Why Your Company Needs a Context Graph (and How to Build It) — Gil Feig, Merge

Therefore. Thank you to everyone who Therefore. Thank you to everyone who came. My name is came. My name is came. My name is Gil Feig. I am the Gil Feig. I am the Gil Feig. I am the co-founder and co-founder and co-founder and CTO of CTO of CTO of Merge. Merge helps Merge. Merge helps Merge. Merge helps companies easily companies easily companies easily create AI for create AI for create AI for manufacturing. We are the manufacturing. We are the manufacturing. We are the connecting connecting connecting infrastructure to make infrastructure to make infrastructure to make everything work together. everything work together. everything work together. Today I will explain Today I will explain Today I will explain why you need a why you need a why you need a context graph and how to context graph and how to context graph and how to build one. I won't build one. I won't build one. I won't talk about talk about talk about specific technologies specific technologies you should you should you should use. You use. You use. You can discuss this can discuss this can discuss this with your agent. He will with your agent. He will with your agent. He will help you. There are help you. There are help you. There are many good many good many good technologies that technologies that technologies that can can can provide this. The provide this. The provide this. The focus is on focus is on focus is on why you why you why you need a context need a context need a context graph, how to graph, how to graph, how to structure it at a structure it at a structure it at a high level, what high level, what high level, what its components should be, its components should be, its components should be, and how and how and how it all comes together to it all comes together to it all comes together to create the best “ create the best “ brain” for your brain” for your brain” for your company. By the way, company. By the way, company. By the way, if you haven't if you haven't if you haven't received a cap yet, my received a cap yet, my received a cap yet, my team is going around handing team is going around handing team is going around handing them out. They them out. They them out. They say "well say "well say "well connected". So connected". So connected". So let's get started. I won't let's get started. I won't let's get started. I won't spend much spend much spend much time talking about time talking about time talking about Merge, just a quick Merge, just a quick Merge, just a quick overview. We have over 400 overview. We have over 400 overview. We have over 400 corporate corporate corporate clients: OpenAI, Ramp, Netflix clients: OpenAI, Ramp, Netflix , Perplexity, many , Perplexity, many , Perplexity, many companies that companies that companies that use us use us use us for both internal for both internal for both internal and client and client and client AI solutions. We have three AI solutions. We have three AI solutions. We have three different products: Unified, different products: Unified, different products: Unified, Agent Handler, and Gateway, Agent Handler, and Gateway, Agent Handler, and Gateway, covering many of the covering many of the covering many of the components of this components of this components of this system. We founded system. We founded system. We founded the company 6 years ago the company 6 years ago the company 6 years ago with my with my with my co-founder Shen co-founder Shen co-founder Shen Xi. We have raised three Xi. We have raised three Xi. We have raised three rounds of investment.

rounds of investment. rounds of investment. We have 130 employees We have 130 employees We have 130 employees in three of the most beautiful in three of the most beautiful in three of the most beautiful cities in the world. Good. cities in the world. Good. cities in the world. Good. So, a brief outline of the So, a brief outline of the So, a brief outline of the speech. First, we'll speech. First, we'll speech. First, we'll talk about what a talk about what a talk about what a context context context graph is and why it's so graph is and why it's so graph is and why it's so important for important for important for every company to have one. We will every company to have one. We will every company to have one. We will look at different look at different look at different levels of context, what levels of context, what levels of context, what types of data, and what types of data, and what types of data, and what logic your logic your logic your company needs to company needs to company needs to get into the graph. get into the graph. We will talk about the context selection strategy: how do we context selection strategy: how do we select select select the most relevant the most relevant the most relevant content for our content for our content for our agents at any given agents at any given agents at any given time? time? time? Traceability and Traceability and Traceability and origin. How origin. How origin. How to find out where this to find out where this to find out where this data comes from, who data comes from, who data comes from, who requested it, when it requested it, when it requested it, when it appeared, whether it is appeared, whether it is appeared, whether it is up to date, and all that other stuff up to date, and all that other stuff up to date, and all that other stuff . And finally, . And finally, . And finally, a few key a few key a few key conclusions. conclusions. conclusions. I'll stay at the end I'll stay at the end I'll stay at the end if you have any if you have any if you have any questions. questions. questions. So, don't hesitate to So, don't hesitate to So, don't hesitate to approach me approach me approach me after the performance. Okay after the performance. Okay . So, what is a . So, what is a . So, what is a context graph in the context graph in the context graph in the context of this context of this context of this conversation? We won't conversation? We won't conversation? We won't delve into the delve into the delve into the nodes and edges nodes and edges nodes and edges connecting each node connecting each node connecting each node in a true in a true in a true graph graph graph infrastructure. We infrastructure. We infrastructure. We will talk about will talk about will talk about corporate corporate corporate intelligence.

