Latest / Microsoft Cloud IT Pro Podcast / Episode 400 – Microsoft 365 Copilot Declarative Agents and Copilot Studio Agent Builder
Transcript
- 0:03Welcome to our four hundredth episode of the
- 0:06Microsoft Cloud IT Pro podcast
- 0:09recorded live 04/21/2025.
- 0:12This is a show about Microsoft three sixty
- 0:14five and Azure from the perspective of IT
- 0:17pros and end users, where we discuss a
- 0:19topic or recent news and how it relates
- 0:21to you.
- 0:22This week, after eight years and 400 episodes,
- 0:26Ben and Scott thought about doing something special
- 0:28to celebrate this four hundredth episode, but instead
- 0:31just decided to talk more about Copilot.
- 0:34This time, discussing getting started with declarative agents
- 0:38in Copilot Studio Agent Builder. They discuss everything
- 0:41from getting started building agents to how you
- 0:44might use them to licensing and costs.
- 0:47Hope you enjoy episode number 400.
- 0:52How was your weekend, Scott?
- 0:54My weekend? Oh, it was
- 0:56a a a quiet stay at home weekend.
- 0:58I've still got a little bit of my
- 1:00radio voice here as I get over my
- 1:02my first spring cold or whatever is going
- 1:05around. I've been getting over one for, like,
- 1:07two weeks. So since I don't have my
- 1:09notes up yet and they do not appear
- 1:12to want to come up, what are we
- 1:14gonna talk about today, Scott? What are you
- 1:16gonna talk about today? Well, I try to
- 1:18get my notes to load. What am I
- 1:19gonna talk about today? Sure. Let's
- 1:22let's have the guy who has no voice
- 1:24be the one who talks a bunch of
- 1:27well, today, we are going to
- 1:29talk
- 1:30about
- 1:33no code declarative agents
- 1:35and Copilot Studio Agent Builder,
- 1:39if that's not a mouthful for you. But
- 1:41effectively,
- 1:42the little Copilot agents
- 1:44that you can take and build for yourself
- 1:47inside of Copilot
- 1:49in a nice little
- 1:51next next next GUI. Perfect.
- 1:54Sounds good. And now my notes are coming
- 1:56back to me. I'm well, in my head.
- 1:58In my head, I have notes. So this
- 2:00is something I've started playing with a little
- 2:03bit. I've actually had clients start asking about
- 2:05this too. So
- 2:06this could be an interesting topic because even
- 2:09I spent a bunch of time even deciphering
- 2:10licensing around this. Like, even licensing around this
- 2:13seems goofy to me. So I wanna build
- 2:16an agent with no code. Where should I
- 2:18start? Should we start with licensing, or should
- 2:21we start with just how to start building
- 2:23it? Why don't we start with how to
- 2:25start building it, and then we can get
- 2:28into
- 2:29a little bit about licensing
- 2:31and things like that as as we go
- 2:33along. Sounds
- 2:35good.
- 2:36So starting to build it, there's a website.
- 2:38So there's a couple ways, I guess, you
- 2:39could start building this, can't you? There is
- 2:43in Teams now or in Copilot, I guess
- 2:46it's actually in the Copilot app, you have
- 2:49agents
- 2:50in the upper
- 2:51right hand corner, and it looks like there's
- 2:53already a few agents in there, potentially,
- 2:56depending on what you set up, but you
- 2:59also have a button up there to just
- 3:00click on build an agent. It turns out
- 3:03that you've probably been interacting with these in
- 3:05some way, shape, or form, and you just
- 3:08might not have noticed. So
- 3:10there's things like a prompt coach. So you
- 3:13you wanna get better at writing prompts, you
- 3:15can go down that path and work with
- 3:17those.
- 3:17There's a generic writing coach.
- 3:20There's all sorts of just
- 3:22built in agents
- 3:24that are
- 3:25ready to ready to go for you should
- 3:27you want to use them. And then you
- 3:30can build out your own agents on top
- 3:31of those and leverage the connectors within
- 3:35your organization.
- 3:37So
- 3:38why would you want to
- 3:40build your own agent? Well, like, there there's
- 3:42a whole bunch of things that you can
- 3:44do
- 3:45within even, like, our world and and our
- 3:47roles and and the things that happen. So
- 3:49for you, I could think about maybe,
- 3:52hey. I would like to build the
- 3:55the the the project knowledge bot for IntelliJunk,
- 3:58and you wanna point it at maybe your
- 4:01document
- 4:02repository
- 4:03for where you've done things like your SAUs
- 4:06or you have artifacts from deliveries, things like
- 4:08that. Like, teach me about past things and
- 4:11help me make this next one better. You
- 4:13could also build probably a pretty good, like,
- 4:15SAU bot, like a statement of work agent
- 4:18based on statements work and things like that
- 4:20that are out there. I've been taking these
- 4:23and building these.
- 4:24So part of my role involves working with
- 4:27products that are released on GitHub,
- 4:29and all the code lives on GitHub, the
- 4:31releases, the wikis for these things. So I've
- 4:35been using them within my organization where the
- 4:37GitHub connector is turned on to point directly
- 4:39at Git repos
- 4:41and get insights on
- 4:43not only the
- 4:44the repos themselves, but also the the code
- 4:47within those repos.
