In the final of our AI webinar series, we walk you through how to successfully roll out AI to a business, without it failing as just another IT project.
Alastair Struthers
Hello!
Thank you very much for joining our webinar today. I'm Alastair Struthers. I'm the Customer Success Manager at IT Foundations. I imagine many of you have probably met me, spoken to me, so know exactly who I am and I probably don't need much introduction. I'm joined today by our partners at Pax8 to take us through a really, well I think is a really interesting topic. We're going to be looking at AI Agents specifically around the Copilot environment because lots of people are asking me about them at the moment, and so I thought I would get the experts to tell you a bit more about it in greater detail and do a nice little demo.
Let me introduce Sydney from Pax8. Hi Sydney. She's going to be doing a fantastic demo for you today so I'm going to be quiet and just hand over to her now. There is a Q&A available, so if you've got any questions, pop them in there. We've got AJ also from Pax8 there to help answer those as we go through and we'll deal with some.
At the end, so Sydney, over to you.
Sydney Chong
Thank you very much. Good afternoon, everyone. My name is Sydney. As mentioned, we are going to look at Agents in Microsoft 365 Copilot today. I'm going to bring you how to build agents from the very beginning and then we can go into something that's a bit more complex, more detailed. And once you feel ready, we can look at the next steps as well.
Just a bit about me. I am Sydney. I work in a company called Pax8. We work very closely with the guys from IT Foundations. We have a really good partnership. And what I do on a day-to-day basis is stuff like things like this. A lot of education, enablement, basically teaching people about Copilot, about AI, why you want to use it, why you might not.
But enough about me. Let's talk more about Agents.
Today's agenda, I'm first going to start off with what's new in 365 Copilot. Microsoft is making quite a few big changes, so I thought it would be a good idea to go through the changes really quickly, just so no one's caught off guard when they actually come into place.
And then we're going to look at the main topic of today, which is AI Agents. Why adopt agents, what is sort of like the gap there, and why you would need agents in comparison to just Copilot chat or just Chatgpt.
After that, we're going to go to a live demo, so I'll show you where to actually find those agents, how to make the agents work for you. And if you want something more, we will also look at Copilot Studio, which is sort of like the full-fledged version of the agent builder. And after that, we will look at some stuff like next steps - Copilot and Power Automate.
And finally, look a bit on the licencing side of things, what to expect. And, from there, we will go to Q&A, and that will be the whole agenda for today. I have with me my lovely colleague, AJ. He will help with any Q&A, so feel free to just pop them in the chat or in the Q&A, the team from IT Foundations will also be on hand to answer questions.
Let's begin. This is just a summary of what is new, what is coming through 365 Copilot in the near, very, very near future. For those of you who were in the last session, we talked about Copilot in 365 apps, we looked at Word, PowerPoint, Excel, that sort of thing. You'll probably be very pleased to note that it is now all going to be combined into what is going to be known as Copilot Home - Microsoft is combining chat and Cowork into stuff like this hub. Everything is coming together. It is going to be called Copilot Home. And you will have your Word, Excel, and PowerPoint all built into Copilot. So no more messing around, trying to find things. Hopefully this just brings everything into the same place.
Next we have Code. For those of you who have been looking at, let's say, Claude Code, Gemini, or something like that, it is also now going to be coming to Copilot. It helps you write code, review, explain things to you, and that also powers GitHub Copilot, which is very similar to how you would just use Copilot. You can ask questions like, for example, this command, this line, what does it do? Please explain this to me, or I'm trying to find some logs from this amount of time passed. Tell me what I must do. So you can ask Copilot those questions when you have a Copilot code.
And lastly would be Autopilot. This would be sort of a personal agent, just a really nice helper that is always online, always there to help you. Sort of like a really nice internal coworker that just has that context, it knows almost everything about what you do as a user, as a worker. For example, it know that I am Sydney, I do a lot of demos around Copilot. It will have that context and it will be able to help me just because it does have that, persistent knowledge of what I do.
Let's just go start with the basics. For those of you who were not in the last session, you might be surprised to know that AI users are no longer in the minority. I would say if you have ever done, let's say, a Google search while you were at work, chances are you have already used AI in some way, just because, when you do a search, the first thing you see at the top right now is an AI summary. That is basically using AI. 71% of people have used unapproved AI tools at work, which means that you might be at risk if you are using this AI tool to look at sensitive information, something that contains PII (personal identifiable information) anything like that. There is always a chance that you are actually not compliant with the policy, or you might just be leaking something that you should not. You might have heard of the term shadow IT, that is when you bring unapproved things or applications into your workplace. And in that same vein, we now have the term shadow AI. If you have unapproved, let's say, note takers, they sit in your calls, they're recording everything you say, who knows where that information is going back to. And despite having this knowledge, 51% of people continue to use this every week, even if they know the risks, they know that if they use unapproved AI, their business secrets could be leaked out, but they continue to use AI. And why is that? It's because AI is so helpful. I would say life changing in terms of productivity, how it assists people. There is a way to get around that, which is to use, AI tools or solutions that are safe for business use.