intelligence. intelligence. Your Your Your company's context graph is made up company's context graph is made up company's context graph is made up of many different of many different of many different components. It components. It components. It contains extraneous contains extraneous contains extraneous data. What other systems do data. What other systems do data. What other systems do you use, like you use, like NetSuite or, say, NetSuite or, say, NetSuite or, say, Jira. It has a Jira. It has a Jira. It has a structured structured structured context. This may context. This may context. This may be data be data be data coming from other coming from other coming from other third-party third-party third-party organizations. This organizations. This organizations. This can be data that can be data that can be data that you have added to you have added to you have added to your your your context level yourself. For example, context level yourself. For example, context level yourself. For example, static documents static documents static documents that you want to always that you want to always that you want to always share with your share with your share with your agents. We have agents. We have agents. We have memories. This is something that memories. This is something that memories. This is something that is formed gradually is formed gradually . We all know what . We all know what . We all know what memory is—things memory is—things memory is—things that are formed that are formed that are formed during the use of during the use of during the use of an agent, when you an agent, when you an agent, when you ask it ask it ask it to remember something or to remember something or to remember something or when the agent when the agent when the agent decides for itself what it decides for itself what it decides for itself what it should remember. And should remember. And should remember. And we also have skills. we also have skills. we also have skills. Skills are a very Skills are a very Skills are a very important part of important part of important part of your your your context graph, which we will get to later context graph, which we will get to later context graph, which we will get to later . OK. . OK. . OK. I see someone I see someone I see someone taking a picture. Okay, taking a picture. Okay, taking a picture. Okay, great. So, great. So, great. So, let's move on. let's move on. let's move on. Let's look at an Let's look at an Let's look at an example. We will example. We will example. We will use this use this use this example to explain example to explain why it is not enough to why it is not enough to why it is not enough to simply connect simply connect simply connect one or a couple of one or a couple of one or a couple of MCP servers and consider MCP servers and consider MCP servers and consider it done. it done. it done. So, let's ask a So, let's ask a So, let's ask a simple question.

simple question. simple question. One of your One of your One of your employees or employees or employees or customers asks customers asks customers asks the agent, "Why was Customer A the agent, "Why was Customer A the agent, "Why was Customer A upset upset upset last week?" A last week?" A last week?" A fairly simple fairly simple fairly simple question to question to question to answer. The agent answer. The agent answer. The agent goes to the goes to the goes to the system, sees that they system, sees that they system, sees that they have access to your have access to your have access to your Zendesk, and can Zendesk, and can Zendesk, and can check: "Hey, what check: "Hey, what check: "Hey, what tickets were created tickets were created tickets were created by this customer by this customer by this customer last week?" last week?" last week?" Perfectly. We saw Perfectly. We saw Perfectly. We saw that the API synchronization that the API synchronization that the API synchronization was broken and was broken and was broken and got specific got specific got specific details of what went details of what went details of what went wrong. We also wrong. We also wrong. We also see that the customer see that the customer see that the customer has submitted a total of three has submitted a total of three has submitted a total of three tickets, and that's it. tickets, and that's it. tickets, and that's it. Great, that answers the Great, that answers the Great, that answers the question. But what question. But what question. But what if we want to if we want to if we want to expand this expand this expand this task? What if we task? What if we task? What if we wanted to ask, "Which wanted to ask, "Which wanted to ask, "Which customers were customers were customers were upset last upset last upset last week?" This becomes week?" This becomes week?" This becomes much more complicated much more complicated much more complicated because now your because now your because now your agent has to agent has to agent has to download every download every download every ticket from Zendesk that was ticket from Zendesk that was ticket from Zendesk that was created in the created in the created in the last week. last week. last week. What if we asked, What if we asked, What if we asked, "Which customers have been "Which customers have been "Which customers have been upset in the upset in the upset in the last year?" We last year?" We last year?" We should be able to should be able to should be able to answer that. Our answer that. Our answer that. Our agent needs to be able to agent needs to be able to agent needs to be able to respond, but they respond, but they respond, but they won't be able to if they won't be able to if they won't be able to if they have to make have to make have to make real- real- real- time queries and pull time queries and pull time queries and pull thousands and thousands of thousands and thousands of thousands and thousands of tickets, a hundred at a time, tickets, a hundred at a time, tickets, a hundred at a time, until they time out until they time out until they time out and and and simply can't simply can't simply can't respond. Therefore, MCP does not respond. Therefore, MCP does not respond. Therefore, MCP does not fully solve fully solve fully solve many of these problems many of these problems many of these problems because the underlying because the underlying because the underlying API access patterns of API access patterns of API access patterns of these platforms do not these platforms do not these platforms do not support support real-time search methodology. If real-time search methodology. If you could semantically you could semantically you could semantically query your query your query your ticketing system, if ticketing system, if ticketing system, if Jira provided an endpoint Jira provided an endpoint Jira provided an endpoint where you could where you could where you could ask in English ask in English : "Which customers were : "Which customers were : "Which customers were unhappy unhappy unhappy last week?" and last week?" and last week?" and it would return this it would return this it would return this data, that would be great data, that would be great . But no platform is . But no platform is . But no platform is interested in interested in interested in this because this because this because vectorization costs vectorization costs vectorization costs them money and reduces them money and reduces them money and reduces the number of the number of the number of visits to their visits to their visits to their platform. So you platform. So you platform. So you need a need a need a synchronization layer on top of synchronization layer on top of synchronization layer on top of that. You can't

