AI-3018: Master Microsoft Copilot Studio – Build AI Chatbots!

0:01[Music]

0:08hey guys I'm back with another video so

0:10let's get started with Microsoft copal

0:12Studio in this uh we already understood

0:15what is generative Ai and what are the

0:17different Microsoft copil which are

0:19available in different Microsoft tools

0:21now it's the time we have to decide

0:24whether we are going to go with

0:25Microsoft co-pilot studio uh with the

0:27low code approach or we are going to go

0:30with Azure AI Foundry where we are

0:32actually going to develop Uh custom

0:34agents with the help of azure AI Foundry

0:37portal now in this video we are not

0:39going to talk about Azure AI Foundry

0:41this is totally focusing on Microsoft

0:44co-pilot studio so if you are a

0:46developer and looking for a pro code

0:48kind of an environment this video is not

0:50for you but if you are looking for a low

0:53code kind of an environment this is for

0:55you let's get started the first thing in

0:58this agenda of this video is what is

1:00co-pilot Studio we will get familiar

1:02with the UI of copilot Studio as well as

1:04there are couple of terminologies which

1:06we need to understand in this like the

1:08first we are going to see what is topics

1:11and then how we can work with topics in

1:13copala studio after that we'll see that

1:15how we can use generative Ai and the

1:18configuring a copala studio based

1:20prompts and then we'll talk about

1:22automating task with actions uh there

1:25are various actions which you can use to

1:27trigger uh we have different types of

1:29topics also we will discuss about all

1:31those things in this and then after this

1:34we will talk about publishing and

1:35distributing our co-pilot agents finally

1:38in the last thing we are going to talk

1:40about

1:41analysis so let's start with what is

1:43copilot studio now if your organization

1:46makes the decision to customize

1:47Microsoft co-pilot or developing custom

1:50agents Microsoft is actually providing

1:52two different tools which you can use I

1:55think we have discussed this thing in

1:56the previous video already so I'm not

1:58going to repeat this but but as you can

2:00see in this slide you have copil Studio

2:03which is designed to work well for low

2:05code kind of a development scenarios and

2:08on the right side you have aure ai

2:10Foundry which is a platform as a service

2:13and it's going to provide a development

2:15portal where you can actually have a

2:17full control over the language model and

2:19you can customize it as you want in this

2:22particular video we will talk about low

2:24code kind of a model that's why we are

2:26going to talk about Microsoft co-pilot

2:28Studio we will not use a aure AI Foundry

2:31Azure AI Foundry we will see after a few

2:34videos Azure AI Foundry we are going to

2:36see in the coming videos but right now

2:39let's focus on co-pilot Studio as I

2:41mentioned this is fully managed hosted

2:44software as a service which is available

2:46from Microsoft it's going to give you a

2:48dialogue and conversational

2:49orchestration so basically the topics

2:51and the configuration which we are going

2:53to see will help you to customize your

2:55logical step-by-step workflow with that

2:58orchestration mechanism you going to

3:00have built-in analytics with security

3:02and governance control so you don't need

3:04to worry about any of those things

3:06because it's all managed by the studio

3:08itself and then you will have an options

3:10to deploy your common chat channels like

3:14web apps social channels and even

3:17teams now if you have a question in your

3:20mind that why you're going to use

3:21co-pilot Studio consider this benefits

3:24which are visible on the slide which are

3:25showing you that how co-pilot Studio can

3:28help you to create custom agent

3:30the first thing is quick deployment your

3:32agents can be created and embedded into

3:34your website with just few clicks and

3:37it's super fast you can Empower your

3:39subject matter experts so your smmes

3:42with the little or no development

3:44experience can actually create agents

3:47quickly and easily using an intuitive

3:50codefree graphical user interface they

3:52just need to have a domain knowledge and

3:54your product knowledge they don't need

3:55to have any programming knowledge in

3:57this case next we have enable rich and

4:00natural conversations so your Microsoft

4:03powerful conversational AI capabilities

4:06enables your users to have Rich multi-

4:08turn conversations which quickly guide

4:11them to the right solution with no need

4:13to retrain AI models so model training

4:16will not be required next we have take