- 4:48So I'll give you an example. We have
- 4:50a
- 4:51product that gets distributed to customers. It gets
- 4:54compiled and and pushed out. That product has
- 4:56a bunch of
- 4:58hidden flags inside of it. And those hidden
- 5:01flags are they're they're sitting there in the
- 5:03code if you know where they are, but
- 5:04I don't always remember, like, which file to
- 5:06go into at which time to look up
- 5:08which code.
- 5:09So
- 5:10enter my little my little product helper agent
- 5:14that can tell me those things. And then
- 5:16it can also maybe be augmented with documentation
- 5:19from internal wikis, SharePoint,
- 5:21things like that as well. So
- 5:24these become powerful little helpers
- 5:28against your knowledge bases. So you can effectively
- 5:30think about it as a or the way
- 5:33I think about it is a very quick
- 5:35onboarding,
- 5:36no code flow
- 5:38to be able to
- 5:41leverage
- 5:42RAG, retrieval augmented generation.
- 5:44So RAG with my own documentation, my own
- 5:47code basis, things like that, to ground these
- 5:50models
- 5:51and further refine the responses before they spit
- 5:54some stuff back over the fence to me.
- 5:55Me. Yeah. And I think even the part
- 5:57you brought up there about, like, looking at
- 5:59GitHub is
- 6:01it's also a way to
- 6:03bring in some of that additional
- 6:06context
- 6:07or
- 6:08refine these Copilot agents or build these agents
- 6:11to look at other data sources. Like, as
- 6:13you go into these Copilot agents, it's not
- 6:15just I wanna go look at this specific
- 6:19SharePoint site or
- 6:21look at
- 6:22this specific,
- 6:23I guess, context for the Microsoft three sixty
- 6:26five. It's go pull in data from
- 6:28GitHub. Let's go look
- 6:30at Confluence wikis. Let's go look at Trello
- 6:33boards. Let's go look at a SQL database.
- 6:35There's lots of other date or this allows
- 6:38you to build these agents that are not
- 6:41only just specific, but also pull in data
- 6:44from a lot of other data sources
- 6:46to,
- 6:47help with these Copilot responses, these
- 6:50conversations
- 6:50you have with Copilot.
- 6:52Yes.
- 6:53So
- 6:54I I I think, like, there there's a
- 6:55world where
- 6:57some of this stuff can get more complicated,
- 6:58and you can go down a path of
- 7:01needing
- 7:03to actually build an agent
- 7:05and do some code around it. And then
- 7:07there is the
- 7:08Copilot agent studio
- 7:10experience, which is
- 7:12zero code, super easy to onboard. So you're
- 7:15kinda getting out of the world of those
- 7:18more traditional
- 7:19language bots
- 7:20that are gonna require some measure of I
- 7:23don't wanna say, like, extensive coding, but they're
- 7:25gonna require some knowledge
- 7:27and things to get in there versus these
- 7:29things, which are
- 7:32declarative. Like, you just go in and and
- 7:34and declare what you want
- 7:36via the Copilot agent studio, and then it
- 7:39effectively
- 7:40ends up in this what you define is
- 7:42what you get kind kind of
- 7:45bot that that's out there. The other cool
- 7:47thing is because
- 7:49it's all built on top of m three
- 7:51sixty five Copilot,
- 7:52if you think about things that you might
- 7:54want
- 7:55in
- 7:56a workflow
- 7:58so let's take, like, kinda like one of
- 7:59the easier examples. So employee onboarding. So when
- 8:02I do a new hire for an employee,
- 8:05there's typically a slew of documents that employees
- 8:07need to sign when they come in, like
- 8:09get this going, get this going, get this
- 8:11going kinds of things.
- 8:13There's probably also, like, security groups they need
- 8:16to sign up for. There might be
- 8:18meeting series that they need to be a
- 8:20part of, like an all hands, things like
- 8:23that. So you could build an agent that
- 8:25has those assets inside of it,
- 8:27like, point it to your HR onboarding documents,
- 8:31have a baked prompt in there, a system
- 8:33prompt that you've already worked on, and and
- 8:35you've kind of gotten to where it needs
- 8:37to be. And you could do that across
- 8:40a series of system prompts that are baked
- 8:41into that agent and augmented with the data
- 8:44that you need for your onboarding process, and
- 8:46you could just make that employee onboarding super
- 8:49easy to do right there.
- 8:51And then if there's even things that you
- 8:52wanna do outside
- 8:54of just
- 8:56retrieval augmented generation, like, with documents that are
- 8:59sitting, like, in SharePoint or, like, my example
- 9:02would, like, GitHub, things like that, Because this
- 9:04is m three sixty five Copilot, and it's
- 9:06connected to the rest of your data sources,
- 9:09like,
- 9:10if if I'm chatting with that agent, that's
- 9:12my team's chats, that's
- 9:15my mail sitting in my mailbox kinda thing.
