Transform customer experience with an array of AI agents

Hi, everyone. Good afternoon.

Um, so I'm going to take a few minutes and uh walk you through um generative agents that we have now deployed into production in multiple enterprises.

Um, and when we look at the latest developments over the last couple of months, we have seen gen um apply in multiple areas like marketing copies, brand, creatives, advertisements, analytics, code and productivity and also video and gpt agents are coming out as pretty solid developments.

At this point in time, we are seeing a lot of enterprise use cases that can quickly unlock some value for you within like two or three months. Literally, I mean, you can deploy them into production and they start working from a business standpoint of view in, in, in under three months. And that's kind of the platform that we have developed. It's max dot a i, it's available on the aws marketplace. It works in conjunction with aws bedrock to create and deploy these agents in a very compliant manner from your uh from within your enterprise.

And when we look at some of the use cases that are coming out, there is obviously the baseline um question and answer engine kind of aspect that is there. But beyond that, if you want to look at unlocking voice of customer, which is coming from various streams, and then how do you look at the sentiment that's coming in? Finding out some of the issues that might be affecting your customer experience on an ongoing basis. We are able to do that effectively.

Uh content marketing is a core use case that we have seen. Um and I'm going to show you a case study about it on how that's driving click through rates and conversions um across a range of email and mobile notifications as we go by.

These agents are also really powerful in doing training of employees, especially if you have a larger enterprise which has a lot of reps or service staff. What's the kind of training regimen that's going on? What are the topics which are resonating? What is driving nps in an effective manner? That's another set of agents that are there.

We're also seeing a lot of work that's going on from a customer service standpoint of view. So if you have l one l two reps that are running around um especially in the l two layer. When a question comes up, usually your agents have to spend a lot of time in figuring out what the actual issue is. And then what is the resolution for that is from technical documentation and other ways. And we are seeing gp t agents being able to do that fairly fast come out with the top three or four resolutions and then that speeds up the rep productivity. In this case.

Similarly, there are a few other things from a digital marketing standpoint of view and productivity standpoint of view. Those are specific use cases that are easily deployable and that that can generate the returns that you're looking for.

Now, the way we are able to do this is because the core agent architecture that max is able to deploy on top of aws. So for all the foundational models, we typically use what is available from a bedrock stand point of view, which is already approved from enterprise governance standpoint. So whether you want to use gb d four for a particular one or anthropic for a particular one, or for example, an open source set of models, we are able to pick and choose what's necessary for a given use case.

And once you have the model framework in place, um one of the things that the platform uh brings to the table is looking at a wide variety of data sources which are sitting inside your enterprise and being able to quickly pull that together into a dynamic mesh which then gets converted into vector embeddings and, and that powers the agent as you go by.

So that process, whether it's generative a i or classical a i or some of the other applications we have always seen that that has been a really painful part with enterprise. And that's where we have done a lot of innovation to speed that process entirely.

And then that goes from a lm inference standpoint of view. And then we are putting a tool chain on top of it which provides agents long term memory prevents them from hallucinating on the block. There is a line chain that has been connected so that you can connect to various other applications that are necessary for the agent to operate.

And when we are talking about generative a i agents, the way we are looking at it is you have a lm core that is powering the agent. But then you also want to use either traditional a i or traditional automation techniques to be able to connect the agent to the rest of the enterprise systems that are there and then allow it to execute a digital task.

So essentially take the generator by i core, connect to the traditional um a i or machine learning aspects and then it's able to execute a digital task in its entirety. And that task can be, you're sending a set of emails that are going on from uh from sales force. You're executing something as a marketing campaign from an adobe standpoint of view, you're executing a task in zendesk from a customer service standpoint of view.

So there is a wide variety of productivity tasks that agents are able to execute in this case. And then all this is available both with the user interface that is easily usable by business. At the same time, they are also available as api end points, which means that if you have a legacy a application, then we are able to plug these agents into those and then they automatically get upgraded with the with the power of the llm. That's that's doing it.

So all this entirely is something that we are able to deploy via aws bedrock on top of it within roughly 6 to 8 weeks at this juncture. So, so the reason why that is the case is because these are fairly easy to set up from the aws marketplace itself. And then we are able to connect to a range of other systems that are there within enterprise and then push it out um from a us uh from a business standpoint of view.

