Atlan + AWS: Building Context-Aware Enterprise AI Agents

Razi Shafiq

0:00 All right. Thank you everyone for joining us today for another Atlan in Action webinar. Today we actually have the head of product for Amazon Quick, Shekhar, here with us joining us from Amazon. And what we're gonna do today is take you through the philosophy of context with Atlan, how that fits into with Amazon Quick, and then show a quick demo at the end.

0:31 So it's building context aware enterprise AI agents. From the Atlan side, we have Chandru, who is our director of product, as well as myself, a partner sales engineer, and I will be taking us through the demo

Chandru Sugunan

0:48 Hey. Hello, team. Good morning, good evening, good afternoon, everyone. Thank you for joining our webinar series. All right. So I just wanted to say that I'm sorry for a little late. I have a internet is-issue kids issue. Here we are. So let's get going.

Chandru Sugunan

1:13 Today I wanted to start by saying every enterprises is going through have this hundreds of agents. They're gonna run these hundreds of agents across dozens of platforms. And we have been talking to many customers and in some of the customers' environment enterprises, it's already happening.

1:37 And if you look at this map, as you can see in the right, there are a bunch of copilots you know, vertical AIs, low-code builders, and it keeps growing on and on. One of the CAOs that we discussed this issue with have quoted that everyone wants to be their agentic platform and, you know, they expect to have dozens of platforms and if not hundreds of agents at steady state.

2:07 And in action, the way it looks like something like this. Agent one it thinks revenue as a term in the company means bookings. Agent two may think revenue means ARR, something from finance, and maybe agent three doesn't even know what revenue means. And as the old data joke goes, "Ask finance and sales for revenue, you get two different numbers."

2:31 Now, every agent is going to do the same thing, and it's gonna compound. The problem is going to compound at scale. And an assistant in every app multiplies this fragmentation. You can, you can, you, you know, like, And it's same it's true for the context also. And if every agent with its own definition say revenue means different things, and it multiplies and it causes more confusions.

2:58 And we cannot fix this problem by picking one platform, and this is where making context open really helps. You know. I'm curious how many agents platforms are in your company as we speak. Just please drop the number if you can. The way we think about the simple truth is context isn't one thing.

Chandru Sugunan

3:22 It's a collection of things: knowledge, expertise, norms. And let's look at what's knowledge. Knowledge means actually things. What is true. Entities, definitions, metrics, relationship, the actual map of your business. And expertise is how a work gets done, the playbooks, the workflows, how you break down the tasks.

3:46 And the norms is what are allowed, the policies, the permissions. So most AI conversations stop at, you know- The first one, the knowledge. But the agents to be successful needs all three

4:00 So the AI is complicated. As you, as you can see here at practice, this is how it would look. A store manager asks, let's say, "What's the reason for the drive-thru numbers going up this week?" It's a simple question, but the data is probably sitting in your, you know, AWS estate. And actually, the data isn't the problem.

4:23 You have the right data, to be honest, right? But the knowledge is what here. The knowledge is what does drive-thru time means. Where is this column is even sitting? It's not the POS number finance uses. It's the actual operational number. And what does this week mean? It's a term. Like, for you, that this week is Monday through Sunday, and it's not a rolling seven days.

4:48 If either gets wrong, your answers or the agent's gonna confidently be wrong coming up with the response. The expertise. The expertise here is validate your premises, the seasonalities. Is there any launches of any food items or whatever that drive-thru is, weather, local events? The best analyst in your company will think about all these parameters.

5:12 Unless you tell the agent to think about these parameters, it's not gonna think about it, right? And finally, the norms. The store manager sees his store, the VP of Ops sees his chain, and so on and so forth, right? There are six pieces of context has to coming together in order to answer this simple question, and none of this actually lives in the table, the data itself.

5:37 I'm also curious to know what do you think about this knowledge, expertise, and norms. You know, if you have any comments, thoughts, please drop in the notes. And finally, the way we think about this context path is crossing these three walls. The first wall is the agent bootstrapping wall, the context for the agent bootstrapping like it's basically a Fortune five hundred CDL as building the agent for them took about five minutes, but giving it a business context took five months.

Chandru Sugunan

6:12 Another global bank they stood up like thousand Genie rooms. Ninety percent of those are abandoned because the business didn't trust them. And this is why we do a AI-led bootstrapping for the context. And the next wall you will hit is how do you get to three to thirty agents, the life cycle and the governance?