- 9:17You can also bake those system prompts so
- 9:19that they further go out and spider and
- 9:22get information from those sources as well. So
- 9:25not just go get the HR doc, but
- 9:28or go get the blah blah blah project
- 9:30start doc, but maybe go back through my
- 9:32mail and look at the last five projects
- 9:34I did and help me write up my,
- 9:37introduction for this one kinda thing. Absolutely. And
- 9:39as you're going to build these, like, there's
- 9:42a few different steps once you go in
- 9:43here to build these agents. So we've kinda
- 9:46talked about, like, what you can do with
- 9:48them. I think the fascinating part is when
- 9:50you go in to build these agents, like,
- 9:52you click on that. Like you said, you
- 9:54go to the the agent builder, and the
- 9:57first thing you do is you actually use
- 10:00Copilot to try to build the agent. Like,
- 10:02it'll ask you,
- 10:03what do you wanna do with this agent?
- 10:05What are you trying to accomplish with this?
- 10:07And you can just type in what you
- 10:09wanna build to get started building these agents.
- 10:12So you'll actually see the system agents in
- 10:15there, So, like, that prompt coach, things like
- 10:17that. So you can get a sense not
- 10:19for, like, how those are built if you're
- 10:21looking for a little bit of information there.
- 10:23The the docs are pretty good here as
- 10:25well about kinda giving grounding and and what
- 10:28each of the options are that are available
- 10:30to you. But it is kind of just
- 10:32as easy, like, licensing aside.
- 10:35Once you've got the ability to create these
- 10:37things, it's literally just go to Copilot, and
- 10:39that could be Copilot the app, the m
- 10:42m three sixty five Copilot, whatever the business
- 10:44one is. Like, it's pinned to your desktop
- 10:46automatically in most cases,
- 10:48you've got that one, or you can just
- 10:50go to office.com/shop.
- 10:52And
- 10:53right up there in the upper right, like
- 10:54you said, you've got your list of kinda
- 10:56system agents, and then you can just go
- 10:58ahead and build a new agent right on
- 11:01top of that. So you just go ahead
- 11:03and click create agent, and then you're kinda
- 11:07off to the races
- 11:08from there. So it it's it's meant to
- 11:10be
- 11:11super
- 11:12quick. Like so so that agent builder is
- 11:15there to kinda show you the template, show
- 11:17you what can be done, get you through
- 11:18that creation process
- 11:20pretty quickly,
- 11:22and then get you to publish and then
- 11:25maybe come back and and make updates and
- 11:27refine and things like that as you go.
- 11:29Yeah. And those different things. So once you
- 11:31do that, you get your agent built or
- 11:33as a process of building it, if you
- 11:35wanna add different things,
- 11:36it is. It's going in, giving it a
- 11:38name. Like you said, this could be the
- 11:40internal knowledge base. This could be
- 11:43employee onboarding,
- 11:44give it a quick description, and then you
- 11:46can give instructions
- 11:48for your agent too. So this is, like
- 11:50you said, how you direct the behavior of
- 11:52this agent, different tasks, how it's supposed to
- 11:55complete different tasks,
- 11:57but all those prompts around
- 12:00how it builds agents,
- 12:01and then it gets into that other thing
- 12:04we said, the knowledge.
- 12:05Where is this agent gonna start looking for
- 12:07stuff? Is it gonna look in
- 12:10SharePoint? Is it gonna look at a public
- 12:13website? Is it gonna look at a SQL
- 12:14database?
- 12:15Adding all those different sources
- 12:18into your agent,
- 12:20and then even some starter prompts. So you
- 12:22can go in and add some number of
- 12:24starter prompts for your
- 12:26different users so that
- 12:28you can help them get started with the
- 12:30agent, help them figure it out. But it
- 12:32really is. It's just that GUI
- 12:34work through the
- 12:35different
- 12:36the different steps in it to really refine
- 12:38that knowledge, instructions,
- 12:40even capabilities. Is it gonna have a code
- 12:43interpreter, image generator as a part of it,
- 12:45and then what those starter prompts are? And
- 12:47then you just hit publish, and it'll be
- 12:49there to start using, and then you can
- 12:51go in and tweak it, refine it, configure
- 12:53it as you go. Some of that stuff's
- 12:55relatively
- 12:57new or it kinda, like,
- 12:59evergreen SaaS product. Right? It's always changing underneath
- 13:02you kinda thing. So some of those connectors
- 13:05that are available like, I noticed, like, in
- 13:06my organization,
- 13:08like, on day one, the number of
- 13:10pre enabled
- 13:12Microsoft Graph connectors
- 13:14was pretty
- 13:15low, like, low to nonexistence.
- 13:18And now there's a whole bunch of them
- 13:20in there, things like connecting to ADO, connecting
- 13:23to
- 13:24the ServiceNow, connecting to SAP. Like, there's a
- 13:27whole bunch of them that have just been
- 13:29added in there, and then some of those
- 13:31capabilities you talked about. So
- 13:33code interpreter is an interesting one.
- 13:36It uses
- 13:37Python
- 13:39in the background
- 13:41to analyze data that's being input by users.
- 13:44So maybe if you're, like, trying to build
- 13:46an agent where you're trying to have the
- 13:48agent help you solve a math problem, things
- 13:50like that, That's where the the code interpreter
- 13:53comes into play. And then the image generator,
- 13:56just image generation one zero one, and I
- 13:58think the things that folks are probably used
- 14:01to from from seeing most of these systems.
- 14:03You do have to keep in the back
- 14:04of your head that this is all built
- 14:06on m three sixty five Copilot,
- 14:08so it's going to kinda behave like m
- 14:12three sixty five Copilot.