4:19actions so your agents can chat with

4:21your users are great but agents that can

4:24act on their behalf are even better and

4:27that's a reason we have a co-al studio

4:29where you you can create an agent that

4:31not only respond to the user but also

4:33takes an actions to an automated task

4:36which you can configure inside that next

4:38point we have generative AI your

4:41co-pilot Studio makes it easy to include

4:43generative AI capabilities into your

4:45agent so that it can generate natural

4:47language responses based on the joural

4:49knowledge which is retrived from the web

4:52or maybe a specific knowledge source

4:54that you provide based on the relevant

4:56information for your users which can be

4:59coming from your organizational data or

5:01maybe your product related data last but

5:04not the least we have Monitor and

5:06improved agent performance

5:07configurations available your co-pilot

5:10Studio includes this kind of a built-in

5:12benefits where you can do analytics you

5:14can easily monitor the performance and

5:16user satisfaction of your agent and that

5:19basically gives you an Insight so that

5:22you can improvise your

5:23agents now in this slide we are going to

5:26talk about some of the key Concepts

5:28which you need to understand before you

5:30explore copilot Studio the first thing

5:32is topics topics are the core of the

5:35conversational interactions which your

5:37users will have with your agent each

5:40topic Define a dialogue flow which can

5:42be triggered by a specific phrase in the

5:45user input or by that event which is

5:47going to occur your agents can use

5:50generative AI to create natural language

5:52responses that are grounded in knowledge

5:55which you specify for example you can

5:58create an agent that answers employee

6:00questions about expense claims based on

6:03your organization's expense policy

6:05documentation in addition to answering

6:08questions and providing information you

6:10can Define actions that enables your

6:13agent to act as an AI agent that

6:16performs task when your agent is ready

6:19you can publish it to make it available

6:21to users through one or more channels

6:24including Microsoft teams emails

6:27websites and social media platforms

6:30to evaluate the performance of your

6:32agents you can view analytics data that

6:35is recorded and reported in co-al

6:38studio now if you ask me how exactly we

6:40can create an agent in copilot Studio

6:43well creating a new agent is very

6:45straightforward in this first you have

6:48to ensure that on the right top Corner

6:50your power apps environment is selected

6:52which is actually allowing you to create

6:54your agent inside that your environments

6:57are created by power apps administrator

6:59and you have to define a security

7:01boundary for your apps in that

7:03organization after selecting the

7:05environment you can create new agent by

7:07using a natural language to describe the

7:10functionality which you want it to have

7:13when you take this approach copilot

7:15studio is going to initiate a chat

7:17conversation to ask you about how your

7:19agent should behave source of

7:21information which it should use and

7:24other features which you want to include

7:26inside that alternatively if you you

7:29want you can use an agent templates

7:32which are available here there are

7:34templates which are provided for some

7:36common agent scenarios such as travel

7:38planning store or product information or

7:41maybe Financial Insights and there are

7:43many more now let's Deep dive into

7:46topics and triggers you can think of

7:48topics as predefined conversational

7:50interactions that your agent can engage

7:53in each topic is triggered either by a

7:56phrase which is entered by a user or by

7:59event such as an error the topics which

8:02you create for an agent determines the

8:04kind of interaction it is capable of as

8:07you can see in this slide that if you're

8:09going to provide an input like some

8:10greetings like hi and hello is going to

8:13revert with the message that hey I'm a

8:15virtual agent and then how can I help

8:18you with that same way you can see some

8:20other triggers which are focusing on

8:23weather of information or maybe just a

8:25goodbye kind of a message or if you are

8:27getting some kind of an error is going

8:29to return with I'm sorry something went

8:31wrong with that and this is how your

8:33agent is going to

8:35behave when you're working with topic

8:37one of the most important things which

8:38you have to take care is defining the

8:40topic flow because topics defined in the

8:43conversational flow are going to made up

8:45of multiple nodes inside that you can