6:34 Who owns the context and who approves it? Who maintains it? You know, the marketing position changes. The agent still pitches the older version. Nobody knows what else depends on it. And when an agent gets it wrong, you can trace it, which context caused it. So this is how we think about the life cycle and the governance of the context for the agents.

7:01 And the wall three is the sprawl and the drift. Every platform has its own way. Nothing flows back to the shared layer. One of the feedback from a CDO from Fortune five hundred is that they don't want their IP the context as an IP to be logged in one of these platforms. So I'm curious again, like which wall-- if these walls make sense, and which wall you are in at this point.

7:35 And finally so and this is exactly why we think AI context platform is crucial to build the agentic flows, a shared enterprise brain. On the left, you know, where the context lives today, the system of knowledge, system of records, the data, the policy. In the middle, you can see the you know, the AI-ready data, the semantics and ontology around it, the agent skills underneath, the memory what is an active context graph that feeds and improves everything about it.

8:14 And then we mine the context from it, from the lineage, the connectivity that we have, we govern it, we observe it, deliver it to the agent, and further close the loop. The agent learns with the evals and flows back into the whole flow. And with this central context layer or, you know, centralized context layer the copilots, the analytical agents, the operational agents, everything consumes the same context.

8:45 All the serious agent platform is building semantic understanding and, you know, Shekhar will show further in the quick how he thinks about it. And the last stage is how we mature. How do customers actually get there? This is like a four-step process. The way we think about it, it's like stage one, which is the AI-ready data.

9:10 Can a human or an agent find the right data, use it correctly? This is where our lineage, data quality, governance, SQL intelligence functions come in place. The stage two is the AI-ready business meaning. How do you interpret your business, the business terms, ontology the glossaries without inventing it, without changing it?

9:34 The semantic models and metrics for analytics agents are available for the agentic flow. And the stage three is the context engineering. Can the context do the work? Right procedures are in place, the tools are in place, and build the skills for the for the agents, the catalogs, and agentic operating procedures.

9:57 And finally, the self-improving context. Can the agent you know, write it back the traces, the corrections, the shared memory? So this is how we think about a four-step process, four-stage process to mature your context. And with that I would like to hand off to Shekhar. Please you know share your thoughts on how you think about it in AWS and Amazon Quick.

Shekhar Kopuri

10:31 Thanks, Chandru. Hello, everyone. Great to be here. I'm Shekhar. I lead the product for structured data and analytics within Quick. Quick is For those who are unfamiliar with Quick is a fairly broad agentic platform where our focus is how do we put the power of AI in the hands of end business users within an enterprise, but without compromising what you, the tenets of a great enterprise application stack which is, you know, security and governance are well handled.

11:21 So today I'm gonna just take a like, take a step back from where Chandru left. A key aspect of utilizing AI in an enterprise is, of course, the context that the AI agent is subjected to. But I want to sort of first maybe introduce Quick to you and then tie it back on how the integration with Atlan sort of works out.

11:53 As Chandru said all the organizations are already using AI. There are greenfield agents that are built by typical business units in the organization. Data is still fragmented and although each agent is built to do a great job at a particular task or a set of tasks the fragmented data, the context switching of the end user from, you know, tool A to tool B, maybe all of them powered by the AI, still creates a bottleneck in terms of the value that the end users, the business users, the analysts in an enterprise get out of the out of out of their AI deployments.

12:52 So we built Quick to sort of think a little differently. Quick is a unified AI platform that allows you to bring all of your enterprise context, potentially the team's context, some of it the personal context, all of that together under a single layer. The Quick platform is model agnostic, and it still has all the guardrails in place that ensure security and governance.

13:32 You're not dealing with multiple product setups to enable AI across all job functions inside the company. And yes, you get a bunch of compliance certifications out of the box when your users utilize Quick. So what are all the aspects that we have inside Quick? You have a common shared knowledge foundation that powers everything.

14:09 Quick comes with its own orchestrator. It's a multi-agent orchestrator. We have our own harness. And then we have a layer of features that allows end business users to do a whole bunch of activities starting from the simple chat, ask a question all the way into conducting really deep research that you can let the agent go about and sort of mine volumes of data, including on the Internet, and build deep research reports and everything in between.

14:49 Quick is also a service that has a full-featured BI and analytics foundation layer built into the product, and therefore enterprise structured data, which is coming out of data lakes paired with enterprise context, allows Quick to build accurate answers for the questions that a person might have and actually take those answers and deploy them into actions.