- 14:13And the reason I call that out is,
- 14:15like, image generation and, like,
- 14:17just chat GPT
- 14:18versus, I don't know, maybe going out and
- 14:20using, like, stable diffusion OSS versus
- 14:23mid journey versus Copilot are all very different.
- 14:26And they all have their kinda baked in
- 14:28system prompts along the way and and their
- 14:30own flavors, so you can't override some of
- 14:33that stuff. Like, I don't know, maybe there's,
- 14:35like, a prompt injection thing or something. But
- 14:37for the most part, like, yeah, I'm I'm
- 14:39not hacking away at these things like that.
- 14:41So you can't override them. Right? You kinda
- 14:43live with the outputs
- 14:44in some cases that those systems give you.
- 14:46Yeah. And I noticed even with setting up
- 14:49these connectors to your point of different ones
- 14:51showing up, even setting them up, I was
- 14:55working through some of the documentation the other
- 14:57day, and they're even set up a little
- 15:01it's interesting how you set them up. Well,
- 15:02how about that?
- 15:04It's not necessarily straightforward as I thought. Like,
- 15:07I can go into mine right now when
- 15:09I see
- 15:11just one of those knowledge sources. And I
- 15:13was going in and trying to figure out
- 15:14how to add more, and you can, like,
- 15:16click to add more knowledge sources and you
- 15:18see them there, but you can't really add
- 15:20them. It's like you have to go connect
- 15:22to them first. And you actually have to
- 15:24go into, like, your Microsoft three sixty five
- 15:27admin and go into the search settings
- 15:30in Microsoft three sixty five admin and go
- 15:33configure these additional
- 15:35sources
- 15:36in your Microsoft three sixty five search setting,
- 15:40and that's where you set up these connectors
- 15:42that can then be used in Copilot Studio
- 15:45to add those additional
- 15:47or that additional context, those additional knowledge sources
- 15:49to Copilot Studio. So there is a little
- 15:52bit of that. I mean, that's how it
- 15:54works under the covers. Right? Like, if you're
- 15:55gonna exclude something from Copilot using sensitivity labels,
- 15:59you also exclude it from Microsoft Search. End
- 16:01of the day, from all this different knowledge
- 16:04that gets pulled into Copilot, it's very heavily
- 16:06using the Microsoft three sixty five search under
- 16:09the covers for everything from connectors to pulling
- 16:12your Microsoft three sixty five data. Yes. Search
- 16:15plus. So so it depends. Sometimes it's using
- 16:17the
- 16:18existing indices that are already there. Like like,
- 16:21hey. They're all ready to go. In the
- 16:23other cases,
- 16:24it's
- 16:25let's take, like, a graph connector, for example.
- 16:28So you go and configure a graph connector
- 16:31for
- 16:32for a ADO kinda thing. Part of that
- 16:34experience of turning that on is actually not
- 16:37only lighting up the connector, but, like you
- 16:40said, lighting it up to search, and then
- 16:41not only lighting it up to search, like,
- 16:44just from the top level of saying, like,
- 16:45oh, it's enabled within m three sixty five
- 16:47search. It's also going so far as to
- 16:50do things like
- 16:52enabling your connectors for inline results.
- 16:55So you're gonna go through, like, your search
- 16:57verticals and make sure that they're enabled
- 17:00up up where they need to be that
- 17:01way. So, like, it is, like, embedded in
- 17:04there. So I think you do have to
- 17:05think about it a little holistically if, like,
- 17:07before you turn this stuff on. Like, you're
- 17:09not just to your point exposing it to,
- 17:12Copilot. You're exposing it to search,
- 17:15and you're kinda built in search experience. But
- 17:18then it's still security trimmed, all that kind
- 17:21of
- 17:23stuff
- 17:25around me. Do you feel overwhelmed by trying
- 17:27to manage your Office three sixty five environment?
- 17:30Are you facing unexpected issues that disrupt your
- 17:33company's productivity? Intelligink is here to help. Much
- 17:35like you take your car to the mechanic
- 17:37that has specialized knowledge on how to best
- 17:39keep your car running, Intelligink helps you with
- 17:42your Microsoft cloud environment because that's their expertise.
- 17:45Intelligink keeps up with the latest updates in
- 17:48the Microsoft cloud to help keep your business
- 17:50running smoothly and ahead of the curve. Whether
- 17:52you are a small organization with just a
- 17:54few users up to an organization of several
- 17:57thousand employees,
- 17:58they want to partner with you to implement
- 18:00and administer your Microsoft Cloud technology.
- 18:03Visit them at inteligync.com/podcast.
- 18:07That's intelligink.com/podcast
- 18:14for more information or to schedule a thirty
- 18:16minute call to get started with them today.
- 18:19Remember, Intelligink
- 18:20focuses on the Microsoft cloud so you can
- 18:23focus on your business.
- 18:26I would say the other interesting thing I
- 18:28was talking to a client about this the
- 18:29other day, and they got all excited when
- 18:30I mentioned public websites as a data source.
- 18:33This is not just any public website. I
- 18:35cannot go point this to, like, learn.microsoft.com
- 18:38and have that be a knowledge source source
- 18:40for my bot. Like, you have to confirm
- 18:42this is a website you own, DNS records,
- 18:46all that good stuff with public websites. That
- 18:48was a funny example for me because I
- 18:50can build declarative agents. You can burn learn.