8:48create a flow that orchestrate the

8:50conversation by connecting multiple

8:52kinds of nodes such as triggers messages

8:56questions variable management condition

9:00topic management and actions now each

9:02one is very useful like triggers are

9:05going to be nothing but phrases or

9:07events that initiates the topic most of

9:09the time triggers will be the starting

9:11one after that you have a message which

9:13is going to send a message to the user

9:15question which is going to ask user a

9:18questions and assign the response to a

9:20specific

9:21variable variable management which is

9:24going to allow you to set get or reset

9:26the variable values with that condition

9:29which are logical conditions branched

9:32into a flow based on the variable if

9:34condition is true it will go into the

9:36true Branch if it is false it will go

9:38into the other Branch then you have

9:40topic management which is going to end a

9:43topic redirect a different topic or

9:46maybe a transfer to an external system

9:48also and then at the end of this we have

9:51a last one which is action which can

9:53call an action to perform an automated

9:56task now if your question is how many

9:59typ of topics can be there well the

10:01answer is two you can have custom topics

10:04or system topics custom topics are those

10:07topics that you create or you maybe

10:10inherit from an existing template custom

10:13topics represent the anticipated

10:15conversational dialogues that your agent

10:17can support when you create an agent it

10:20usually includes some pre-defined topics

10:23for some common conversational dialogues

10:25such as greetings and farewells it may

10:28also include some lesson topics that are

10:30provided as an example to help

10:32developers understand how to build their

10:35own topics on the other hand system

10:38topics are triggered by the system

10:39events such as conversation initiation

10:42errors or some unexpected input you

10:45can't create or delete system topics but

10:47you can disable them if you do not

10:49require in a particular agent scenario

10:52if more than one triggers qualifies

10:54triggers get executed in order to

10:56Creation with the oldest first if you

10:59would like to change that behavior you

11:01can set the priority in the nodes

11:03priority span we will try this kind of

11:05things when we do the lab part of this

11:07particular

11:08module this diagram is actually showing

11:11you a fallback topic the fallback system

11:14topic is particularly useful because it

11:16enables the agent to respond gracefully

11:19when a user enters an unexpected input

11:21that doesn't match the trigger of any

11:23custom topic you can follow the flow on

11:26the slide to see how the fallback topics

11:28can handle the unexpected input so you

11:31can see in the first case in the left

11:33side we have a user input when user

11:35submits an input the agent is going to

11:37use a natural language understanding

11:40model to pass that input into the user's

11:43intent and then based on the intent it's

11:46going to attempt to find a custom topic

11:49with the corresponding trigger with that

11:51if we found the custom topic great but

11:54if there is no matching custom topic

11:56available then it will go into the

11:57fallback topic the fallback topic is

12:00triggered through the unknown intent

12:02trigger the fallback trigger can then

12:05ask the user to rephrase their input a

12:08reasonable number of times before giving

12:10up and redirecting the conversation to

12:13the escalate topic you can see in this

12:15slide it's showing you that it's going

12:17to give users three attempts to rephrase

12:20this and after that also if it is not

12:22done it's going to going into escalate

12:24topic section which can be used at the

12:26end of the conversation and it's going

12:29to to connect to a human operator where

12:31you can understand simply that your

12:33chatbot agent is not able to answer the

12:36question so that human can answer that

12:38this is a typical fallback topic logical

12:41workflow which is always suitable in

12:43most of the cases of

12:46chatbots hey guys sorry for Interruption

12:49my name is maruti and I'm here to make

12:51an very important announcement I hope

12:53you are liking our videos and you're

12:55doing a continuous learning with us on

12:57an Azure cloud and AI related topics if

13:00you are enjoying this thing I'm going to

13:02announce skill tech. club which is our

13:04upcoming website which is going to be

13:06launched very soon we are here to tell

13:09you one thing that everyone who is a

13:11subscriber of this particular channel

13:13will get asure cloud and asure AI