Now, this is a case study of where the agent is able to do subject line testing as well as marketing copy testing at scale. And this is something that one of our customers has production to almost 8 million emails per week. It's driving double digit incremental from a click through as well as conversion standpoint of view. Forrester recently spoke with them and validated it and published it in their report a couple of months back.

And the way this works is that there are a lot of inputs that we can look at it from a marketing standpoint of it could be um about the campaigns that they're running the traditional aspects that they have in terms of what they're feeding as a marketing content standpoint of view, the length parts of speech, sentiment score and a range of other things that the large language model is able to take as an input and then generate the different subject lines.

But then the subject lines are only one part, then you are to figure out who are the different cohorts of customers that you want to test them against and then take the power of lm combine it with experimentation, connect it all the way to something like an adobe target or something else which then powers through a range of campaigns on a weekly basis to get some insights.

And one of the things that it's also able to do is for those specific customer segments that you're targeting, we're also able to get more psychographic information. What do they like? What's a tone that will work for them? What is the kind of type of subject line or pros that works with them? All, this is technically new information which is not available about the customers in the enterprise today. So it generates a lot of behavioral insights along with the marketing content that's going on.

So what i'll do is essentially we have developed a set of these agents which are ready to deploy. So there is a core um agent called maverick, which does um q and a kind of engine and this applies very well on multi multiple documents on multimodal use cases. There is also raven which does voice of customer you got kit which does more around customer complaints kind of issues as well as uh customer success and agent productivity and then content marketing.

So these are a couple of different agents which are available today on aws marketplace, which you can deploy very, very fast. I mean, within a month or two depending upon your enterprise um governance protocols and everything else.

Um, we'll go through a couple of these agents. Um we can't do the demo but i have a quick video of each one of those so that you can see how they work in action. For example, if you take my break, a lot of times we are seeing close to 20 to 25% improvement in productivity as you use some of these agents from a knowledge services kind of a task and point of view.

So if you look at maverick, um the way it works essentially is that you're taking a lot of the information that is available in the enterprise and then you're able to create a set of collections and then you can switch between the collections as you go by and then query them. And these collections at the end of the day are controlled, user access control. So specific groups can access only specific collections and that's something that you can implement from a enterprise it standpoint of view.

These collections are very easy to manage. You can just upload a new collection, tell the agent what is the role of the description that it is doing? This can connect all the way back to amazon s3. And then the user interface gets deployed on top of it and it's available for the business users to use the data, right? And then it connects to a wide variety of fal sources too. So that's kind of the way the agent functions in this particular in this particular case.

So this is kind of a core question and answer q and a engine kind of a concept which can be deployed really fast, especially if you have different departments like within marketing, within branding, they want to take advantage of existing knowledge base and then drive productivity from there.

A secondary. Um and, and a more advanced one is called kit. This works very well from a complaint standpoint of view or if you have um let's say a large amount of information in the past, what you got to do is at the end of the month, you got to kind of do a really big n lp type of a project where you've got a team which goes and gets all the data figures out what the sentiment is, what the topics are and then create a dashboard out of it.

So now all that can be done with the generative a i agent. And then the way it works largely is you got a complaint summary, how the agent performs. You can define what these different topics are on an ongoing basis, right? from a user interface saying that i'm interested about a product interest or are there customers who are talking about security concerns that they're looking at and other things that they're, that they're asking and these questions or these topics can be defined directly by a business user.

And then the agent and the generative agent keeps monitoring for these topics on an ongoing basis. And that reports back saying that here are all the issues that have appeared in the last one day, one week or however, you want to see from a business standpoint of view.

Similarly, we're looking at um content marketing. So this one is an interesting one. This takes um generative a i to generate content and then combines with experimentation. This is a case study that we spoke about. It can do a range of things, it can create images, it can create full marketing copies, it can generate subject lines given a topic.

So the idea here is uh use the lms to generate the content and then combine it with an experimentation engine, look at the target set of groups and then push those content all the way to the marketing engine and see how that's uh that's scaling up.

