking forward to is the extension of agents, but pre prepared agents. So to give you an Example one called facilitator. So that's being a facilitator, this is Copilot inside of the meeting or the group chat or the chat itself. And then being able to understand based on context what are the topics being discussed inside of that meeting to see the action items which are being tracked as well inside up meeting. So using facilitator, using that agent to be able to do that. So this is one which is in the process of being released. But it's really interesting to see from a internal perspective to see how facilitator is being added into the conversations and some of the questions which are being asked now.
Manesh Mistry [00:48:48]:
Sorry, the unanswered questions based on the conversation is fascinating to see how facilitator works there inside.
David Bowman [00:48:55]:
Is it going to highlight that I've done all the talking in a meeting?
Manesh Mistry [00:48:59]:
No, it won't do that. I can't do with that prompt sometimes. Yes, David really did talk a bit too much and didn't let David is talking too much. No, it doesn't do that. It tries to be as democratic as possible in terms of its approach. No, let's think about responsible AI here and the applications and the features that are put together. We still need to make sure that we're following that on there. The other one is on Copilot actions.
Manesh Mistry [00:49:28]:
So being able to create scheduled prompts which can support what you do on a day to day basis. But these Copilot actions can also send messages to other people on your behalf. So again, this is really starting to bring in that agent backend process capability and then pushing for responses to asks to other people and then for you to get the responses and to see them in one place, adding automation to that.
Jarbas Horst [00:49:56]:
So it's actually a preview currently like this feature that you mentioned, if I'm not mistaken.
Manesh Mistry [00:50:01]:
Yeah, the rollouts of these things can sometimes take time to do. So it's the same with the SharePoint agents. Some places have got it and some places haven't. But this year it's this agentic process. It's understanding business processes and looking at the back end of the services and seeing how those business processes can be supported using generative AI. That's some good stuff coming on that side. And I know it doesn't sound like much, but being able to reference other documentation inside of Excel and even referencing Excel, referencing third party Excel data in documentation, the creation of graphs inside of the Copilot chat service. So being able to take some data and produce a graph right there on the screen.
Manesh Mistry [00:50:46]:
So one of My colleagues has done this. He's looked at the working hours of several nurses and doctors and as part of the agent that he produced, he got it that. So he was able to grab the data and then for the agent to produce the graph inside of the conversation and then to start tweaking the graph by saying, well, remove the doctors, keep the nurses or show me those who are just working too many hours. So it was able to do that. So having visual representation inside of charts inside of there to be visually represented.
David Bowman [00:51:19]:
You say, you know, it sounds like it's not much. These are the kinds of things that make a big difference for people. Right. You know, those kind of individual use cases of being able to achieve something like that that's, you know, super powerful. Yarbos predictions, things that you're excited about in this space.
Jarbas Horst [00:51:32]:
I'm also like looking at the agents. And one thing that you didn't mention, Manesh, is when it comes to agents talking to other agents.
Manesh Mistry [00:51:39]:
Yes.
Jarbas Horst [00:51:39]:
I think this will also be quite interesting when you can have maybe a collection of different agents, each focused on one specific task. But when that agent hits its limitation, it maybe then reached out to the other agent to complete the task. And I think this is what I want to see coming this year, this possibility to have this interaction between the agents. What do you think, David, when it comes to the Internet space?
David Bowman [00:52:06]:
Yeah, well, you know, we're spending a lot of time thinking about how does Copilot, how do agents affect intranet? And you know, I guess the kind of current position for us on this is it will be another channel, right, for content, for users to consume intranet content through. You know, we were doing a session on predictions for this year and you know, we don't see Copilot replacing intranets anytime soon. But you know, I think what is more likely to happen is employees, people consuming content that is in the intranet through Copilot. And you know, I think there's scope to have a series of agents that represents kind of common scenarios, use cases in the intranet, you know, a news agent, a people agent, because, you know, you can define the edges of the intranet within the agent. Right. The more context that you can give this thing, the more likely it is that people are going to get the outcomes that they are expecting. So I think that brings us to the end of our session today. Manesh, really appreciate you coming and joining us, taking the time to talk to us about COVID 19 Copilot today.
Manesh Mistry [00:53:06]:
It's always fun. I love, I don't know why? But, yeah, I really love talking about Copilot. Very passionate about it, and hopefully that comes across because I want other people to get excited about it as well.
David Bowman [00:53:17]:
I think that I'm more excited about it now than I was an hour ago. So mission accomplished, I think, because it.
Jarbas Horst [00:53:23]:
Connects also with your creativity aspect, Manesh, isn't it like, you're looking at all of these paintings here behind you, so it's something that you do, right?
David Bowman [00:53:30]:
It's.
Manesh Mistry [00:53:31]:
Oh, it's such a strange one. It's. You've got to be creative outside of work. There's a balance that you've got to have. So having that creativity there. So for me, that manifests itself as doing painting. So this is what I've been doing for the last seven years, by the way. I've painted seascapes, landscapes.
Manesh Mistry [00:53:47]:
Behind me you can see visual representations of machine learning, artificial general intelligence and also transformation. So this is what I've been doing. I've thought about this, but then I needed to visually represent that in paintings. So that's been my approach. But I noticed over the last 18 months that my painting output has reduced, which is bad. It shouldn't be. But then I'm thinking, well, why is that the case? It's because my creativity has gone into prompting. It's gone into just, what can generative AI do for us? What the outcomes that we can achieve from it? And that requires creativity.
Manesh Mistry [00:54:28]:
So if you have creativity and you have logic together, you will love, love, love Copilot, because it brings all of those different elements together.
David Bowman [00:54:38]:
Well, that's a great close. That's a great close. For today's episode. Manesh Yarbus, appreciate the time. Speak to you soon.
Jarbas Horst [00:54:45]:
Thank you very much, Manesh. Thanks, David.
Manesh Mistry [00:54:47]:
Bye for now.
Jarbas Horst [00:54:48]:
Cheers.