13:15related certification courses free of

13:17cost in Skil tech. Club so you will be a

13:20part of the Skil tech. Club kind of a

13:22membership automatically free of cost

13:25and everyone who's a subscriber of this

13:27particular channel will get those

13:29benefits which are available in that so

13:31what are you waiting for I request you

13:33to please subscribe to this Channel and

13:35share it with your friends and families

13:36if they are also interested in Azure

13:38cloud and AI learning that's it from my

13:41side now you can carry on with your

13:43learning thank you now this is a very

13:46important slide because it's showing you

13:48how you can create generative answers

13:51using a generative AI now generative AI

13:54is a type of artificial intelligence

13:56system that is using large language

13:58models to generate natural language

14:00responses based on prompts and we all

14:03know this thing if you're not familiar I

14:05request you to check out my other videos

14:07which are focusing on generative AIS to

14:10use generative AI in a topic you can use

14:13create generative answers node as you

14:16can see in the slide we have a create

14:17generative answers node in which we are

14:20configuring it to send promp usually to

14:22the users input to an llm and then it's

14:25going to return the AI generated

14:27response basically this is going to make

14:30a call to your llm and then processing

14:32is going to happen at the large language

14:34model the response which is generated

14:36from the large language model will be

14:38provided here in this particular flow

14:41you can set additional configurational

14:43options also where you can include

14:45knowledge Source on which based on that

14:48you can provide your own data and you

14:50can generate answers you can also

14:52specify custom moderations to set the

14:55balance between generating creative

14:57responses versus more relevant answers

15:00and you can also do prompt

15:01customizations to add additional

15:03instructions to the prompt that affects

15:05that how your llm will respond now as we

15:09have seen in the previous slide that one

15:11of the option which was mentioned was

15:13knowledge Source Now by default

15:15generative AI functionality in copilot

15:18studio is based on the general knowledge

15:20which is actually available from the web

15:22however you can add a custom knowledge

15:25source as it is visible in this

15:26particular slide you can actually add

15:29your knowledge source which can be

15:30coming from Azure AI search kind of a

15:32service or maybe an Azure SQL database

15:35or maybe you can also connect with some

15:37Salesforce or Zenex kind of an account

15:39with that remember this kind of

15:41Enterprise data connections are right

15:43now in the preview but very soon is

15:45going to be available for Global

15:46availability you can configure this in

15:49your create generative answers node to

15:51use only a specific knowledge sources

15:54and optionally you can fail over to

15:55General Knowledge from the web if there

15:57are no answers which are are found in

15:59the knowledge sources which you have

16:01provided when you are using a website as

16:03a generative AI resource there are few

16:06things which you have to keep in mind

16:07first make sure that you use a real URL

16:11and not the one which is redirecting to

16:13another page when you using a real URL

16:16then only it's going to allow you to

16:18retrieve the knowledge base from that

16:20you cannot point to a website that

16:22requires authentication because there is

16:24no automated authentication mechanism

16:26available with this kind of an agent so

16:29your websit that requires authentication

16:31or Not indexed by the Bing cannot be

16:33used with this also you need to have a

16:36website URL for generative answers can

16:39have up to two levels of depth so you

16:42can see right now as visible in the

16:43slide your URS can have two levels of

16:46depth like a trailing forward slash

16:49however is allowed for example if you

16:51have this kind of retail SLT SL us then

16:55is something which is going to be

16:56working like this so up to detail and is

16:59going to work but not with the multiple

17:01depth with that so up to two levels

17:03depth after Bas URL you can have one and

17:06two slashes it is fine but not more than

17:08two so this kind of a slashes and

17:10further categories will not work so

17:13please keep all these things in mind now

17:15when you are going to use documents for

17:17generative AI you can upload files that

17:20can actually contain knowledge this

17:22files can be documents with unstructured