Human labour is reaching its limits. 80% of workers, they agree that
Yes, we don't have enough time to do as much as we want to do with the time that we have been given. But 53% of leaders agree that productivity must increase. What do you do then? There is a huge gap between where people are feeling at their limit and leadership demands more. What do you do then? You turn to AI to help you.
People who use AI are actually producing work that they could not have done 12 months ago. Just think about this slide deck? I've made a slide deck for this webinar, something that would have... taken me a week without AI in the past. I could actually pump this out in a day with the assistance of AI.
Give AI the boring repetitive tasks, let's say data entry or something that I personally
do not enjoy doing and use that time to do something that's a bit more meaningful, something that makes you enjoy your work. Like for me, it is talking to you guys instead of just making slides. So that is what I like to do. That is what matters most to me.
How do we do that? We augment our workforce with agents. If you were in the last session, we looked at how to use Copilot in your 365 apps. We learned how to use it in Teams, in Word, in PowerPoint, in Excel. Let's take it one step further. We are now going to look at how to make Agents, but before we do that, we need to understand what are agents. Think of Copilot as the UI for AI. What this means is that on the left here you have you have me as a user, I put in a prompt to Copilot and then Copilot gives us response. But now if you bring in agents into the equation, we have Copilot that can also connect and talk to the agents. You would see that it just unlocks so much more possibilities, so much more functionality, it is just so much more useful in that sense.
As I talked about before, we talk about work IQ, which is what Microsoft terms as the context behind each person as a user. Again, it knows who I am, it knows
what I do on a day-to-day basis, it knows who I work with, it knows who my bosses are, who my customers are. It will be very specific to me as a person. Someone else who works in design and marketing, for example, miles away from my role, they will also have their own personalised experience and results in terms of what Copilot gives them. Right. I would say, think of it as a really nice assistant. And also, you still have to coach it to give you the best results. You can't just take the first thing it spits out at face value, it'll be like, yep, I'm going to send this to my boss. He'll be happy. Everyone's happy. It doesn't work like that.
You still have to coach it, you still have to make sure you know it's giving you what you want, and you still have to refine it. Agents are more like helpful assistants that you would be able to communicate with using Copilot. You give them a set of instructions and they do that task for you automatically. They can gather information, you can ask them questions, or you can give them maybe like a set workflow, a simple one, and they will do that for you. Think of this as a new way to complete a task.
You are sort of like outsourcing that task in advance. You write it down, you tell the agent what to do, and in the future, the agent will do it for you.
This one is just an example of how to prompt. The reason I have this on the screen is for those of you who were not in the last session, this is what a prompt might look like. On, I would say, 80% of my prompts with any sort of AI, I would still be going with a basic prompt, like summarise this, I throw in a document, it's done, it works, it's not wrong. So that would be a basic prompt, you tell the AI what to do and it gives you a result. And if you want something that's a bit more specialised or a bit more accurate you can go with an engineer prompt. Look through my inbox and my team, summarise any actions for me today and tomorrow, and present them to me with bold formatting if something has been flagged. So you're giving it a bit more details on what to do, exactly how to do it, and how you would like it to be shown to you. Focus on the task or follow up itself. I would say this takes slightly longer to do compared to the basic prompt, but now we look at the next level prompt. It is a lot more detailed and I would probably be able to type this out one day, but forget it by the next day. There's no way I can remember this whole thing.
I could, of course, copy and paste this, save this in a notepad somewhere, and just paste it into Copilot every day. But what if I told you, you could actually turn this next level prompt into an agent? You don't have to remember this anymore. It sits in the instructions, it sits in the agent's memory, and it just does all of this for you automatically.
The way to think about Agents is, let's go back a little bit into the past. We have had GPT 4. So that was maybe, I would say two years ago. And GPT 5 came out. We are now at GPT 6. As the model improves, you can think of it like the brain. So the brain power increases and it gets a bit smarter, it understands a bit more. So you write less in terms of the prompt. In the past, we wrote a long prompt and GPT 4 couldn't understand it very well. The answer wasn't that good. Now we are at GPT 5 or 6, we write a slightly shorter prompt and we get a much better answer in return.