- 18:53Okay. Yeah.
- 18:55Look at me. I'm special. I can use
- 18:57learn.Microsoft
- 18:57for my agents. Yeah. There we go.
- 19:00No. I'm not coming up with another example.
- 19:02But I also know that because I've I've
- 19:04played with this even with search, like, pointing
- 19:06the Microsoft three sixty five search to my
- 19:08own website is there's also limits on search
- 19:11and how much you can crawl with search
- 19:12before you have to start paying. So we
- 19:14haven't talked about licensing Copilot Studio yet, but
- 19:17I think depending on these agents and how
- 19:19they interact with search, there could potentially
- 19:22be some additional
- 19:24cost considerations
- 19:25there because
- 19:27you start running into limits on what search
- 19:29can actually
- 19:30go
- 19:31pull access based on licensing, etcetera, from things
- 19:34like public websites. So it does start to
- 19:37kinda run away a little bit, like, once
- 19:39you start to light up all the
- 19:42all all all the little options that are
- 19:44available for you. So
- 19:46graph connectors,
- 19:47they do require a license or some kind
- 19:50of meter usage.
- 19:52SharePoint, obvious, you gotta have a license to
- 19:54SharePoint, OneDrive, things like that.
- 19:57Web search, funny enough, does not require a
- 19:59license,
- 20:00but scoped web search does.
- 20:02So you gotta kinda think through maybe a
- 20:05little bit of nuance there. And then there's
- 20:07some things that are outside the agent builder,
- 20:11but they're available through
- 20:13Teams and and building declarative agents in Teams.
- 20:16So things like a Dataverse connector for your
- 20:19dynamic CRM data,
- 20:21email people, Teams messages. That that stuff's all
- 20:24part of the Teams toolkit,
- 20:25and then that all requires licensing as well.
- 20:28So
- 20:29it it it can be, it can be
- 20:31a thing. So why why don't we talk
- 20:32through some of
- 20:33the other
- 20:34licensing constraints here? Because you mentioned Copilot Studio,
- 20:38and there there is some stuff to do
- 20:40there. It's interesting how they license it. And
- 20:42I don't know that I would say the
- 20:44documentation is clear. Imagine that. Right? Yeah. I'm
- 20:47shocked. Yeah. So there's three different ways
- 20:51based on all the reading through the documentation
- 20:54and looking at this that I did that
- 20:56you can license Copilot Studio to go build
- 20:59these agents. And it's not it's not actually
- 21:01like you're licensing Copilot Studio. It's more like
- 21:04you're licensing the interaction of your users with
- 21:07these agents, but it falls under this Copilot
- 21:10Studio license
- 21:12umbrella. So first, you have your pay as
- 21:14you go option. So I wanna go build
- 21:16an agent. I want my users to be
- 21:18able to use it, but I just want
- 21:20to pay when they use it. And this
- 21:23is where you go
- 21:24set up a PayGo connection,
- 21:27and it is 1¢
- 21:29per message
- 21:30for licensing
- 21:31or leveraging these agents. And I'm gonna come
- 21:34back to messages here in a little bit.
- 21:36Then you can also buy message packs. So
- 21:38instead of paying on a
- 21:41per message
- 21:42basis,
- 21:43you can go buy a message pack for
- 21:45$200
- 21:46per tenant per month. And this essentially gives
- 21:48you pooled messages. And if I remember right,
- 21:51I think it's 10,000
- 21:54messages
- 21:55no. 25,000
- 21:56messages
- 21:57in that $200 plan. So now 200
- 22:00a month, you're buying 25,000
- 22:02messages that you can use within your company.
- 22:05And as users use it, it will deduct
- 22:08from
- 22:08this message pool. And then the other way
- 22:11is you already have Microsoft three sixty five
- 22:13Copilot,
- 22:14the $30 per user per month plan,
- 22:17and then your Copilot Studio is just included
- 22:21within that $30
- 22:22per user per month. It's kinda sort of
- 22:24included. So so some of the limits you're
- 22:27talking about, like, you're talking about message inclusion.
- 22:29Yep. PAYGo is PAYGo. Like you mentioned, the
- 22:31message packs come in 25,000
- 22:34message
- 22:35increments
- 22:37or Copilot Studio with m three sixty five
- 22:39Copilot unlimited.
- 22:41But that's just for the messages component. It's
- 22:43it's a little weird to say
- 22:46that it comes with everything because the the
- 22:48one thing that
- 22:50stands out that Copilot Studio
- 22:53usage rights when you're using Microsoft three sixty
- 22:56five Copilot,
- 22:57oh my gosh. It's a mouthful,
- 22:59is that you are not allowed to create
- 23:02and publish your own agents anywhere.
- 23:04So with that PAYGo and with that message
- 23:06pack model, you can absolutely create and publish.
- 23:09But with the Copilot studio usage rights, you
- 23:12can't publish them anywhere. They're only within your
- 23:15tenant and your users, that that kind of
- 23:17thing. So
- 23:18you you do have to decide, like, what
- 23:20kind of business you're in and and what
- 23:22the ultimate, like, intention is for these things
- 23:24along the way. There's some other things as
- 23:26well that change around, like, the channels you
- 23:28can publish in and and stuff like that.