17:25text or maybe a structure data which is

17:27maybe in the format of CSV comma

17:29separated value files also make a note

17:33that when you upload files these files

17:35are going to be stored in data verse and

17:37it's going to be indexed which can take

17:39some time before the knowledge which

17:42available in the file is available for

17:44generative AI usage you should note that

17:47while the files are stored securely to

17:49prevent the direct access to them in the

17:51data ver the contents of the file are

17:54visible to all users of the agent so

17:56there is no way to set permissions that

17:59do not allow some of the agent users to

18:01access the knowledge while preventing

18:03the access to by other users so you

18:05cannot have this kind of permission

18:07based configurations also keep in mind

18:09the file size is limited to maximum 512

18:13mb per upload also only the text b files

18:16are recommended for this kind of a

18:18knowledge base your images audio video

18:21and other executable files or binary

18:23files are not supported in this now this

18:26is a diagram which is actually showing

18:28you how you can use conversational

18:30boosting for the topic the

18:32conversational boosting system topic is

18:34going to provide a way to generate AI

18:37answers to your user questions for which

18:39there is no matching custom topic

18:41trigger it inserts an attempt to use

18:44generative AI to respond to a unknown

18:47intent before failing over to the

18:49fallback topic if the create generate

18:51answer node in the topic generates a

18:53relevant AI response it is written to

18:56the users otherwise the conversation

18:58flow moved into a fallback topic and

19:01that is exactly what which is visible

19:03here now in the previous slid when we

19:05have seen this thing as you can

19:07understand this is a user input which is

19:09provided if we found the Matched input

19:12with the custom topic it will go into

19:14that custom topic if not is going to be

19:16going into conversational boosting Topic

19:19in this case it will create a generated

19:21answer with that if we got the generated

19:24answer great is going to find the

19:26relevant answers generated by the AI if

19:28if not then it's going to go into

19:30fallback Topic in that also is going to

19:32try multiple times so maybe multiple

19:35attempts are going to be made to

19:37rephrase that from the user side if it

19:39is still not possible to get the proper

19:41answer it will hand over that to a human

19:44operator this bottom configurational

19:47diagram is actually exactly same which

19:49we have seen in the fallback topic the

19:51only thing is conversational boosting

19:53topic is just adding one additional Loop

19:56inside this which can be possible when

19:59you want to make sure that you want to

20:00boost your conversation with your system

20:02topic now let's understand how we can

20:05enable generative AI for intent

20:08recognition identifying the intent of

20:10the user is a very important task by

20:13default co-pilot Studio use classic

20:15language understanding model to pass and

20:17interpret users input and determine the

20:20users's intent with that as an

20:22alternative approach you can enable

20:24generative AI to determine the most

20:27appropriate top

20:29and actions to trigger based on the user

20:31input you should also keep in mind that

20:34all the generative AI functionality in

20:36the topics that we have covered so far

20:39is actually available in either of this

20:42particular mode however if you want to

20:44have an agent which is triggered action

20:47automatically based on the user input

20:49then you need to enable generative AI

20:52automatic task with actions actions are

20:55going to enable you to move your agent

20:58Beyond and purely conversational

21:00solution so it's not always true that

21:02you're going to give answer in the

21:03conversation sometimes you have to do an

21:05automated action with that and that is

21:07where actions are going to help you and

21:09that is where your actions are going to

21:11act as an AI agent and is going to

21:14automate the task on users behalf your

21:16actions can be implemented as a power

21:18automate flows or maybe a custom skill

21:21you can also use connectors to automate

21:24task in an external service one way you

21:27can accomplish this is to include a

21:29Callum action kind of a node in the

21:31topic flow this approach Works

21:33regardless of whether the agent is using

21:36generative AI or a classic language

21:38understanding model to determine the

21:40user's intent in case of connectors to a

21:43service you can use connectors to maybe

21:46manipulate a spreadsheet in an Excel

