Power Virtual Agents in the wild with Abhi Rathinavelu

[mark]: 0:52
in this episode we're going to be focused on power virtual agents and their application in the real world i would like you to introduce to my guests today is from washington the united states works for microsopt as a technical program manager you can find links to his bio and social media in the show notes for this episode welcome to the show Abhi

[abhi_ratna]: 1:16
thank you mark thanks for having me

[mark]: 1:19
great to have you on the show i always like to get my my guests well before i start so tell me what do you do for food family and fun what do they mean to you like everything you do outside of microsoft

[abhi_ratna]: 1:36
you started with my favorite topic food so i am fully allowed to try new casines and the seattle where i'm based out is a really great place to do that so we have a wide range of prasines o low since i'm from india i love india quizine spicy food but i've been also branching out to other types like you know most obvious ones like we have like great terek in seattle and a lot of other cisians like iteopand cuside chinese and greek it we have it it's always fun to go around and try to experience new casing i love that and next family yes i live here with my wife and daughter and we have a four year old dog that kind of keeps me keeps me busy when i'm not at work there is always things to do there in the probably the fun part is you know like definitely enjoying the great outdoors of washington state rights and that is really blessing living in this place seeing these beautiful spots going on hikes and near to water bodies so that is one thing we mentally enjoy and and in general i think just especially now the fall has started and we can see the beautiful colours and always want to you know go around and be in the outdoor sets good for us

[mark]: 3:07
very good very good tell me what's your journey to microsoft how did you how long you been there and and what have you done in the time you've been at microsoft

[abhi_ratna]: 3:16
yeah so with microsoff have been pretty much like right out of the college and i've been with the company for little over fifteen years but i've been working across multiple teams different roads and started off in engineering role developer then i moved around and then i moved to a programme it petrol and this current told what i am doing is very customer facing role and part of the team called a power cat some of my colleagues have already been the guests of this show and power cat is a unique team because they have they are part of the engineering and we help the customers of power platform products like to be successful and that's our charter and and we have a lot of subject matter experts who influence the product through the customer feedback and also we generate a lot of i p part of the team itself so that we can scale better so other customers can replicate from whatever our learnings artifacts we created for making our particular customers successful have been part of this team for the last two years and i personally am specializing on power virtual agents for the last two to fears and part of the scheme and really enjoying the rick

[mark]: 4:43
so so tell me a bit about power virtual agents what is it and you know for those that perhaps listening have not used it can you give us kind of like how how do you explain p v a two customers and in the typical applications they use it for

[abhi_ratna]: 5:02
yeah the simplest way to think about p v is it's a service used for building chat parts enabled chat but without writing code writing very minimal so that is basically our us pits we have the part virtually and product has a very simple easy to use designer which will help even a business user to go in without any deep knowledge of you know data science how the models work they can actually go and create a bard and deploy themselves nature and that is that's what i would say as powers agents in the it's a great mix of you know the benefits of the case and also the easy to use you know experience what we have created so that customers coming you know the organizations who have only business users who are really subject matter experts right they can come in and then just stitch up a conversation using the tool and then actually re and bought for their own purpose

[mark]: 6:16
so so what are the elements that make up a chat but like you know is it purely just a question and answer situation where someone just asks a specific question and a key word and that trigger is a response or is it more advanced than that

[abhi_ratna]: 6:33
yeah that's a great question like when we look at chart boards right like it has gone through it's gone through evolution in the last few years i would say and when we started off i think one of the most popular products in microsofter when it comes to chator was like the and maker chanot you have this what we call the single turn effi type of chat boards where we ask a question and then the chat board gives an answer and then they use understan information right so that was like one of the most popular ones because people thought instead of putting it in a page why don't i just expose it in a chart bout so people can ask the question to the chart bout and directly get to that in stut crawling through the page are trying to find out where the pages and things so that is that's a great thing but then as the chart boat he was right like now we're seeing that yeah more and more customers want to build the actual conversation experience which similar to lie how we want to talk to a human so it's like the we want the chart boards to be smart and context aware so that when the user comes and talks with a but it's not always giving a can response but depending on the uses response that is able to take them to different topics different show them different information or even perform like self service actions and also be intelligent to know like when it's not able to solve the issue it's actually escalating or transferring the customer to a human agent so that they can you need to get help so that's so that's what we are seeing in the market that jampboards are like becoming smarter and the customers one that capability s create because most of the times these chat bats are like extension of the customers brand right so they're putting it in their website and it is if the chart board is not smart enough if it's not able to perform like the basic operations that reflects poorly on the brand a rate so it's a very visible component of the organizations as we see it