Why not just turn your whole prompt into the instructions for the agent. And I write a very short prompt. One or two sentences, like in the beginning, the basic prompt that I show you, I could just say, do your thing, do your job. And it's going to look at the instructions. It's going to know, okay, Sydney is asking me to do this and this and this and this and it knows because it's all in the instructions. Ta-da, we now see that, okay, we have just taken all of that work in writing a prompt and we've made it the agent's job instead. So the most important part is now stored in the agent's memories. All we need to do is know how to use the agent and how to trigger the agent.
Right. Sounds good.
Last thing here before I go, I think, into the demo. Okay, two more slides. Prompt, just to recap this, is you have to rebrief it every time. Every time you start a new chat, you have to type in, I want you to summarise this document - great, it works, but it is not that reusable. You still, even if you save it somewhere, you still have to retrieve it, you still have to paste it into the chat, and then what happens the next day? You do it all over again. An agent would be able to remember your instructions. If you want a specific type of output, it will be able to do it.
It has its own knowledge base. You can connect it to sources of knowledge. And you can also share this with your team. For example, I could share it with the sales team, the finance team. I can have it run on a schedule. Let's say every day, 5pm, I want it to do that summary for me. And the best thing about an agent is no code is required.
I don't really know how to code, I'm going to be honest, but I can actually make an agent just by using natural human language. If you can write a GCSE prompt, you can build an agent. It is nothing to be scared of. Now you started to think, okay, what can I actually do with an agent? I'll tell you that there are probably 3 main things.
types of agents, basic types. The first one would be a retrieval agent that basically retrieves knowledge from a knowledge base. You ask a question, it gives you an answer, and then it helps you to make a decision.
The second one would be a task-based agent. So you can tell it to, let's say, send an e-mail for me and it will actually do those tasks on behalf of you as the user. And the third one is an autonomous agent. This one works without you actually being there to press the buttons if you wanted to. For example, you could say, give it a workflow, give it a bit of logic if this happens through this. I could say, if an order comes in, look through my knowledge base, gives me the result and then send out the quotation. If you wanted to, you could actually give it that sort of freedom. It's all up to you based on how much control you want it to give, right, to get.
We are very close to the live demo portion.
Before we begin, I just want to tell you a few terms. The first one would be instructions. As I mentioned earlier, these tell the agent what to do. It directs behaviour and tells the agent how to complete those tasks. Knowledge space would be a source of knowledge that you connect to an agent. For example, a SharePoint or a OneDrive, you give it access to a few different folders, and that is where it draws that knowledge from. Capabilities tells the agent what it is able to do. For example, interpret code or maybe it's allowed to generate images. If you turn it off, then it can't generate images. Suggested prompts, these tell the agents, no, these tell the user what this agent does. The very moment they open up this agent, I can see suggested prompts. Prompts, this tells me that, okay, this agent is allowed to do this, this agent can help me with that, and so on. And tools, these will actually tell the agent what it is allowed to do with more of like external systems. If I connect it to the same output.
Look, I can tell the agent, okay, you're allowed to look at my emails, you're allowed to draught an e-mail for me, but you are not allowed to send an e-mail on my behalf. So you could give it many tools, just think of each tool as a sort of permission to do a certain thing. Right? Triggers, they tell your agent when certain things happen. For example, if someone sends me an e-mail, the agent will be triggered and it will come alive and start doing its job. So that's a trigger. And the last thing would be, yeah, agents. You can connect an agent to another agent. Just think of it like a chain of people working together to do a task or a workflow. I promise you that is the boring bit. Now let's go into the live demo. Let me stop sharing and I will share my other screen.
Okay, so this is my screen. Do give me a thumbs up or a happy emoji if you see this just to just so I can make sure. This would be quite familiar to you, hopefully. Oh, Code is here today, so it looks like the changes have come in already. We are right on time.
The first thing you notice is that, on my left, I have this tab called Agents. In the last session, we kind of looked at the pre-built one. If I click into this now. Let's go to more Agents. This brings us to the agent store. Like, just like you have the Apple Store, you have the App Store, we also have an agent store. This is where you can get new agents. In the previous session, we looked at the ones made by Microsoft. Researcher, analyst. And we also have agents made by our team or other third-party apps out there. Like Adobe, Asana, things like that.
Today we are going to learn how to make a new agent. Here you can click on Create Agents or if you click on this arrow again and go to here, New Agent, it is the same thing.
This is Agent Builder, so it says here Agent Builder, and you notice that it looks just like how you would chat with Copilot. I could say, I want an agent that helps.
Me with my employee handbook. I want an agent that is going to help me with my employee handbook. And here you can add managed sources. I can add content. And let's do my sample agreement letters. So that is going to have my terms and conditions inside.