that. You can't that. You can't just just just use MCP use MCP use MCP to live search for to live search for to live search for things you want things you want things you want to dig into or to dig into or to dig into or query query query semantically, you semantically, you semantically, you need to actually need to actually need to actually sync a copy of the sync a copy of the sync a copy of the data locally. data locally. data locally. Of course? This does Of course? This does Of course? This does not apply to everything. not apply to everything. not apply to everything. For example, you For example, you For example, you might have static might have static might have static requests to Stripe. We will requests to Stripe. We will requests to Stripe. We will always always always ask Stripe ask Stripe ask Stripe about a specific about a specific about a specific customer. We will always customer. We will always customer. We will always perform one perform one perform one action in Stripe. It could action in Stripe. It could action in Stripe. It could be MCP. This can be in be MCP. This can be in be MCP. This can be in real real real time. But if you time. But if you time. But if you need any need any need any kind of semantic kind of semantic kind of semantic search, any search, any search, any analysis that goes analysis that goes analysis that goes beyond a simple beyond a simple beyond a simple query but requires query but requires query but requires looking at a looking at a looking at a huge huge huge data set, you should data set, you should data set, you should synchronize the copy synchronize the copy . So this becomes the . So this becomes the . So this becomes the next part of next part of next part of your your your synchronization layer or synchronization layer or synchronization layer or your, uh, your, uh, your, uh, context layer. context layer. context layer. So you have your MCPs, So you have your MCPs, So you have your MCPs, your live queries, and your live queries, and your live queries, and your your your synchronized or synchronized or synchronized or cached queries. And the ones cached queries. And the ones cached queries. And the ones you get from a you get from a you get from a third party, you third party, you third party, you put into a vector put into a vector put into a vector database locally. database locally. database locally. Again, AI can Again, AI can Again, AI can create this vector create this vector create this vector database for you relatively database for you relatively database for you relatively easily. Uh, you easily. Uh, you easily. Uh, you need to work need to work need to work on a strategy for on a strategy for on a strategy for how you store data how you store data how you store data in the context layer in the context layer in the context layer and how you retrieve it and how you retrieve it and how you retrieve it from that from that from that synchronization and synchronization and synchronization and caching layer. But that's kind of the caching layer. But that's kind of the caching layer. But that's kind of the next next next component here. But the component here. But the component here. But the next problem is: next problem is: next problem is: okay, now we have okay, now we have okay, now we have all this relevant all this relevant all this relevant data. We extract data. We extract data. We extract data from our data from our data from our vector database to vector database to vector database to get semantic get semantic get semantic information, and we information, and we information, and we query in query in query in real time for real time for real time for other data we other data we other data we may want to may want to may want to add. Now we add. Now we add. Now we need to pull all need to pull all need to pull all this data together because this data together because this data together because we can't just we can't just we can't just dump the huge dump the huge dump the huge amount of amount of amount of context our context our context our agent has received. This is where we agent has received. This is where we agent has received. This is where we introduce introduce introduce the router and the the router and the the router and the generalizer.