21:48online or maybe you want to send an

21:50email or maybe you want to get a forast

21:52from an MSN weather information API when

21:56your agent is configured to use

21:58generating AI to interpret the user's

22:00intent you can create actions that are

22:03triggered automatically by the agent in

22:06much the same way the topics are

22:08executed in that in this approach

22:10actions are triggered directly by the

22:12agent and do not need to be included in

22:15the topic flow these kind of actions are

22:18going to be triggered automatically by

22:19your agent now let's talk about

22:22publishing and distributing before you

22:24publish you have to prepare to publish

22:27so when you're all the agent agent has

22:29all the topics and actions which you

22:30need you have configured it properly you

22:33have to make sure that your generative

22:35AI functionalities are properly tasted

22:38in the co-pilot Studio environment so

22:41once you're ready to publish it and

22:42distribute to the users you have to

22:45first configure the security settings

22:48particularly for authentication methods

22:50which you want your users to restrict

22:52the access of your agent as well as the

22:54authentication method which you want to

22:56choose to determine that which check

22:58channels you can deliver it through

23:00authentication options which are

23:02included in copilot are no

23:04authentication where basically anyone

23:06can use the agent through any channel or

23:09Microsoft entra authentication which you

23:12can deliver through the agent where they

23:14are going to pass through Microsoft

23:17teams and power apps or maybe Microsoft

23:19365 co-pilot and then you can also have

23:22manual authentication where you will be

23:24implementing a custom authentication

23:26solution using any particular identity

23:29provider such as entra ID or any oo 2.0

23:33provider after all this thing when you

23:36are ready to publish you can publish

23:38your agent and deliver it to the users

23:40through the available channels the

23:41channels which are included in this are

23:44Microsoft teams websites email or social

23:47media platforms as you can see in the

23:50screenshot social media platforms like

23:52Facebook Skype line twio telegram are

23:57actually available here as a the partner

23:59channels and please keep in mind each

24:01Channel needs to be configured

24:03independently obviously once publishing

24:05is done our work is not done once your

24:08agent has been published successfully

24:09you have to focus on monitoring and

24:12agent analytics data in our copala

24:15studio you can view the usage metrics

24:17which includes summary customer

24:19satisfaction sessions billing and boost

24:23conversation kind of thing most of the

24:25time customer satisfactions are going to

24:27associate with rating from customers

24:29using your agent your sessions are going

24:32to be showing you downloadable

24:33transcript of agent sessions your

24:36billing is going to show you details

24:37about the cost which is incurred by your

24:39agent while the Boost conversations are

24:42going to show you the details of how

24:44often generative AI found relevant

24:47answers and it's going to break down of

24:49reasons for the times it fails to do so

24:53so basically how good your agent or how

24:55bad it is in giving a conversational

24:58response respons to your user is

24:59something which you can analyze using

25:01this and basically that's what which is

25:04going to be visible in the summary the

25:06right side screenshot is showing you the

25:07summary tab where is actually showing

25:09you an overall usage summary associated

25:12with this obviously you can dig deep

25:15into any of the particular tabs and you

25:17can see more detailed summarized

25:19information about that now last but not

25:21the least at the end of this particular

25:23module I want to add on that if you

25:26really want to Deep dive into my

25:28Microsoft copilot studio and Microsoft

25:30copilot this is a link which I'm

25:32highlighting here this link is going to

25:35help you to start your journey to

25:37explore Microsoft co-pilots you can see

25:40these are just four steps which are

25:41mentioned on this page step one you

25:43understand co-pilot step two you addopt

25:46co-pilots step three you extend it and

25:49step four you build it for your own

25:51purpose and remember at the below Links

25:54of all four sections you have couple of

25:56videos by which you can watch and learn

25:58these things so what are you waiting for

26:00start exploring Microsoft co-pilots

26:02today and I'm going to see you in my

26:05next video thank you

26:07[Music]