Um and one of the things that we are noticing is that the large language models keep getting more and more powerful on an ongoing basis. right. So any of the use cases that I've showed so far, some of these are going to get commoditized and more new and advanced use cases will appear.

So what we want to do is that can we give you a build your own age and capability where pretty much anybody in your enterprise can go in and create your own agent and it will just function like a gp d store, but in a very compliant manner, it will work inside your enterprise with the governance protocols that your it as well as the security team approves.

And the way it functions is that you go and do a couple of clicks and it makes a new agent for you. And um and then you are able to use it as a business user in this case.

So the way it works essentially is that you can select what models that you want, whether it's gp four or anthropic or certain other things. There are like factory settings, you can use them. And then the entire architecture that i showed you before becomes like a tool chain which gets printed.

Um and then it makes a new agent out of it. And now there are different types of capabilities. These agents can do, it can be q and a, it can be like a chatbot. kind of functionality, but it can also do basic data analysis very well. When you're using the gp d four code interpreter capability at the back end or equivalent llm model, it can build that agent.

Now, one of our customers are like 100 different agents that they built and uh they're just being used by different um business users pretty much on an ongoing basis, right?

So, so technically, it's like an army of a i agents which are getting created for different business user groups on an ongoing basis. And the rate at which the progress is happening from a large language model standpoint of view, what we expect is that there are going to be more and more basic use cases which will get commoditized and they will be served by this kind of interface.

Uh this is a very affordable, easy to use um agent. So it's just a flat fee on a monthly basis. There is no limitation on how many agents get created and then they can be used for different purposes, whether it's marketing or branding or a set of data analytics, copilots that you're doing and each one can be deployed into specific departments.

So supply chain gets its own one within supply chain. Let's say there are somebody who's doing safety swap, then they get their own thing. There is marketing which is doing some sort of a specific marketing analytics, they get their own pocket agent, there is product teams which are using product documentation, they get their own agent.

So it's kind of an army of a i agents that can be deployed within your enterprise across all your departments on an ongoing basis. And then just available for you and for your it teams to be able to manage that along with the business user groups.

So it kind of brings the full power of generative agents to your doorstep, which you can use. And then the other agents that i've shown earlier were essentially more advanced use cases which cannot be handled by this sort of agent builder in this case, right?

So that's what i want to share with you today. This is available on the aws marketplace and uh which can be downloaded from there directly and you can install it within your own enterprise environment. It's an enterprise only product. Here is the remaining transcript formatted for better readability:

And when we look at some of the use cases that are coming out, there is obviously the baseline um question and answer engine kind of aspect that is there. But beyond that, if you want to look at unlocking voice of customer, which is coming from various streams, and then how do you look at the sentiment that's coming in? Finding out some of the issues that might be affecting your customer experience on an ongoing basis. We are able to do that effectively.

Uh content marketing is a core use case that we have seen. Um and I'm going to show you a case study about it on how that's driving click through rates and conversions um across a range of email and mobile notifications as we go by.

These agents are also really powerful in doing training of employees, especially if you have a larger enterprise which has a lot of reps or service staff. What's the kind of training regimen that's going on? What are the topics which are resonating? What is driving nps in an effective manner? That's another set of agents that are there.

We're also seeing a lot of work that's going on from a customer service standpoint of view. So if you have l one l two reps that are running around um especially in the l two layer. When a question comes up, usually your agents have to spend a lot of time in figuring out what the actual issue is. And then what is the resolution for that is from technical documentation and other ways. And we are seeing gp t agents being able to do that fairly fast come out with the top three or four resolutions and then that speeds up the rep productivity. In this case.

Similarly, there are a few other things from a digital marketing standpoint of view and productivity standpoint of view. Those are specific use cases that are easily deployable and that that can generate the returns that you're looking for.

Now, the way we are able to do this is because the core agent architecture that max is able to deploy on top of aws. So for all the foundational models, we typically use what is available from a bedrock stand point of view, which is already approved from enterprise governance standpoint. So whether you want to use gb d four for a particular one or anthropic for a particular one, or for example, an open source set of models, we are able to pick and choose what's necessary for a given use case.

And once you have the model framework in place, um one of the things that the platform uh brings to the table is looking at a wide variety of data sources which are sitting inside your enterprise and being able to quickly pull that together into a dynamic mesh which then gets converted into vector embeddings and, and that powers the agent as you go by.