generalizer. generalizer. The router The router The router first sends first sends first sends things to the right place things to the right place . A request is received. He . A request is received. He . A request is received. He decides who should decides who should decides who should process it, where it process it, where it process it, where it should be sent. Once should be sent. Once should be sent. Once all this data is collected, all this data is collected, all this data is collected, our our our summarization tool takes it summarization tool takes it summarization tool takes it from the synchronized from the synchronized from the synchronized and current and current and current context, context, context, summarizes it, and summarizes it, and summarizes it, and sends it back to sends it back to sends it back to the agent who the agent who the agent who sent the request. But sent the request. But sent the request. But you all know this, you all know this, you all know this, we have one thing we have one thing we have one thing we haven't talked about yet we haven't talked about yet - it's memory and skills - it's memory and skills . They can . They can . They can exist separately. Why is exist separately. Why is exist separately. Why is this important for your this important for your this important for your level of context? level of context? level of context? The reason is that The reason is that The reason is that getting context getting context is actually a set of is actually a set of is actually a set of corporate corporate corporate principles and processes. principles and processes. principles and processes. For example, which For example, which For example, which customer was customer was customer was unhappy unhappy unhappy last week may last week may last week may mean something completely mean something completely mean something completely different for one different for one different for one company compared to company compared to company compared to another. This could be the another. This could be the another. This could be the number of open number of open number of open tickets. But what if tickets. But what if tickets. But what if these tickets were these tickets were these tickets were positive? What if we positive? What if we positive? What if we don't use a don't use a don't use a ticketing system and ticketing system and ticketing system and we know that to measure we know that to measure customer satisfaction, we always customer satisfaction, we always need to need to need to look at Qualtrics first and look at Qualtrics first and look at Qualtrics first and check check check survey data because that's how survey data because that's how survey data because that's how our company works? our company works? our company works? We are different. That's why We are different. That's why We are different. That's why skills are important, because skills are important, because skills are important, because you need to be able you need to be able you need to be able to determine if to determine if to determine if a customer is satisfied.

a customer is satisfied. a customer is satisfied. So now your agent So now your agent So now your agent has to first has to first has to first test the skills test the skills test the skills before moving before moving before moving to the synchronized to the synchronized to the synchronized or current level or current level or current level to figure out what to figure out what to figure out what data to retrieve, data to retrieve, data to retrieve, summarize it, and summarize it, and summarize it, and return a result. return a result. return a result. So, the process is this: So, the process is this: So, the process is this: a request comes in, it is a request comes in, it is a request comes in, it is routed to the routed to the routed to the right agent. right agent. right agent. The agent checks whether The agent checks whether The agent checks whether he has memory or he has memory or he has memory or skills related to skills related to skills related to the question or current the question or current the question or current task. Based on task. Based on task. Based on this, it knows where this, it knows where this, it knows where to look in the cache to look in the cache to look in the cache database and through the current database and through the current database and through the current MCP requests to MCP requests to MCP requests to summarize and summarize and summarize and send a response. There is send a response. There is no no magic here. She really is magic here. She really is magic here. She really is n't there. It's all physics. n't there. It's all physics. n't there. It's all physics. If you know how If you know how If you know how to calculate data to calculate data to calculate data within your within your within your use case, you'll use case, you'll use case, you'll be fine. This be fine. This be fine. This process, where we again process, where we again process, where we again looked at the number of looked at the number of looked at the number of open tickets, would not have been open tickets, would not have been open tickets, would not have been possible possible possible with just current with just current with just current requests. But with requests. But with requests. But with local search, it local search, it local search, it was very simple: we was very simple: we was very simple: we were able to collect all were able to collect all were able to collect all customers, the number of customers, the number of customers, the number of their tickets, their tickets, their tickets, summarize them, and summarize them, and summarize them, and provide a status for provide a status for provide a status for each. So, it's a each. So, it's a each. So, it's a combination of current combination of current combination of current queries and queries and queries and cache lookups. All of this together cache lookups. All of this together cache lookups. All of this together forms very rich forms very rich forms very rich data for any data for any data for any question. OK. There are question. OK. There are question. OK. There are different levels different levels different levels of context, and one of context, and one of context, and one might argue whether might argue whether might argue whether query skills and query skills and query skills and memory are context.