If you want, you can dictate as well. You don't have to type in. You can just use your microphone, but I'm just going to click send. And now it is going to help me build an agent. On the top left, you can see that you have access to a few different models. The first one would give you a quick response. Maybe it doesn't think as much. Or if you want it to think more and give you a better answer, you can choose that.
Or if you can't decide, just leave it on auto, it's fine.
So it is now trying to determine what kind of agent it needs to build in order to give me what I want. You can see it pop up on the screen for one second there. I think it's almost done. Okay, what's new? Let's get out of that. Now you have the screen. On the left with be your chat with the agent builder. It tells you, okay, what it's done. Employment guide in plain English, sample UK employment agreement, which does work. It also tells you what else you might be able to do. Add the document as in source, test it out, or make the agent. So that is how you can continue to refine your agent. And on the right would be your agent itself. I'll just take you through all the sort of like different sections. The first one would be the name. You can give it a logo as well. For example, you can say “Helpful looking robot with glasses and a cap” and I can actually get it to generate an icon for me. While this is generating, maybe we can start thinking of what agents you might want to build on your own.
This part tells you as the user what this agent does. It does not affect how the agent behaves in any way. It's just for us as the human users to understand. This next part is the main part. This would be the system instructions.
Copilot has very clearly laid out, what is the purpose of this agent, what are the guidelines, step-by-step instructions, error handling and limitations. If you notice, none of this is code at all. It is all in plain, proper English. You can write this by yourself, you can write it by hand, or you can also get Copilot to do it for you as I've done. Like this. And if you don't like certain parts, you can always say, I would like to, I don't like this one, suggest improvements, make this more friendly in tone, use a lot of emojis. My employees are all Gen Z just in case. So it makes it a bit friendlier, I would say relatable to your employees. If you don't like something, you can always go back to the agent builder, tell you, okay, this is what I do not like, or this is what I want to refine, please help me. Yeah, now it's losing a lot of emojis, so I like that. Apply both, and it's going to now change the instruction. You should see it come true in just a bit more. Working on your request. If you don't want to do it this way, you can of course just switch to the pencil and you can do it manually.
The next section would be the skills. As I mentioned earlier, think of the skills as things that is allowed to do. For example, I can add one that says, okay, you have to write my documents in a certain way. And so explain my holiday entitlement and
rules that is how you can test it. Knowledge sources, this would be, so for example, it has already told me here that I can connect this to my OneDrive, I can connect this to my Outlook or my Teams or any other connectors that I think would be useful. I could...
Also, just add a document or a folder, and that just tells Copilot or the agent where that knowledge should be from, so it doesn't just go onto the web, it's not going to grab employment agreements from another company and suggest it to my own.
Employees, like holiday rules, I like the emojis, and yeah, that looks pretty good for me. Suggested prompts, as I mentioned, these tell the users what to expect with your agent. For example, I can go in and I know, okay, it's going to break down my leave for me. It's going to tell me what my probation is. And once I'm happy,
If everything, I will go ahead and create that agent. On the top right, you have this button, create. It is private and only available to you. And I will now start a chat. Once you go into this step, you have officially just made an agent within less than 5 minutes. Again, you can see that I have my suggested prompts. If I shared this with my colleague, for example, the first thing they'll see is the screen. They will see the suggested prompts, and then I can also ask you questions. For example, “what is my dress code?” like so, and it is going to give me answers based on that document that I have uploaded. And it should work here. See, okay, so I put in my document actually that Hawaiian shirts are mandatory on Friday. So it has actually pulled that through. Yeah, just to show you what it looks like, that document I have, let's see, where is it?
I have my documents stored on SharePoint. So it would be in a folder here and then Copilot is going to tell the agent, okay, this is where you can get those documents that you need for the knowledge and then it's going to pull that information out.
out and give it to you here. What we see here is an example of a retrieval agent. It asks a question, it gives you a reply in return. Very quickly, that would be Agent Builder. And we now move into Copilot Studio. So you made an agent, you made it in Agent Builder, and you've had it run, quite successfully for the last few weeks and maybe you think, okay, what else can I do with an agent or how much further can I take it?
This is Copilot Studio, which is the full, I would say, full-fledged version of Agent Builder. Ready to automate, it looks the same, but also slightly different. So you have a few options here. You can make an agent with GitHub Copilot, you can make a workflow which is multi-step, but for today, we can actually go with a standard one. Standard one comes with predefined rules and flows. It looks like this. Let me just click in here. It will ask me for a name, so I can say a demo. And it is going to create
that agent for me.
All right, so let me close this out.