So that process, whether it's generative a i or classical a i or some of the other applications we have always seen that that has been a really painful part with enterprise. And that's where we have done a lot of innovation to speed that process entirely.

And then that goes from a lm inference standpoint of view. And then we are putting a tool chain on top of it which provides agents long term memory prevents them from hallucinating on the block. There is a line chain that has been connected so that you can connect to various other applications that are necessary for the agent to operate.

And when we are talking about generative a i agents, the way we are looking at it is you have a lm core that is powering the agent. But then you also want to use either traditional a i or traditional automation techniques to be able to connect the agent to the rest of the enterprise systems that are there and then allow it to execute a digital task.

So essentially take the generator by i core, connect to the traditional um a i or machine learning aspects and then it's able to execute a digital task in its entirety. And that task can be, you're sending a set of emails that are going on from uh from sales force. You're executing something as a marketing campaign from an adobe standpoint of view, you're executing a task in zendesk from a customer service standpoint of view.

So there is a wide variety of productivity tasks that agents are able to execute in this case. And then all this is available both with the user interface that is easily usable by business. At the same time, they are also available as api end points, which means that if you have a legacy a application, then we are able to plug these agents into those and then they automatically get upgraded with the with the power of the llm. That's that's doing it.

So all this entirely is something that we are able to deploy via aws bedrock on top of it within roughly 6 to 8 weeks at this juncture. So, so the reason why that is the case is because these are fairly easy to set up from the aws marketplace itself. And then we are able to connect to a range of other systems that are there within enterprise and then push it out um from a us uh from a business standpoint of view.

Now, this is a case study of where the agent is able to do subject line testing as well as marketing copy testing at scale. And this is something that one of our customers has production to almost 8 million emails per week. It's driving double digit incremental from a click through as well as conversion standpoint of view. Forrester recently spoke with them and validated it and published it in their report a couple of months back.

And the way this works is that there are a lot of inputs that we can look at it from a marketing standpoint of it could be um about the campaigns that they're running the traditional aspects that they have in terms of what they're feeding as a marketing content standpoint of view, the length parts of speech, sentiment score and a range of other things that the large language model is able to take as an input and then generate the different subject lines.

But then the subject lines are only one part, then you are to figure out who are the different cohorts of customers that you want to test them against and then take the power of lm combine it with experimentation, connect it all the way to something like an adobe target or something else which then powers through a range of campaigns on a weekly basis to get some insights.

And one of the things that it's also able to do is for those specific customer segments that you're targeting, we're also able to get more psychographic information. What do they like? What's a tone that will work for them? What is the kind of type of subject line or pros that works with them? All, this is technically new information which is not available about the customers in the enterprise today. So it generates a lot of behavioral insights along with the marketing content that's going on.

So what i'll do is essentially we have developed a set of these agents which are ready to deploy. So there is a core um agent called maverick, which does um q and a kind of engine and this applies very well on multi multiple documents on multimodal use cases. There is also raven which does voice of customer you got kit which does more around customer complaints kind of issues as well as uh customer success and agent productivity and then content marketing.

So these are a couple of different agents which are available today on aws marketplace, which you can deploy very, very fast. I mean, within a month or two depending upon your enterprise um governance protocols and everything else.

And um we'll go through a couple of these agents. Um we can't do the demo but i have a quick video of each one of those so that you can see how they work in action. For example, if you take my break, a lot of times we are seeing close to 20 to 25% improvement in productivity as you use some of these agents from a knowledge services kind of a task and point of view.

So if you look at maverick, um the way it works essentially is that you're taking a lot of the information that is available in the enterprise and then you're able to create a set of collections and then you can switch between the collections as you go by and then query them. And these collections at the end of the day are controlled, user access control. So specific groups can access only specific collections and that's something that you can implement from a enterprise it standpoint of view.

These collections are very easy to manage. You can just upload a new collection, tell the agent what is the role of the description that it is doing? This can connect all the way back to amazon s3. And then the user interface gets deployed on top of it and it's available for the business users to use the data, right? And then it connects to a wide variety of fal sources too. So that's kind of the way the agent functions in this particular in this particular case.