[mark]: 8:52
yeah that's that's an interesting way to look at it that is a brand representative when when it's in the public domain you mentioned the word there context aware how does context awareness work or could you give us an example of it in p v

[abhi_ratna]: 9:09
ah yeah so p b has a lot of in build features through which you can get this context avarness so one simple use case like when we have this concept of p v a working with the dynamic three sixty five only channel we have a first party integration out of box integration of the product with the dynamic amy channel so dynamic amply channel is the agent helped us engagement nature that's where the human agent receives the various charts and then replace them so now we have this option of passing context variables in the p v a so what that means is like you could potentially collect information from the chat bar uses problem or the on the uses name or any specific information you want to collect from the user and then when the let's say at some point the bat is not but to solve the issue and and it wants to the user wants to escalate so now the bard is escalading the conversation to the human agent on the other side of the channel right so now that we thought we thought the context awareness of without the context variables the custom the human agent might have to ask the same questions again to the user find out what's happening but since we are passing this context variable capability were having this context variable capability anywhere passing this collected information as part of that so the the human agent is going to get a very clear picture of what was the issue and everything which is passed as a context right directly so that is one straightforward example of how we are doing it and even otherwise context variables like we have bought enabled as part of teams pvbogts which can be built as part of teams so there have the same concept like some of the inn built information about like for the example the users name or email department those things can be pulled directly from let's say the activity three and then surface to the user right so that the body becomes more smarter and smarter so that's one way to do that and another important feature like what we have in terms of entity and slot filling where you can con figure the bars are designed the bats in such a way so that you can have your own system entities or custom entity is like late so that whenever they use her types let's say and information of friend as part of as initial query like you know i want to know i want to know the availability of let's say a black m w or something like that right so so some of these entities can be set up in such a way that you can avoid re askings questions like the bard can be smart enough and understand that okay you're already mentioning the cars color the cars model rate are probably the year so that the subsequent questions where you will have to ask okay what is the cars model watch wot color car do you want all that information can be collected as entities and then used the and the power which all agents he model does this thing called heart feeling right and then it automatically fills those information and helps the user to skip this redundant it's right and then directly takes them to the import the relevant message on the relevant part of the conversation so these are some examples of how you can use text yeah i can keep going on because there's so many applications of how we can do this

[mark]: 13:08
this is good tell us about integration what options are there available integrating into other ther sets that the organization might have

[abhi_ratna]: 13:17
yeah so the the simplest easiest way for us today we have extensions through like one of the popular extension is through the power auto mate and also the connector sequels system the whole power connects so be our own seven hundred plus connect us as part of this ecosystem and you can pretty pull data from you any of these sounded connects using power automat action which can be incoreincorporated as part of the p v so that is that's one popular way of doing it there are also other ways like historically customers have used to like other extension points like but framework skills and use the use skill boat and then you can actually call a skill boat from a p and then continue the conversation and get some information or access late at you know topping to like the business a pas and all those things and and bought composer is another way to do that with the g t p connection you can actually hit the past points collect information from there so there are these will these are the most common ways people try to bring in the data from from the area uses

[mark]: 14:44
m and so what about where does a i play a fit is it possible for me to you know create a um an m l model that i can then provide to the chap bout as well how does or is a i fitting into the mix in any part of what you can do with p v a