memory are context. memory are context. Context is everything that is Context is everything that is Context is everything that is passed to the agent. passed to the agent. passed to the agent. We'll put it all We'll put it all We'll put it all here because it all here because it all here because it all belongs to your belongs to your belongs to your level of context. So level of context. So , you have , you have , you have query skills and memory. We query skills and memory. We query skills and memory. We talked about how talked about how talked about how they will determine they will determine they will determine where we get our where we get our where we get our data from? You have your data from? You have your data from? You have your live API calls. They are live API calls. They are live API calls. They are easier to create, easier easier to create, easier easier to create, easier to add, and to add, and to add, and cheaper to operate than a cheaper to operate than a cheaper to operate than a synchronized synchronized synchronized context. So, you context. So, you context. So, you should should should prefer them when you prefer them when you prefer them when you know you will be know you will be know you will be dealing with simple dealing with simple dealing with simple queries and searches. queries and searches. queries and searches. Cached data is Cached data is Cached data is absolutely essential absolutely essential absolutely essential if you need if you need if you need that that that level of semantics, complex level of semantics, complex level of semantics, complex search, or deep search, or deep search, or deep queries about queries about any data. But any data. But any data. But keep in mind that it's keep in mind that it's keep in mind that it's a little more complicated a little more complicated a little more complicated to set up and to set up and to set up and build. This is build. This is build. This is more expensive to operate more expensive to operate more expensive to operate because you need to because you need to because you need to have a vector have a vector have a vector database and constantly database and constantly database and constantly synchronize this synchronize this synchronize this data. So it's harder data. So it's harder data. So it's harder to keep them up to to keep them up to to keep them up to date. date. date. Finally, something we have Finally, something we have Finally, something we have n't discussed yet is n't discussed yet is n't discussed yet is derived context, derived context, derived context, which is any which is any which is any additional context additional context additional context you create when you create when you create when retrieving data from retrieving data from retrieving data from third-party third-party third-party developers. It might developers. It might developers. It might look like this: we look like this: we receive all tickets in real time, receive all tickets in real time, and every time a and every time a and every time a full full full set of set of set of customer tickets arrives, we customer tickets arrives, we customer tickets arrives, we summarize them and summarize them and summarize them and store a summary of the received data in our store a summary of the received data in our store a summary of the received data in our structured structured structured context context context . So . So . So in the future, if a in the future, if a in the future, if a common common common question arises, we won't question arises, we won't question arises, we won't have to have to have to search again and search again and search again and summarize everything all over again summarize everything all over again summarize everything all over again . Perfectly. So, . Perfectly. So, . Perfectly. So, all four all four all four components work components work components work together. You saw together. You saw together. You saw the diagram of how they the diagram of how they the diagram of how they do it. And yes, these are those do it. And yes, these are those do it. And yes, these are those levels. Here is a quick levels. Here is a quick levels. Here is a quick flowchart of how it flowchart of how it flowchart of how it works when a works when a works when a request or prompt comes in