In here, you will see that it looks almost similar to what we had in Agent Builder. You would still see the description, you will still see the instructions, but you now have a few more options. At the top here, you have a lot more in terms of monitoring it, evaluating the agent and how it performs. As it sets up, you will see that you would have a lot of different things you are able to do, but It is almost the same. So it is not too much more difficult compared to what's in the agent builder. I would say once you start thinking of, I want to change this to something else, another knowledge base set up in another system, that is when you start thinking of using Copilot Studio.
Here you have instructions, you are able to select different models. In case you do not like or you do not prefer Chatgpt, OpenAI, you can go for Claude, you can go for Grok, there's also Mistral. Depending on what your admin allows you to use, right now, it's provisioned, so I can now start.
I can give instructions, I can add the knowledge sources again, like here. If I click into here, you can see that it's not just my OneDrive or SharePoint that it's able to access, but you can connect it to, say, Salesforce, to an SQL server, anything else that you might be using. So you have a bit more flexibility there. Tools, as I mentioned, would be permission to do a certain thing. If you go into Outlook, you can see that, let's just give it a bit of time to load, it's actually connecting to my account. You would see that there are different tools, for example, the ability to read my emails, to write draughts for me, to reply, and to send emails. So that is where you can access all of this. And in the future, or even right now, if, for example, you have access to an MCP server, what this does is actually tell you, so for example, if I go back to tools, MCP, what an MCP is, is it tells the agent what tools are available to it. So it kind of gives them a menu. And then let's say you have a menu of 10 different tools, the agent is able to decide what tool it wants to use for this task. So you give it a list and it can make those decisions for themselves.
The next thing is triggers. When do you want this agent to come into life? Is it when someone puts a message into a Teams channel? Is it when someone creates a form or maybe someone sends you an e-mail? This is how you can add those triggers. It could be based on a schedule as well if you want.
Connect other agents. As I mentioned, you can have multiple agents working together to do a whole workflow. Maybe one does the approvals, one does the paperwork, and one does the quotation. Multiple different agents. And topics would be, I would say, for example, if you have, let's say, an HR agent, so one could be about leave, one could be about dress codes, one could be about probation. Different topics with different answers inside. And lastly, it would be the suggested prompts. So it is the same as we saw in Agent Builder, tells the agent or it tells the user what you can do with this agent. Let me just show you one that I've made in advance.
Let's go to Parts Manager. This is a very, I would say, general example. Just imagine that you are now in a manufacturing company, people are sending you emails asking for parts, let's say ball bearings, springs and whatnot, and they want you to give them a quotation by e-mail. Manually, this process, just imagine I would have to go to the warehouse or I would have to ask someone, how many ball bearings do we actually still have in stock? That person would have to reply to me, it might take minutes or an hour, depending on how busy everyone is.
And now I have to write out the e-mail, make sure everything's correct, and then I have to send that e-mail out. If someone needed this urgently, it might just not be ideal for them. I've made an agent to help me with that. You can choose the different models you prefer. I've gone with Claude for this one and I have given it some instructions. The main thing to understand here is that you have a limit of 8000. So it means that, you can't just put an entire procurement system or any sort of system into an agent and expect it to take over that smoothly, but I would say keep it small, keep it simple, and keep it targeted. For mine, it's just a four-step workload. The first one is to analyse my emails, look for the parts, and if there's no parts, try to identify what the customer is asking for.
And the second step is to look through my knowledge base. This could be my warehouse, it could be my inventory system, look for something that matches, make sure it's available. And step three, I've given it sort of like a template on how to respond. If the part is found, respond like this. If multiple matches are found, respond like this. If something is not found, respond like this. Don't make it up. Just say, not found. Don't make it up. And also, if the cost is more than 10,000, I want you to escalate this to a manager or someone in charge. I don't want you to handle this.
And step 4 would be to give me or give the customer an e-mail response. This is the structure of my e-mail. I would like you to actually do steps one to three, write that e-mail and send it out to my customer. So it is going to do that autonomously without my input if I set this up correctly. So a lot of time saved. And step 5, as a bonus, I also want you to hand off your work to the next agent in the workflow. One that is going to help me with my documentation, is going to write me the bills and statement of works and so on. And that's it. That's all I want out of my agent. If I wanted another task, if I wanted another workflow, I would just build another agent to do that. Right. And again, I told it, please make sure to escalate this when the order is more than 10,000. So that would be my instructions. My knowledge base, as you saw, it would be be in the SharePoint that I showed you earlier. So it is all sample data. It has my inventory, it has my sample emails and whatnot. It's all there. And the tools. If you click into here, you can see, okay, my agent has tools to create worksheets, draught and e-mail get emails from my inbox, reply to them. I have sent emails turned off just because it was actually trying to send emails to my customers, so I had this turned off. Otherwise, it would make people quite confused.