[abhi_ratna]: 15:10
so so has that unique unique feature here in the sense like ah we have a generic a model which is which serves across multiple tenise multiple tenants across multiple customers right it's a pre trained pre trained model optimised for customers serves and most of the customers service business related words antic so it's like and so that so there is no real need for you to actually go and train the model for your topics so the expectation is you give a sample set of trigger phrases for your intent or topic and then the model is pre trained enough so that i can identify what you are talking about when the user comes in with the user query the model based on this ogrfhaces will be able to match the right topic so the need for you to actually have training the model from the scratch like using a customer model is is no longer required in the in the p v scenario so that is a unique advantage but having said that there are also customers who say that in spite of the main in spite of the generic model they might want to argument it with additional mail capabilities like in they want to their own customer models so in that case they can use the like the extension points which we just talked about right like you could potentially use you know louise model and then call it from the p v from the fall back and if if they the part actually agents generic model is not able to catch it then it will go into the fall back and then the louis modelinia which is configured in ye fall back and do the accilary intent matching and to find out if it is able to resolve it and if not then you can decide the experience so there are few ways to do that

[mark]: 17:21
so where does search come into it and and what i mean is that using p v a to let's say search a repository of information and provide it back so let's say internal stay holder so that's not public facing but let's say i needed two and what am i asking for here is this possible to have you seen something like it so i'm wanting to search or carry data that might be for example sitting in share point as a data repository is that possible and like say you know hand back here's a p d f document about what you just asked for is that type of is that supported

[abhi_ratna]: 18:00
yeah so that is supported think about that kind of let's let's play back that scenario which you just talked about right so in this particular use case basically the use might come and say that you know i want to i want to learn about les odcastfor an example it can you have a document on paspodcasting somewhere in in the shot point so when the user makes that query the maximum response you can potentially give is a single turn response right you you will be able to look into the chart point and throw one link back saying that okay go read the top so this is really not if you look at it it's really not a conversational experience it's an effect experience so so they affect you the challenge what they're doing effect experience is you know extensively using pa for that matter any any tool which is more optimates for multi actual real conversation building experience right because the more and more effect use you just create typically if you look at it using you know tools like you i make they run in like thousands of packs so the maintaining of those effect use the integrity of those triggering trigger phrases become more and more conflixso you need to become your model needs to become like more more sophisticated and and to be able to always pick the right topic like

[mark]: 19:48
correct

[abhi_ratna]: 19:49
so if you have even a slight overlap it might not give you the right so so you can definitely do it in this case how it will be like yes you could we have built feature called suggest suggested topics you can use that feature to mind the short point page in directly you can actually export it to a a a lot of document and then mind it and then it will create to you this single term questions and answers and you can do that nd but the thing to keep in mind is the more and more becomes larger you love to maintain the trigger phrasing integrity so that you can maintain the good triggering performance and the other way to do that is you can you can use the p more for the actual conversational multi turn experiences which involves more complex actions and more back and forth conversations right and for these kind of a single turns you could use to fall back and use your existing can maker or custom which already has the trepocity so that you need not reinventthe weel you just can tend your t v bought to pull the information from your unit maker and display in p so that that

[mark]: 21:10
that's good that's good i like it i like it tell me in your experience with working with the various customers as part of the cat team what are the if i could get you to give me the top five use cases that that companies are using without mentioning the company names what are the most common use cases that you're seeing either publicly in other words outside the fire well we're where a customer is using p va to engage with it with their customers and then the other one which is where organizations are using it maybe as part of h or part of ou know internal processes around you know how do i find a policy on travel if i want to travel you know overseas for example based on my location what what are you saying what are the typical use case is four p v a that you're seeing in the market at the moment