request or prompt comes in request or prompt comes in . The first question . The first question : do we have a match in : do we have a match in : do we have a match in skill or memory? skill or memory? skill or memory? If so, then we want to If so, then we want to If so, then we want to perform this skill. perform this skill. perform this skill. Does the skill require Does the skill require Does the skill require external data? If external data? If external data? If not, we simply not, we simply not, we simply generate a response generate a response generate a response and send and send and send it back immediately. Unfortunately, it's it back immediately. Unfortunately, it's it back immediately. Unfortunately, it's not always that simple. not always that simple. not always that simple. So what happens So what happens So what happens if we don't have a if we don't have a if we don't have a match in skill or match in skill or match in skill or memory? None? Well, memory? None? Well, memory? None? Well, then let's move then let's move then let's move straight to straight to straight to data selection. The same if data selection. The same if data selection. The same if the skill requires the skill requires the skill requires external data. So external data. So , we say "yes," and the , we say "yes," and the , we say "yes," and the first thing we do is first thing we do is check our cache check our cache check our cache or synchronized or synchronized or synchronized data store and data store and data store and ask: is this ask: is this ask: is this data up to date? Do data up to date? Do data up to date? Do they meet they meet they meet our SLAs? Can we our SLAs? Can we our SLAs? Can we use this data use this data use this data because it is considered because it is considered because it is considered fresh enough? fresh enough? fresh enough? If so, we If so, we If so, we use this use this use this cached context. cached context. cached context. No? Then, unfortunately, we No? Then, unfortunately, we No? Then, unfortunately, we have to go have to go have to go lower and call the lower and call the lower and call the live API. Once we live API. Once we live API. Once we have enough have enough have enough context, we will do context, we will do context, we will do several iterations until we several iterations until we several iterations until we have the have the have the necessary base. necessary base. necessary base. Once we have enough Once we have enough Once we have enough context, we can context, we can context, we can store in memory store in memory store in memory everything we just everything we just everything we just found out or found out or found out or understood, then understood, then understood, then generate a generate a generate a response and response and response and send it. Same send it. Same send it. Same here: if the context is here: if the context is here: if the context is not enough—I not enough—I not enough—I just showed this—we just showed this—we just showed this—we need to call the need to call the need to call the active API and active API and active API and then return the then return the then return the generated generated generated response to the client.

response to the client. I'll leave this here for a I'll leave this here for a second. I see some people second. I see some people second. I see some people watching and watching and watching and taking pictures. OK. taking pictures. OK. taking pictures. OK. So, the last So, the last So, the last element, almost element, almost element, almost at the end: at the end: at the end: traceability and traceability and traceability and provenance of data. provenance of data. provenance of data. At first, it doesn't At first, it doesn't At first, it doesn't seem important. Don't seem important. Don't seem important. Don't lose information lose information lose information about the data source. about the data source. about the data source. Always keep it. Always keep it. Always keep it. You get it through You get it through You get it through MCP. This is also in everything MCP. This is also in everything you synchronize you synchronize you synchronize in vector databases—it's in vector databases—it's called called metadata. You metadata. You metadata. You should keep it should keep it should keep it there. This is related to the there. This is related to the there. This is related to the data itself. This is data itself. This is data itself. This is critically important. This is critically important. This is critically important. This is similar to how it similar to how it similar to how it works in ChatGPT or works in ChatGPT or any other any other any other interface: when it interface: when it interface: when it states a fact or states a fact or states a fact or claims something, claims something, three sources are listed at the end three sources are listed at the end with with with links to them. links to them. links to them. This is even more important when This is even more important when you are making you are making critical critical critical business decisions based on this data. When business decisions based on this data. When business decisions based on this data. When this data has versions. this data has versions. this data has versions. When this data may When this data may When this data may become outdated. When you become outdated. When you become outdated. When you allow the system allow the system allow the system to make decisions to make decisions to make decisions based on data based on data based on data synced a year synced a year synced a year ago. You don't ago. You don't ago. You don't want that. So, you want that. So, you want that. So, you need to know need to know need to know where they came from. where they came from. where they came from. Where did we get them from Where did we get them from ? When did we get them? ? When did we get them? ? When did we get them? This is not here. Under This is not here. Under This is not here. Under what identifier what identifier what identifier and within what limits of and within what limits of and within what limits of authority? Who had authority? Who had authority? Who had access to the data? What access to the data? What access to the data? What permissions permissions permissions were used were used were used for access? Were for access? Were for access? Were they transformed? What did they transformed? What did they transformed? What did they look like they look like they look like at first? And lastly: at first? And lastly: at first? And lastly: what should be the what should be the what should be the basis for their basis for their basis for their annulment?