And if I go back to the main page, triggers, so it is set to trigger when a new e-mail arrives or when someone puts a message into a Teams channel. And for the next steps after this workflow, it will also connect to my approval agent. So that sends it to a manager of approval or maybe to an agent that helps me with my documents. Now we can actually test this agent. I'm going to ask it to try out something. For example, I want to buy 100 ball bearings. Let's see what it says. This is me actually testing on my agent after I built it, and I wanted to, yeah, do its job, do the workflow. Sending, let's try this.
It's going to bring you into this page so you can see what is happening. It should come up with stuff like, yeah, this flow chart here. It's telling me that I'm now searching through the knowledge sources. I will give you an answer very quickly, very soon. In the sources, you can see that I have included some sample emails. I have
database and it's actually looking through all of that to give me the answer. I have on the right the answer that I want, that I need. Ball bearings is actually found two different ball bearings, so multiple options found. Please clarify and I can say, I want the first one. And it should give me the quote and like so, so very professional looking e-mail and then I can say yes, draught that e-mail for me. Yeah, that is the first workflow just done, just like that. Without me having to actually ping someone else in the warehouse team and ask them, hey, how many ball bearings do we have? Or how many types of ball bearings do we have? So you give it permissions. It has permissions that are available to you as the user. I wouldn't be able to, let's say, look at Alastair's emails or Stacey's emails and act as them. It would just be me acting as myself. Now I have this e-mail that looks good. I mentioned in the instructions. If it is more than 10,000, I have to escalate this. Let's try this. I want to buy 10,000 ball bearings, which should be more than 10,000 pounds.
Okay, and since it is a huge amount, it is now going to escalate this to my manager. So that works. And the last thing I told it was, if something doesn't exist, tell me that it doesn't exist. Don't just make up something. Instead of ball bearing, does this...
ask for bears, like I want to buy 10 bears. It should hopefully tell me that no, we don't sell bears. Yes, so it is not a product in our catalogue, even though we did have ball bearings, but it knows the difference between ball bearings and bears. Yes, it will now say, okay, we don't have bears, we have ball bearings.
Is that what you wanted? Yeah, it looks like it looks to me that everything is in order, so it is following my instructions correctly, is helping me make those escalations as required. Yeah, it looks pretty good to me. I would say you have now had a pretty good look at what agents can do. We went from Agent Builder, which is like the light version, and we went to Copilot Studios, which is the full version.
We have about... 13 minutes left. Very quickly, let's just look at where we can go to monitor those things. If you ever find yourself thinking, okay, after Copilot Studios, what else can I do? Let me introduce you to Foundry. I'm not going to go into this today, but just to make you aware that it exists. Foundry, instead of just those few models, you actually have access to just 11,000, but I believe the number now is around 16 to 18,000 different models. New ones are coming out every single day. This is where you come if you want to really build like the powerhouse kind of agents. The really, really big ones you have access to all these agents and you have that much more in terms of what you can do. Copilot Studios is not the end. In fact, it's just the beginning. If you want, you can actually go further. But we will not look at this today. I just want to let you guys know that there is such a thing.
Right. I also want to show you where you can go in terms of monitoring. People are going to ask, I've made agents now, where do I go to monitor them? Right. I will show you two places. The first is in Intune. I am in my Intune admin centre. You can see here on the left.
Agents, if I click here, I don't have any at the moment, but you will see this is where you can go to self-like monitor what your agents are doing, where things are happening. You also have reports as well, and yeah, let's... go here. The next thing would be in the admin centre. You also have Copilot. This is where you can go to see an overview of what people are doing with Copilot within your tenant. For me, I don't have too many users. If you have access, you would be able to see this much more. So you can see what people are doing with Copilot, what are the top actions. The important thing to note is that it doesn't go into the sort of detail, like, okay, Sydney is doing 10 prompts a day and all she wants to see is puppies. So that is not the level of detail you would get, but instead you'll be able to see, okay, Sydney is using Copilot a lot in Excel and in Word. So that is how you can self track the adoption. You can see the usage, how many licences are assigned, and how many people are actually using Copilot.
So that gives you ideas on how you want to structure your training, your rollout, things like that. And search usage. So it is possible to monitor certain things. There's also logs and etc. If you want to know more about this, feel free.
come to the IT Foundations team and so on, and we'll give you a good demo on where you can find all of this information. Let's just jump back into the slides. I just have a few more to present and we will go for the Q&A.