15:23 The actions may look like simple automations. It can actually be as complex as full-fledged workflows that can be automated, similar to how RPA works. Or an end user just using natural language can curate their own custom AI agents, starting from the data to the context to the final actions that they want to carry on, which translates data and context not into just decisions, but actually downstream business impact.

16:04 Quick also supports the ability to build custom apps. So now your biz ops challenges that typically that you would go and talk to your IT to build automations is also something that we democratize and put that in the hands of the business users Quick is a multi-surface product. So today literally a few weeks ago, we launched Quick for desktop and Quick for mobile.

16:37 So you can, as an end user, you can initiate a task or you can spawn a task associated with an agent on your behalf on a desktop and the activity happens remotely on cloud runtime. Therefore, you can literally switch over from desktop to mobile. Web works standalone today, but shortly we are going to introduce the same unified sort of backend runtime that powers the web also.

17:12 So it becomes something a service that an end user can continue and make the power of AI operate for them irrespective of what their task is during the day and what surface that they are using. Quick also has a whole bunch of extensions that you can utilize. Because one of our foundational philosophies is that you allow users to operate in the product that they spend most of their time on without compromising the governance and security layers, and more importantly by utilizing the same unified platform, no matter what the final usage surface looks like.

17:58 So you could be in your Word and you could be writing your document, you could ask Quick a question, and Quick will give you the exact same answer or the visualization or the summary that it would give you on your mobile or a desktop, and you can literally plug and play all of those insights into your document and continue to utilize the Quick agent to adjust your document and improve your productivity.

18:30 So Quick like Atlan, comes with lot of connectors. These connectors span from structured data connectors all the way to ERP softwares, data analytics you know, connectors. We also have vertical-specific connectors. So foundational piece of Quick being a common platform or a common ecosystem is, is this ever-increasing sort of ecosystem of connectors As I said the end user can build agents for themselves.

19:06 The you know, you can write your own custom applications, and you can do things like, you know, write automation tasks and literally take over the onus of a classic sort of analyst to not go to their IT for solving routine but tasks that reduce productivity in their primary shop or their primary stakeholders

19:39 So now let me tie this back to what Chandru sort of started this conversation with. How do we think about where context what context powers in an enterprise? And Quick tries to be the AI platform that facilitates the connectivity from your data sources. And we typically think of data sources as your structured data sources or knowledge bases, but even your email is a source of data as far as the end user's personal context is concerned.

20:23 So your data sources could be your RAG that's built with your Wiki pages or other knowledge bases. It could be your deeply governed data warehouses on structured data that has critical data, and all of that form a semantic layer in Quick. Now, if your enterprise-- as an enterprise, if you have utilized Atlan to curate the semantics of your own businesses or your own enterprise at scale, then Quick ingests those semantics into its own graph and then associates that with the personal knowledge graph or the personal context that it discovers and curates constantly on your behalf as an end user.

21:19 And then it powers various AI applications that end users can just go and build themselves within a governance layer that your enterprise engineering teams or your BizOps teams can configure to the degree that they want to give access and empower specific business units, specific personas within a particular company, or all the way to the line sort of workers within a company.

21:55 So this is where data is deeply important to utilize and connect context coming from products like Atlan, which you might utilize to configure and curate your entire business's business context as well as structured data context. So with that, I would invite Razi to actually show this whole setup in action and sort of showcase, give you a taste of what Quick can do for your enterprise with Atlan.

Razi Shafiq

22:31 Perfect. Thank you. And so with that, we're actually gonna jump into our demo portion. And what we're going to do here is show how to curate and collect context in Atlan and then activate it inside of Quick. So with that, let me start here in Atlan. As with anything we're actually gonna start here in our connector marketplace, and this is where we would be able to connect into all the different places that you would like to see within Atlan.

23:08 So things like Amazon products, Redshift, QuickSight, S3, et cetera as well as other pieces. You can think of your CRMs, your ERPs like SAP, Salesforce, et cetera. Once we connect into all of those different systems, we actually collect the metadata and create what we call our enterprise data graph.

23:28 And what that actually looks like, I'm gonna show a node on that graph, is not only the context and metadata around it, but we'll also start collecting lineage and how these pieces are being transformed. So here's one example. This is a customer three-sixty table in Redshift, so this is one node on that graph.

23:48 And if I click into it, I'm able to see the description, I'm able to see the columns, as well as things like read-mes, et cetera. So these are metadata pieces on what this is, how to use it how to protect it, et cetera. But from here we also create end-to-end column-level lineage. So here I am in our lineage view, and we can see that this is that customer three-sixty table, and I can go ahead and open this up all the way upstream as well as downstream.