- 23:31Generally speaking, like, that Copilot
- 23:33Studio usage rates, because you're already a
- 23:37Microsoft three sixty five Copilot
- 23:39subscriber,
- 23:40that's going to give you, like, your most
- 23:42bang for your buck across all of these
- 23:45things that are out there. Well and I
- 23:46don't think so you mentioned building them in
- 23:48Copilot Studio. There's a table. We'll link to
- 23:51this in the show notes. It actually says,
- 23:53like, create and publish your own agents anywhere.
- 23:56It's like it's weird because Copilot message packs
- 23:59and PAYGo, it's included.
- 24:01Studio, it's not included. But create and publish
- 24:03your own agents and plugins to extend Microsoft
- 24:05three sixty five, it's not included with Studio,
- 24:09PAYGo, or message packs, but it is included
- 24:11with the $30 a month plan. So they,
- 24:14like, flip flop. Well, there's an unwritten word
- 24:17that you have to extend in. So create
- 24:19and publish your own agents and plugins to
- 24:21extend Microsoft three sixty five Copilot
- 24:24in your tenant. Yeah. So it kinda becomes
- 24:26in your tenant versus multi tenant. Versus anywhere.
- 24:29Yeah. So, like, if you were a builder
- 24:31and you wanted to build and deploy an
- 24:33agent across multiple tenants, then you you might
- 24:36be in one of the other kinda
- 24:38one of the other buckets there. So it's
- 24:41it's weird when they talk about extensibility
- 24:43that way, but it's really, like,
- 24:45extensibility
- 24:46within your tenant
- 24:48is is the way that I I think
- 24:50about it in my head. You'll look at
- 24:52it there. Yeah. And then we talked about
- 24:53messages. This is the other thing that you
- 24:56need to be aware of is you may
- 24:57think messages is
- 24:59I send the agent a message. I ask
- 25:02it a question is a message. Kinda like
- 25:04maybe you'd chat in Teams and you'd send
- 25:06someone a message.
- 25:08But different
- 25:09actions
- 25:10in
- 25:11Copilot
- 25:12count for different messages. It's like another one
- 25:14of Microsoft's weird, we're gonna use this to
- 25:17measure your usage without really
- 25:20making it clear
- 25:21how that's what that unit is.
- 25:24So in the billing rate table, and we
- 25:27can throw this out there, like, a classic
- 25:29answer is one message. So maybe that's where
- 25:32this started is I asked Coppola the question,
- 25:34it gives me a classic answer, that's a
- 25:36message. But if now I'm doing a generative
- 25:39answer, it's two messages. If I'm doing an
- 25:42agent action, it's five messages. If I'm sending
- 25:46agent flow actions,
- 25:48it's 13 messages per every 100 actions.
- 25:53Text and generative AI is a message, but
- 25:55text and generative
- 25:56or basic
- 25:57is
- 25:58per 10 responses
- 26:00as one message. But if you're doing text
- 26:02and generative
- 26:03AI tools premium,
- 26:05it's a hundred messages per 10 responses.
- 26:08So this is, I feel like, definitely a
- 26:10little bit more of an art than a
- 26:12science, and you could have some
- 26:15huge fluctuations,
- 26:16I would say, in some of this PAYGo
- 26:18pricing or your message pools
- 26:21based
- 26:23on how you're allowing your users to leverage
- 26:26Copilot within these agents that you build. I
- 26:28mean, Happy Path is certainly, like, you've picked
- 26:31up m three sixty five Copilot like you
- 26:33talked about. Right. The $30 per user per
- 26:35month or or whatever number that eventually ends
- 26:37up being.
- 26:39And and and you've gone down that rabbit
- 26:41hole, but that's happy path. Like, to your
- 26:44point, not everybody's in the happy path. And
- 26:46some of this stuff is a little confusing.
- 26:48Like, really, when they're talking about messages, they're
- 26:50talking about tokens. Like, as much as I
- 26:52hate talking about tokens and context windows with
- 26:54LLMs,
- 26:55at least, like, token makes sense to me.
- 26:57Right? Or you could say, like, fuel up.
- 26:59Like, hey. I'm I'm going to
- 27:01in The US, we have, like, Dave and
- 27:03Buster's, and you're gonna go refill your party
- 27:04cards so you can go hit more video
- 27:06games. Same kinda thing. Right? Let let me
- 27:08go let me go fill up my card
- 27:09with tokens, and then boom, I'm gonna go
- 27:11spend 25
- 27:13tokens to play this game. Or same kinda
- 27:15thing here. I'm gonna spend one token
- 27:18to to to generate an answer. I I
- 27:20think you do have to understand a little
- 27:22bit about
- 27:23your patterns for these things and and how
- 27:26they come across and and plan for them.
- 27:29So
- 27:30the the the things here that start to
- 27:32get you
- 27:33are the
- 27:35agent to agent actions. So, like, if you're
- 27:37autonomously executing agents or you're having one agent
- 27:40interact with another one, like, great. That's five
- 27:42tokens.
- 27:43The other one is the graph grounding stuff.
- 27:47It can hit you.
- 27:48So you're saying, okay. Hey. I'm I'm hooked
- 27:50up to my tenant graph, like Microsoft Graph,
- 27:53and I've I've turned on these connectors to
- 27:56extend
- 27:57Copilot within my tenant. Well, that stuff starts
- 28:00to get expensive as well. And then from
- 28:03there, you might actually wanna, like, turn things
- 28:04off. Like, you might not want to allow
- 28:07for, like, generative images because of the high
- 28:10cost and and the tokens associated with them.