What comes next? So you've seen my demo, you've seen the things that an agent can do. What else do you want to, what else comes next? Agentic Toolkit, it is not just, in fact, what I showed you was a very simple sort of example. You can in fact do so much more. For example, you could connect it to your customers' data, you could get it to identify opportunities for you, you can connect it to your policies, to your intranet, to your processes, help it just get it to help you with all those different tasks you have lying around, and just solve all those pain points you have everything here is a knowledge source. Everything here can be connected to. It's just a matter of what you want to do with that knowledge and what you tell the agent to do. I would say the limiting factor is probably creativity or maybe about how much you actually want to hand over to AI.
We looked here at Agent Builder, we looked at Copilot Studio, and I'm pretty sure we looked at Cowork in the last session. We pretty much covered all three today. And yeah, so each one has its own pros and cons. For Agent Builder, it is like the light version, it is all no code. And give it one narrow, repeatable task. Don't try to make it do too much. And it is very useful for things that you have a pretty much predefined scope that doesn't move around too much. And if you want something that is a bit more huge in scope, you want to develop it a bit more, you have external data connectors. I would recommend going for Copilot Studio just because you now have access to those things. You have access to many more. And also if you want to do something that's like a workflow, something that requires many steps, go for Copilot Studios. Cowork is not an agent, but it's designed to do multi-step. Think of it as you give it a task that has many steps, like, write me an agenda, save this as a Word document in my SharePoint, e-mail this to my boss, and then set a meeting with my boss. All of this is within 365, many steps in the workflows across different apps, but you want it to be done with one single prompt. That is when you go to cowork. Right. Next slide. How to go about this. Most of what you saw is, I would say, included in your Copilot Licence, but if you go to the Copilot Studio Builder, you would see that it does actually require tokens to use once you publish the agent. So that means you are free to test it, but once you publish it, you share it with the wider company or the team, that is when it starts to consume tokens. And how these tokens are consumed is based on the task you ask it to do, for example, agent responses, a generic basic one would cost about one credit per response. If you ask it to do an action, for example, create and update a record, for example, it would cost 1 to 10 credits. If you want it to generate something like an image or write you a spreadsheet or maybe make a chart that would cost a bit more. And lastly, the workflows. If you wanted to do something multi-step, you put in a lot of logic and stuff, that would cost a bit more. There would be different costs depending on the complexity of the task, how much context and how much knowledge you wanted to draw from. Just bear that in mind as you design your Agents.
A little bit on licencing. We have, I would say, 3 main ways of going about this. The first one is business premium with Copilot. So that is the cost there. And essentially, if you're not making two complex agents, this one is good to go with. You have access to Copilot chat and also the agents, the pre-built ones, would all be free to use.
They are all included in price, so you have nothing to worry about. For Copilot Studios in the middle, there would be a cost there. So you would have about 25,000 credits per month just to play around with. And if you feel like, I'm going to burn through that, you can actually just top up with another package of 25,000. Depending on your usage, it might be what you want. But if you feel like, okay, we don't use it as much, we just want to do pay as you go, that's fine as well. You have options for that. So it is around 1 penny, give or take, per credit. Depending on the use case, depending on what you're trying to build, it is worth looking at all three just to see which one fits in the most. Definitely talk to the guys at IT Foundation. They can sort you out. They can tell you which one is the best based on what you're trying to build. And again, so think about the consumption. It is not based on how many people are using, but rather how you design the agent, what sort of features you want how you plan to use this. I would say heavy usage, or if you roughly have an idea of how much is going to be used, definitely go for the credit packs. It's a good value. But if you feel like, just a few people per month are going to use this, try out pay as you go, monitor the usage for maybe a month or two, and then decide if you want to switch on to the credit pack or not.
There's also people asking about what Power Automate plays in this kind of scenario. I would say that it is actually a separate layer. So you could have an agent layer, which is what we looked at today, so Copilot Studios, Agent Builder, 365 Copilot.
this all would be the first layer. And if you feel like after that you want to go on to some automation, you can actually build this into Power Automate so it can make it part of workflow. And next would be if you want something even more. I showed you Foundry, right, which is like the big boy version of this.
So you could have it connected to those sort of things, other things like Dataverse, maybe you have other apps as well you want to connect it to, but it is a journey, it's a step-by-step thing. You don't just go from step one to step 100 in a day. Play around with the first layer for now build up your confidence, make sure you know and understand what you want from it, and then we can move on into the future.