24:19 If I track this upstream, we can see that this data actually started here in Postgres, was loaded into S3 and then from S3 back into Redshift here. A few different transformations happened in Redshift, and then ultimately this ended up in QuickSight here. This is available at the column level, so if I select a column, I'm able to see where that specific column is coming from, how it's being transformed and where it's going as well as if I wanted to see transformations in between, I can actually see the SQL that exists in between these tables as well Now, one question that always comes up for us is we have hundreds of thousands of tables or millions of tables, columns, et cetera.

25:06 How can we actually curate them without doing a lot of manual work? And that's where Atlan's context agents come into play. What we do here is as we're collecting and connecting to those different sources, we automatically bucket the information into different collections. We also allow customized collections as well, but what this allows us to do is within this collection run Atlan's context agents.

25:32 So here's an example of one of those agents, our description writing agent. If I wanted to add all this documentation, I would just click enrich now. And what this agent actually looks at is not only the schema and naming conventions, we start to look at the relationships, the lineage and transformations.

25:52 Where is it coming from? How is it actually being transformed? We'll look at query history to start to understand what the teams are actually asking of this data today as well as your metrics in your glossary. So using all of that information, our agent actually goes in and writes that description.

26:13 What used to be a very manual process for the governance team and a team of data stewards going into each individual piece, writing that description, writing that readme, et cetera, we have now curated with Atlan's context agents. And this isn't limited to just description writing. We can do things like SQL intelligence.

26:35 SQL intelligence is actually looking at the SQL being run to understand what are popular joins, filters, relationships between tables, between columns, et cetera. Finally what we're also doing here is continuing to build out these agents. So one example would be our metrics agent. We can actually push that metrics agent into all of this information to start to write out and create your business glossary of your terms and your metrics.

27:05 So now instead of having a data steward or someone in the business go in and write all those metrics Atlan can actually discover them and create them for you. So here's an example of one that was done. If I go into our finance glossary here, fiscal year. So here you can see we have defined fiscal year.

27:25 And what's important here is, in this case, for this company, fiscal year runs from October to September, so not the calendar year, January to December, it's October to September. And why this is gonna be important, we'll show here in Quick very shortly. And so with that, let's actually move over to Amazon Quick to show how all of the context that was collected and curated in Atlan can actually help agents answer these questions.

27:54 And I'm gonna jump into two different agents here. I'm just gonna ask this specific question: What was our revenue for fiscal year twenty-four? This is the Atlan plus Quick agent, and then I have one that's just Quick, and I'm gonna ask the exact same question. And while those run we'll go into an overview of what you can do with Quick, what it looks like, et cetera.

Razi Shafiq

28:21 Shekhar I'm gonna let you take this away.

Shekhar Kopuri

28:25 Yeah. Well thanks, Razi. So as I was saying in Quick, as end business user, be it an IT user like let's say a data engineer or a BIE in a, in a, in a classic enterprise, or be it an actual marketing analyst or a sales analyst you can utilize most of the AI features directly using natural language.

28:50 The default you know, landing is, is a chat agent. If you click on New Chat, New Chat is basically gives you a way of just talking to Quick as an entire surface. Quick has its own hardness and its own knowledge graph that it con-con constantly builds. So even if you don't actually curate a specific agent and try to bou-bound its bind its context specifically Quick is able to give you answers out of the box based on its understanding for you as an individual of your combined context.

29:30 We have various con-constructs to contain context, so there are things like spaces where you could build a space, attach structured data, unstructured data, specific third-party connectors you know, add separate instruction documents to nudge the behavior of agents that are operating on this space.

29:54 You could run deep research again, you know, that is attached to a space, attached to a, you know, single instance of source or utilizing web also as a source. End users can build what we call as flows. Flows is simple sort of automation tasks that you can put together as a workflow.

30:18 These are built using natural language. It can have reasoning tasks, computational tasks, discoveries tasks. As long as the connectors are in place that your enterprise has opened up for you, you can literally go and automate a bunch of common context switching or sort of application switching tasks that you perform utilizing flows.

30:44 Like I mean, I use this using browser automation to do stuff that traditionally does not happen without backend integrations across systems. It uses my auth and my identity, acts on my behalf in my full control, and therefore I can automate a lot of the mundane things my team works utilizing flows.

31:09 It's widely used in in the customers where we deploy. And Quick does have traditional BI the entire BI stack. So you can build your datasets, you can build your cross dataset definitions called topics, and you can add additional context if you're sort of not tied into an application or a service layer like Atlan.