- 28:13Yeah. It is
- 28:14definitely
- 28:15something to be careful of, and I would
- 28:18say I've I've started having conversations too is
- 28:20looking at how much you expect or where
- 28:22you start. I would say even the message
- 28:25packs. I did the math here quick a
- 28:27minute. If you're buying 25,000
- 28:29messages for $200,
- 28:30you're only spending
- 28:320.02¢
- 28:34per message less. So as opposed to the
- 28:37PAYGo for 1¢,
- 28:38if you buy a message pack, you're spending
- 28:400.8¢
- 28:41per message. I was talking to a client
- 28:43about this the other day. I'm like, just
- 28:44start with PayGo. Use PayGo until you hit
- 28:47essentially $200.
- 28:49Once if your PayGo is consistently
- 28:52$200 a month, then maybe you switch to
- 28:54the message packs, and you essentially pay the
- 28:56same amount, and you get a few bonus
- 28:58messages. But I don't know that there's a
- 29:00lot of benefit to necessarily starting with message
- 29:03packs unless you're a huge organization,
- 29:05but you're just gonna blow through messages and
- 29:08you're like, yeah, we're gonna be spending a
- 29:10thousand, 2 thousand dollars a month. PAYGo makes
- 29:12sense. Or like you said, in a happy
- 29:13world, you're just licensing Microsoft three sixty five,
- 29:16getting the $30 a month license. Yeah. You'll
- 29:19look at the cost. Maybe you only have
- 29:21five or six users in your organization.
- 29:23Buying six $30 licenses a month is still
- 29:26still cheaper than $200 a month for message
- 29:28pack. So really looking at users usage, where
- 29:31do you wanna start with this when you're
- 29:33getting into some of that
- 29:35pricing and
- 29:37trying some of this stuff out? So you
- 29:38probably want PayGo there anyway,
- 29:40at least if you're doing the
- 29:42the the,
- 29:44sorry, the
- 29:45the consumption
- 29:46or or you're doing kind of the the
- 29:48bigger
- 29:49messaging packs that you talked about. Like, hey.
- 29:51If you buy 25,000,
- 29:53you're probably gonna wanna PAYGo to bump up
- 29:55behind it anyway. So you're gonna go and
- 29:57configure all that stuff on the Azure side.
- 30:00Like, it's it's it's a little weird the
- 30:01way it happens too because you end up
- 30:04taking what's effectively
- 30:07a
- 30:08an a linked environment from Power Platform and
- 30:10bringing that over to Azure, and then that's
- 30:12your you you end up with this weird
- 30:14linked billing construct. Like, yeah, you can go
- 30:16into the Azure portal, and you can see
- 30:18Copilot Studio, but and your, like, message consumption,
- 30:21but then there's other stuff going on over
- 30:22in PowerPlat.
- 30:24And you have to maintain that connection too.
- 30:25Like, if you take that connection away, then
- 30:27the billing relationship is severed. It's it's a
- 30:29little weird still, the the way it gets
- 30:31wired up. Anything else? You've been talking. We've
- 30:33been talking for thirty minutes.
- 30:35You're getting over your cold. You're starting to
- 30:37lose your voice, Scott. You still have your
- 30:39entire week to go. Anything else you wanna
- 30:42touch on before we wrap up for day
- 30:44and you wrap up for today, can you
- 30:46give your voice a little break before you
- 30:48dive into all your meetings? So I think
- 30:49a couple of things. Maybe just some, like,
- 30:51high level tips and tricks and and things
- 30:55to keep in mind. So one is
- 30:57design with
- 30:59when when you're building one of these things,
- 31:00you kinda have to decide, am I building
- 31:02it for myself? Am I building it for
- 31:03my team? Am I building it for a
- 31:05wider audience?
- 31:06So as you're going down that path and
- 31:08you kinda understand your audience and who you're
- 31:10building for, that'll help you do better grounding
- 31:12in your instructions,
- 31:13maybe your
- 31:15your your preprovided
- 31:16prompts, things like that.
- 31:18You do have to think through maintaining your
- 31:20knowledge sources.
- 31:21So, obviously, like graph connectors, things like that,
- 31:23they can be a little bit more evergreen.
- 31:25But say you had augmented with some documentation
- 31:28that you dropped in a SharePoint site and
- 31:29that documentation changes over time, you've gotta remember
- 31:32to go back and update that documentation,
- 31:36and keep dropping it in there and landing
- 31:38it in that place, and potentially
- 31:39updating
- 31:40and
- 31:42refining the prompts within your agent to make
- 31:44it work with that new data.
- 31:47The other thing to think about is who
- 31:48you publish your agents to. So I've I've
- 31:50seen this, like, at least within, like, the
- 31:52stuff that I've done, and and, like, I've
- 31:54built some stuff for myself, some stuff for
- 31:56our team. And all of a sudden, somebody
- 31:58will come back to me and they'll say,
- 31:59like, oh, hey. I was showing off your
- 32:01blah blah blah to a friend, and and
- 32:02they wanted access to it. I'm like, well,
- 32:04they really shouldn't have access to it because
- 32:06it's our agent. It's, like, for our team,
- 32:07and it goes around just this subset of
- 32:09data. Right? Like, back to that, if you
- 32:11built the statement of work thing for yourself,
- 32:13you're probably not gonna give that to every
- 32:15person in your organization. It'll be for you
- 32:17and a select set of salespeople and things
- 32:19like that to to to go drive that
- 32:21through. And then finally,
- 32:23these are
- 32:25still
- 32:26nondeterministic
- 32:27systems.