Copilot Studios, my last, this will be sort of like the takeaway. If you start playing around with Copilot Studios, the first thing you want to do is switch the model. Don't just go with the latest, most expensive one, because that would just... burn up your credits before you even know it. I had a call last week, a customer, they made a workflow, they didn't cheque all these things, and it turns out that they burned 10,000 of their credits within a week without even knowing. So it is very important to evaluate what model you plan to use. There is low spending ones, there are higher ones and the premium ones. Just make sure you are aware of what model you're using before you turn the agent on. Second is to forecast the cost. There are ways to sort of estimate what this agent or workflow would do. You can actually ask the agent themselves, can you estimate how many tokens this action will cost?
It will tell you the third is monitor your spend. As I showed you before, there are dashboards, there are logs, there are things you can do to monitor what those agents are doing and how much it's costing you. And #4, logs, look at the logs, make sure you know what the people that have access to your agents are doing as well. Make sure you know what they're up to. Make sure you know that they're not building some super agent that is going to drain all your agents, all your credits in a single day. And once you have this all set up, you are, I would say, free to play around with agents. You don't have to worry too much but always, always do a cheque. Do not just build an agent, let it loose and assume that it's going to be fine. I am at the end of my slides. Thank you very much, everyone, for your time today. Any questions in the Q&A, I'm going to have a look right now.
Okay, I don't see any in the Q&A. Just look at the chat. Is there, yes, I see you on the screen as well. Perfect.
Alastair Struthers
I thought I'd jump in just at the very end. That was really helpful. Thank you, Sydney. That answers a lot of questions that people have been pointing at in my direction, actually. I figure it was really useful to just spread the knowledge and tell a wider audience rather than just the specific conversations I've had with people. Really hoping every friend that useful.
Sydney Chong
Yeah.
Alastair Struthers
If you've got any questions and you're frantically typing away, we'll hang around for a couple of minutes. I'm conscious of the fact that we're due to finish at two, and it is now two. I can see a couple of people have disappeared off because they presumably got other meetings, but if there's anything that anybody would like to ask, please do feel free to drop that into the Q&A. Oh, there we go. Thank you. There we go.
Sydney Chong
Hey, thank you.
Alastair Struthers
That's wonderful. And obviously, if you've got any questions that you don't want to ask publicly here, feel free to drop me a little e-mail. I think everybody in the attendee list looks like they know me. Drop me an e-mail and I can either answer it myself or point it in Cindy's direction and come back to you. I think, I think it looks like we've got... We've got no questions, Sydney. This is great. You've been comprehensive. Wonderful. Excellent.
Sydney Chong
Yeah.
Awesome. Well, thank you very much for having me. Yeah.
Stacey Williams
I'm sorry, there has actually been one question.
Alastair Struthers
Oh, fire away!
Stacey Williams
Craig Forster, if you build and connect your new agent to your Microsoft e-mail account, i.e. and knowledge, and then share the agent across the firm team, do the other team members get access to your emails or info when using the agent. If yes, is there a way to restrict what info is accessed? And if no, is it restricted to each user of the agent's own e-mail accounts?
Alastair Struthers
Great, did you get the question though, Sydney?
Sydney Chong
Sorry, let me just have a look.
No, so as I mentioned, so it is based on the user's own permissions. For example, if I use the agent, it will use my permission. So it will only have access to, let's say, the emails I personally have access to or the documents I have access to. So based on how you set up the permissions, how you set up your sensitivity labels, these would all tell Copilot what exactly it can pull up and what it cannot pull up. Hopefully that answers your question. And yes, there are absolutely ways to restrict that information.
Stacey Williams
And there was one more actually. If we're looking to download data from an energy metering site, how does the agent deal with websites and passwords?
Sydney Chong
And does it have an API to do that or are you just scraping the website? Yeah, I think it depends on the way that you get that information across. If you have a connector, like a pre-built one, I assume that that is handled, but if you just log in, you get the, get the agent to scrape the website might work a bit differently.
Alastair Struthers
APIs are definitely the way to go if you can, because an API will give you a direct connexion in. It can ask specific questions that the company effectively wants you to be able to ask. I can just pull the information back. If you are, as you suggest, logging into a website and scraping information.
Sydney Chong
Yeah, yeah.
Alastair Struthers
See, correct me if I'm wrong here, Sydney, AI can do that at the moment, but I wouldn't necessarily recommend trying to get it to do it because it could pull the wrong stuff down. It's not a real person and I wouldn't rely on it trying to do that just yet.
Sydney Chong
Yeah, it's not best practise. Yeah.
Alastair Struthers
Perfect. Good. Well, let's leave it there. Thank you very much everybody. Appreciate you joining us. Hopefully you find this useful and we'll see you on the next one whenever we figure out what that's going to be about.
Sydney Chong
Perfect.
Right, thank you.