So this is kind of a core question and answer q and a engine kind of a concept which can be deployed really fast, especially if you have different departments like within marketing, within branding, they want to take advantage of existing knowledge base and then drive productivity from there.

A secondary. Um and, and a more advanced one is called kit. This works very well from a complaint standpoint of view or if you have um let's say a large amount of information in the past, what you got to do is at the end of the month, you got to kind of do a really big n lp type of a project where you've got a team which goes and gets all the data figures out what the sentiment is, what the topics are and then create a dashboard out of it. Here is the remaining transcript formatted for better readability:

So now all that can be done with the generative a i agent. And then the way it works largely is you got a complaint summary, how the agent performs. You can define what these different topics are on an ongoing basis, right? From a user interface saying that i'm interested about a product interest or are there customers who are talking about security concerns that they're looking at and other things that they're, that they're asking and these questions or these topics can be defined directly by a business user.

And then the agent and the generative agent keeps monitoring for these topics on an ongoing basis. And that reports back saying that here are all the issues that have appeared in the last one day, one week or however, you want to see from a business standpoint of view.

Similarly, we're looking at um content marketing. So this one is an interesting one. This takes um generative a i to generate content and then combines with experimentation. This is a case study that we spoke about. It can do a range of things, it can create images, it can create full marketing copies, it can generate subject lines given a topic.

So the idea here is uh use the lms to generate the content and then combine it with an experimentation engine, look at the target set of groups and then push those content all the way to the marketing engine and see how that's uh that's scaling up.

Um and one of the things that we are noticing is that the large language models keep getting more and more powerful on an ongoing basis. Right. So any of the use cases that I've showed so far, some of these are going to get commoditized and more new and advanced use cases will appear.

So what we want to do is that can we give you a build your own age and capability where pretty much anybody in your enterprise can go in and create your own agent and it will just function like a gp d store, but in a very compliant manner, it will work inside your enterprise with the governance protocols that your it as well as the security team approves.

And the way it functions is that you go and do a couple of clicks and it makes a new agent for you. And um and then you are able to use it as a business user in this case.

So the way it works essentially is that you can select what models that you want, whether it's gp four or anthropic or certain other things. There are like factory settings, you can use them. And then the entire architecture that i showed you before becomes like a tool chain which gets printed.

Um and then it makes a new agent out of it. And now there are different types of capabilities. These agents can do, it can be q and a, it can be like a chatbot. kind of functionality, but it can also do basic data analysis very well. When you're using the gp d four code interpreter capability at the back end or equivalent llm model, it can build that agent.

Now, one of our customers are like 100 different agents that they built and uh they're just being used by different um business users pretty much on an ongoing basis, right?

So, so technically, it's like an army of a i agents which are getting created for different business user groups on an ongoing basis. And the rate at which the progress is happening from a large language model standpoint of view, what we expect is that there are going to be more and more basic use cases which will get commoditized and they will be served by this kind of interface.

Uh this is a very affordable, easy to use um agent. So it's just a flat fee on a monthly basis. There is no limitation on how many agents get created and then they can be used for different purposes, whether it's marketing or branding or a set of data analytics, copilots that you're doing and each one can be deployed into specific departments.

So supply chain gets its own one within supply chain. Let's say there are somebody who's doing safety swap, then they get their own thing. There is marketing which is doing some sort of a specific marketing analytics, they get their own pocket agent, there is product teams which are using product documentation, they get their own agent.

So it's kind of an army of a i agents that can be deployed within your enterprise across all your departments on an ongoing basis. And then just available for you and for your it teams to be able to manage that along with the business user groups.

So it kind of brings the full power of generative agents to your doorstep, which you can use. And then the other agents that i've shown earlier were essentially more advanced use cases which cannot be handled by this sort of agent builder in this case, right?

So that's what i want to share with you today. This is available on the aws marketplace and uh which can be downloaded from there directly and you can install it within your own enterprise environment. It's an enterprise only product.

And uh we have a booth here um on the other corner, zs associates. If you're interested to know more, a couple of my colleagues will be there. And um we're happy to take any questions, show you more of the demos if you're interested in. And then, um and then um if, if there is interest, let us know. Thank you very much.