31:37 In this case, because Atlan provides the context but the query that the user-- the system executes on behalf of the user ultimately goes to wherever the data is. In this case Razi's datasets are essentially on Redshift. So the definition of those datasets are created in Quick also. Razi has also created a topic here which is just putting those datasets into a single construct that allows the system to know that these are related to each other in some way, right?

32:18 And that some way can be discovered in runtime through integrations like Atlan, or you can actually also nudge the behavior of the systems by curating custom instructions and topics. And of course, there is more stuff like building apps and whatnot. Here in this particular demo, what Razi has done is actually created two separate chat agents.

32:45 Oh, yeah, sure, connectors. Quick, as I was saying, has a lot of connectors. So they stem from all kinds of connectors from analytics engines to actions like ServiceNow and you know, Jira to your emails and WhatsApp and all kinds of sources, both input and output connectors, right? And so what I was saying was, in this particular demo, Razi created two chat agents.

33:16 If you look at those chat agents, they're very simple definitions. There is one chat agent, this is nothing but a wrapper. So it actually has an Atlan connector mapped to the agent itself. Everything else is auto-discovered based on your past usage of the product or the assets you have access to.

33:40 So in this case, Razi has access to those datasets, Razi has access to those topics. So ACLs or access control or access permissions are evaluated in runtime, and only those assets that you have access to are utilized as sources of data if you go to the other chat agent, Razi, I mean, it's basically an empty definition at this point of time, which means you are utilizing Quick out of the box and the knowledge graph that Quick built on, on the enterprise's behalf to discover you know, the who has what permissions and also what data sources do they have access to give an answer is all coming just out of the box.

34:26 So okay. With that, I'll sort of, Razi- Yeah. Give it back to you.

Razi Shafiq

34:31 And so with those two chat agents we asked all that same question: What was our revenue for fiscal year twenty-twenty-four? So here's that first agent without the Atlan context. It returns an answer but what it did here was assume that our fiscal year was January to December.

34:51 And if you remember as we showed in Atlan our fiscal is actually October to September. So now if I jump over to the Atlan version of this, you can see it also gave an answer. The answer was different, and here are the pieces that it actually looked at. So it looked at fiscal year, determined that it was October to September that's the live definition in Atlan, also how we calculate revenue, and then used those pieces to write that SQL into Redshift and get the correct answer.

35:26 So in this way, we can see as the context is updated and changed inside of Atlan, it will stay live within Quick and your chatbots, your chat agents will continue to use that information as it's proprietary to your business. With that, that is our webinar and demo for today. We will stay on for the next a couple of minutes to answer any questions that we do have in the chat.

35:55 But thank you for joining us for another Atlan in Action with AWS.

Razi Shafiq

36:00 All right. We have one question in the chat. Another question arises, which is in the case of conflicts, what are the rules for memory updates? I'll pass that over to both you, Chandru and Shekhar if we want to jump into that question.

Shekhar Kopuri

36:21 So as the end agent that's making the query to the system in Quick you can tune the behavior of as I posted in the chat, you had tune the behavior of the chat agent that you're operating on to consider the certified definitions.

36:41 And the memory management has multiple sort of layers to it. You as the end user can push specific instructions to the memory. You can literally say, "Quick, remember this for me," and it'll subject that to the long-term memory of the agent for you. You can apply similar behavior to the chat agents that you deploy.

37:07 You can in desktop and in web, we have introduced skills which are going to show up in the Sorry, desktop and mobile which will show up in the web shortly. You can basically put the instructions to nudge the behavior to say what aspects do you subject to long-term memory, more importantly, what you don't want the agents to do, which is how typically agents sort of operate.

37:36 And so you have various levers to manipulate the behavior of both the memory and the agent.

Chandru Sugunan

37:43 Yeah. Yeah, no, that's a great response. And from Atlan's side, the way you can think about is the harness is the Quick, but the context is federated from Atlan. Like, okay now what does that particular term mean in this case fiscal year?

38:01 Like the term for the harness is delivered at scale to the agent. So that's how Atlan plays in this role.

Shekhar Kopuri

38:10 That's right.

Razi Shafiq

38:11 Perfect. I believe those are... A-all the questions now have been answered either through the chat or through the answers that we just gave. And again, thank you everyone for joining us for another Atlan in Action with AWS here.

38:27 We hope to see you again, and a copy of this recording will be sent out to the email that you had registered with. Thank you.