[abhi_ratna]: 21:58
yeah yeah you know like so this is a very interesting question because you know i have worked with a wide range of customers across different verticals from retail to you know like hospitals and you know like tech companies right doing the customer service are mortgage companies you name it right so now now like according to so what i would say is it really depends depending on the customers business process they might choose to auto mate different things as part of the p v but there is one common thread across all this customer right one thing is what we call us the deflection or the deflection rate so most of the time they re talking thinking about how to use the pvabot to deflect the customer users reduced the call volume so that the boat can deflect the incoming charts to the human agent so the usman agents can be more efficient so i think that is the common thread so having said that what they have what they do for that is there always most of them always start with their existing data with their existing called transcripts from the call centers so this i'm talking about the public facing boards right so so they always look into those existing signals to see like where are the areas like it's always we have seen it's like an eight twenty approach like most of the time it's that twenty percent of the use cases which generate the eighty percent of the call wall and they sort of see like what are the what are the most highest all drivers like what are the customers calling about and then they say like how do we build the topics for the how do we add more self service actions for this so that stamer need not actually go and talk to human but get get their own thing get their answer directly and then the important thing is like what are the chan that the bought has to be depending on where the business typically interact with the customer it could be what's up or it could be web channel or it would be like um you know like facebook messenger depending and they meet so so those are the things and languages and other factor depending on the geography and the customer base language might be and other things so this i would say it will be like the outer skeleton of when you are going to look into the use cases so you have to the customers usually make sure that they have these right signals before they jump into the actual use case per set right and i would say at a very high level the common use cases which i ave seen across all these verticals most of these vericles maybe some of them may not be applaiable for some specifical specific verticals like you know like health agencies and things like that but for most most organizations questions about their like their services right like it could be if it's a retail organization like what is the return policy or shipping policy or those kinds of information right and same thing like it there's a service based organizations things around like services las cansolations those kind of things so all that policy information they try to um all the water the services available all that information the companies typically used to use the chart bout to phase that because that's an easy start right so you can you already have the content you just have to create the conversation topics and there you have to go and after that the other big piece what i've seen with customers so it is around a depending on whether they are like a service company retail company or a product company right if its service company like scheduling is a big girl component customers love to use chart boards for because otherwise they have to actually engage a human to do the scheduling right like so they love to use the chart board in this case so that they can use the chart board for scheduling and in the retail context if you look at it checking the shipment shipping status like the order status right order history those kind of things are very popular across different retail companies and across all the companies i count management is another key used case you can see that a lot of companies don't want the customer to call the call center to reset their password or do something else right so so they want to enable all those things through the chat board itself and other key things about the building and payments you know if they want the customer wants to make a payment through the chart board check what is the payment history right what is the monthly building if they want to change it show them the invite all those things are also very popular use case for this kind of external fate in c x bouts in the th these are the top things which come to my mind now

[mark]: 27:43
it's good it's good last question i have for you is around return on investment you know when when a customer is justifying putting in p v a i can see a big factor in that would be that deflection rate right if you can deflect a call going to a human person that could you know last a period of time that's that's a tangible dollar reduction that you can do if you can deflect it because you have answered this question a million times already and you can do it through low cost channel like p v a right electronically are there any other kind of levers that think of around m r i where there's you know the cost to benefit for the customer you can demonstrate

[abhi_ratna]: 28:30
so you know we talked a little bit about the public facing bards i didn't get to the internal bards with but i would have said the thing about the arrows the arrow is different for public facing bats versus internal but you rightly said that most of times the deflection um rate is through one of the key companies in the case of public facing boards it's not going to the human agent it's handled by the bold whereas in the internal facing it's like let's say teams are building for their departments like for benefits or benefits bought or things like that it or department is building apart so there it's a slightly new ance to approach where we have seen that it's not only the way of of reflecting from the human agent but it is more about the cost savings through the perform the improvement in the performance right number of hours saved for example you know a company called automat the who process of getting employment verification letter through the butt that could have saved them like you know ten thousand hours for that particular year and that means that it is bringing them to a significant amount of cost it so and then the speed of getting that thing done so all the slow the efficiency factor placing like what the bought handled time like probably earlier day to take like two days to get the whole thing done into now with the bark they can get it done with like within minutes let's right so that could count so they're talking about not only the dollars figures which is obviously the most important one and most then were also talking about the efficiency with which the task is done the number of hours saved and then how much it is improving the human agent also because if the board is effective it has so in call center in call centers they have important metrics what they call us the data like the time to resolution right like so that is like which kind of says like how long how ever that human agent is it how long it took for him to go and figure out the problem and provides solution so now if you have a bat which is like a vision which is giving him a lot of context information handing a lot of good information for him to finish the to provide a resolution faster so that's improving the productivity of the human agent as well so so there are different angles you can look at it in terms of the way it's not only the dollar out

[mark]: 31:33
m this is good this is good by it's been a pleasure to have you on the show thank you so much for coming on

[abhi_ratna]: 31:42
the pleasure is all mine thanks mark for having me love to talk about this and so glad i could share my experience with this audience