annulment? annulment? For example, the statute of For example, the statute of For example, the statute of limitations. Data limitations. Data limitations. Data older than a month—we older than a month—we older than a month—we shouldn't shouldn't shouldn't consider it. This is extremely consider it. This is extremely consider it. This is extremely important for important for important for display in your display in your display in your platform, and if not platform, and if not platform, and if not for display, then for display, then for display, then for for for incident investigation when something incident investigation when something incident investigation when something goes wrong. And it will goes wrong. And it will goes wrong. And it will definitely go away. definitely go away. So, to summarize, So, to summarize, here's another example of what here's another example of what here's another example of what we can we can we can use. " use. " Give me a brief Give me a brief Give me a brief update on the status update on the status update on the status of our client Acme." of our client Acme." "Include "Include "Include contract renewal risk contract renewal risk contract renewal risk and last and last and last support interaction." support interaction." support interaction." First, we First, we First, we will check the cache for the will check the cache for the will check the cache for the account summary account summary account summary . Let's say there are . Let's say there are . Let's say there are no no no special skills here. special skills here. special skills here. So, we'll So, we'll So, we'll check it out. If check it out. If check it out. If the data is not available, we the data is not available, we the data is not available, we will contact CRM using a will contact CRM using a real-time API request. real-time API request. We will then We will then We will then query the query the query the ticket system for the ticket system for the ticket system for the last interaction last interaction last interaction with support. Again with support. Again with support. Again , this is cached , this is cached , this is cached within the SLA, otherwise you within the SLA, otherwise you within the SLA, otherwise you would have to would have to would have to query in real query in real query in real time. We time. We time. We only want to get five only want to get five only want to get five tickets. We cannot tickets. We cannot tickets. We cannot pull the entire pull the entire pull the entire history because there history because there is a risk of timeouts with real-time queries. And finally , we will generate a , we will generate a , we will generate a generalized generalized generalized conclusion. We can conclusion. We can conclusion. We can save this save this save this conclusion and pass it conclusion and pass it conclusion and pass it on. So, the on. So, the on. So, the main conclusions: main conclusions: main conclusions: tools are not tools are not tools are not context. The context. The context. The context graph context graph context graph will help you will help you will help you analyze analyze analyze information across your information across your information across your entire entire entire company. It's not just company. It's not just company. It's not just tools. This is data.

tools. This is data. tools. This is data. This is synchronized This is synchronized This is synchronized data. These are the processes that data. These are the processes that data. These are the processes that exist in your exist in your exist in your business. Who business. Who business. Who documents what we documents what we documents what we do when a customer do when a customer do when a customer files a complaint? How files a complaint? How files a complaint? How does the system understand where to does the system understand where to does the system understand where to direct direct direct the request? It's a skill. These are the request? It's a skill. These are the request? It's a skill. These are your your your company's skills. These are your company's skills. These are your company's skills. These are your processes. These are your data, processes. These are your data, processes. These are your data, tools, and actions tools, and actions tools, and actions that need to that need to that need to be taken. More be taken. More be taken. More data isn't always data isn't always data isn't always better. You need better. You need better. You need fresh and high-quality data. fresh and high-quality data. fresh and high-quality data. You need to get You need to get You need to get only what is necessary. only what is necessary. only what is necessary. When you When you When you synchronize data, synchronize data, synchronize data, structure it structure it structure it properly for properly for properly for correct access. correct access. correct access. And finally, every And finally, every And finally, every response must be response must be response must be traceable. You don't traceable. You don't traceable. You don't want to be in a want to be in a want to be in a situation where the AI situation where the AI situation where the AI made a bad made a bad made a bad decision and you don't decision and you don't decision and you don't know why. So, the know why. So, the know why. So, the last thing is a little last thing is a little last thing is a little advertising. advertising. advertising. Corporate AI Corporate AI Corporate AI works for us. We works for us. We works for us. We provide these provide these provide these opportunities for opportunities for opportunities for many companies. many companies. many companies. We have a Gateway—a We have a Gateway—a router for LLM. router for LLM. router for LLM. We have Unified, which We have Unified, which We have Unified, which helps helps helps synchronize data synchronize data synchronize data from your clients' third-party from your clients' third-party from your clients' third-party platforms platforms platforms . We also . We also . We also have Agent Handler—out-of-the-box have Agent Handler—out-of-the-box have Agent Handler—out-of-the-box MCP connections for MCP connections for MCP connections for hundreds of tools, hundreds of tools, hundreds of tools, with managed with managed with managed identity, identity, provenance verification, and plug-and-play connectivity for both plug-and-play connectivity for both internal and internal and internal and client-side client-side client-side tasks. So, if tasks. So, if tasks. So, if you want to create a you want to create a you want to create a convenient level of convenient level of convenient level of context, context, context, contact us.

contact us. contact us. Thank you all for your attention.