- 32:28They're gonna have wonky answers sometimes.
- 32:31They're gonna hallucinate. They're they're they're gonna need
- 32:33refinement, all that kind of stuff. So as
- 32:35you publish these out, particularly for anything that
- 32:37you publish that goes beyond the scope of
- 32:38you, whether that goes to a team of
- 32:41two, a team of 10,
- 32:43or your entire company, like, it doesn't matter,
- 32:45you should be prepared to kinda take feedback
- 32:48and iterate on that feedback and and republish
- 32:51these things occasionally
- 32:52so that they can pick up the new
- 32:54capabilities. Like, maybe your tenant admin went in
- 32:56and added a bunch of new graph connectors,
- 32:59or you're just on that kinda journey of
- 33:02of continuous improvement with those. Like, I have
- 33:04found that to be important. Like, you can't
- 33:06just often do the wrong appeal, set it
- 33:09and forget it kinda thing. You do have
- 33:10to kinda come back, care, feed,
- 33:13water, and and make sure that your garden's
- 33:15flourishing. % increase or change those instructions. Your
- 33:19point is data's changing too. You may have
- 33:20to go update the instructions
- 33:23to your
- 33:24agent.
- 33:25I would say even knowing
- 33:28you have to know where these agents are
- 33:29pulling data from because the other thing is
- 33:31you don't want somebody in your company
- 33:33to go drop a particular document in a
- 33:36document library
- 33:37and a particular SharePoint site, publish a particular
- 33:40article
- 33:41in Confluence,
- 33:42and not realize that there's an agent that's
- 33:44open to the whole company that's pointing at
- 33:46this, or there's all that
- 33:48security compliance governance stuff to think about as
- 33:52well because now you may have these agents
- 33:54pointed at very specific spots and allow it
- 33:58it's like almost allowing your employees to know
- 34:01where different agents are pulling data from so
- 34:03they can be cognizant of that as they're
- 34:05updating data. It is a consideration because one
- 34:07of the things you can do is you
- 34:09can take an agent, and because you have
- 34:10access to the SharePoint site, you can point
- 34:13it over to that SharePoint site. And I
- 34:14haven't seen
- 34:16at least in I don't know. I'd have
- 34:18to play around with it. I don't know
- 34:19if everybody that I've shared some of these
- 34:21things with actually has access to the SharePoint
- 34:24site or if it's SharePoint or the agent
- 34:26or like, it's a little weird in my
- 34:28head, right, for how it all ties together.
- 34:30I would think at least for SharePoint, it's
- 34:32going to be using it's gonna respect permissions.
- 34:35Like, if I have permission to the SharePoint
- 34:37One would think. One would think. But other
- 34:39sources, like maybe a SQL database
- 34:42or a Confluence page or something like that
- 34:44or, again, you drop it.
- 34:46Yeah. It's
- 34:47it's definitely something to be aware of. We'll
- 34:50leave it at that. I'll have to go
- 34:52play with that one, Scott, because I'm definitely
- 34:54making some assumptions there. I am too. But
- 34:56SharePoint permissions tend to be pretty wide and
- 34:59open no matter how much you want them
- 35:00to be locked down. So that's why I
- 35:01don't know if it's a me thing, if
- 35:02it's a SharePoint thing, or if it's an
- 35:04agent thing. Or a combination of all of
- 35:05those. Yes. Or all of the above, for
- 35:07sure. We'll throw a bunch of links in
- 35:10the show notes about
- 35:12all the licensing stuff, steps for building these,
- 35:14hooking up different,
- 35:16different data sources,
- 35:18all of that. So it would be interesting
- 35:20too, Scott, to hear how other people have
- 35:23used agents, what they've built. And I know
- 35:25I'm starting to build some. Like, I'm working
- 35:27on building an internal knowledge base one for
- 35:29me where I can look at different solutions,
- 35:32documentation,
- 35:33all of that we've created for clients. But,
- 35:35yeah, hearing how other people are using agents
- 35:37as well, reaching out to us on the
- 35:39socials or any of that and having some
- 35:41discussion out there on agents and usage around
- 35:44it. For sure. Awesome. Well, thanks, Scott. Go
- 35:46get better. Go drink some more tea. Try
- 35:48to let other people talk in all your
- 35:50meetings this week, and we will talk to
- 35:52you again soon. Alright. Sounds good. Thanks, Beth.
- 35:56If you enjoyed the podcast, go leave us
- 35:58a five star rating in iTunes. It helps
- 36:01to get the word out so more IT
- 36:02pros can learn about Office three sixty five
- 36:04and Azure.
- 36:06If you have any questions you want us
- 36:07to address on the show or feedback about
- 36:09the show, feel free to reach out via
- 36:12our website, Twitter, or Facebook.
- 36:14Thanks again for listening, and have a great
- 36:16day.