AI Pulse: May 2026

Santiago Giraldo: Hello, everybody, and welcome to Snowflake AI Pulse update for May. This is a series where we're going to be talking about the latest features and releases for everything Snowflake AI now and into the future. Just a reminder, this is a monthly series, so be sure to tune in every third Tuesday of the month for the latest and greatest from Snowflake. Next slide. Very quickly, the mandatory safe harbor. Everything that you see here is subject to change, but we're excited to share lots of progressive things with you. Next slide here. I'm joined today by a very broad panel of speakers that are experts basically in everything that we're gonna be talking about. My name is Santiago. I'm the Senior Director of Product Marketing for Artificial Intelligence and Machine Learning. I'm joined by Samir, who is our Senior Product Manager here at Snowflake. Shankar, our principal architect for Microsoft and Snowflake and data superheroes. Jesse, our staff product manager here at Snowflake. And Neeraj, another senior product manager here at Snowflake. All these teams have been working very hard on the latest implementations and maintaining the velocity and innovation that Snowflake is known for. So next slide. I'm going to go ahead and level set a little bit on what we're going to be talking about today. So I'll start off with just giving us a quick overview of everything that's been released over the course of this month and everything that's going to be available today. After that, I'll hand it off to my colleagues here who will walk us through some deep dives and demos of everything from AI security to our partnership with Snowflake and Microsoft, and then a deep dive into our latest releases across AI functions. And then we'll end with a live Q&A. So let's start off with what's new. These are the coolest features hot off the press for us. So first and foremost, we've released our AS Security Trust Center. This is a centralized control plane to discover and remediate security identity governance and compliance risks for users and agents. It helps you streamline across cloud security monitoring and put that into one central place that has a lower total cost of ownership. There prevents escalation and overall make sure that you have full visibility over your entire domain and all of your data. In addition to that, you can detect security violations and also near real-time anomalies, enabling you to act as quickly as possible in case of any emergency or any downtime for your systems and organization. I'm happy to announce that this is now available in GA globally. Next up, we have our partnership with Microsoft, where we have a Cortex Agents integration. What this integration basically is, it's a native integration between Microsoft Copilot Studio and Snowflake Cortex Agents. It's delivered through Snowflake's managed MCP server, and makes it very easy for you to interoperate these two systems. Um, what it enables you to do is build conversational agents that combined snowflake data and cortex agents with your Microsoft ecosystem seamlessly. You can build agents powered by court. You can build agents powered by cortex agents inside of Microsoft Co pilot studio and then access them directly inside of Microsoft 365 Co pilot teams. This is also available as of today in general availability. And then on the AI function side, we've got a couple of major announcements in terms of capabilities and how we've invested and expanded in our AI SQL and our AI functions efforts. First up is AI parse document. This function supports extracting of embedded images. It also brings support for over 2000 page documents and comes with very notable quality and speed improvements. AI Complete now accepts PDFs, Excels, and Office documents as direct inputs. And then Cortex Fine Tuning. It brings support for Arctic Extract, which enables everybody to be able to extract models specific to their domains. As of today, all these features are generally available. And then the second piece of our AI functions update are some features that are in public preview today that you can begin to test out almost immediately. First and foremost is AI Extract, which now supports confidence scoring, source citations, and extra long 1,000-page documents. AI Classify now supports document modality natively. And AI Complete now natively supports video and audio analytics, essentially making AI functions here at Snowflake the first organization to be able to power almost every source of information that's coming across from structured to unstructured video and audio all in one place. As I mentioned before, this is all in public preview today. And without further ado, what we want to begin to do is dive more deeply into each one of these areas. So I'm going to go ahead and hand it off. Hand it off to Sameer for RIA security. Sameer?

Samir Kakkar: Hey, Santiago, thank you for giving me my favorite product, a warm introduction. Everybody, I am part of a very large security team at Snowflake, and I'm proud to present some of the work with you all that we have been doing here at Snowflake. Next slide, please. The way we think about AI at Snowflake as it pertains to security is twofold. We think about it as AI for security, using AI capabilities to help improve customer security postures, compliance, governance, both internally and externally. And the second, which is on the right side of this, is security of AI itself. My talk today is focused on the left side of it, which is AI for security. But I have some great teasers for you as to what we'll be announcing very shortly at Summit. So you have to hear the presentation till the very end to hear about that. But when it comes to AI for security, our approach has been to integrate AI workflows into existing customer user journeys and make the art of possible be real. Which in the past, customers had to either build bespoke controls or just wait on us to deliver new features. Next slide, please. So as Santiago was describing, that we have this great native product within Snowflake accounts, which runs natively for all of you without any license to install, any things to configure. This native product, which runs on its own, comes back and tells you how best you can improve your security brochure. It's a single pane of glass that customers can use through UI, APIs, SQL. But now, and that's what Santiago was talking about, now we have AI enabled it. So this is the GA feature Santiago was referring. And this creates a new usability dimension as to how customers interact, but not just with Trust Center, but with every security feature available to customers. So our RBIC controls, our network controls, our encryption, all of them are now possible through Cortex Code skills, and this talk will go into a little bit deep into that. With CoCo, which stands for Cortex Code, that all of you must be familiar with by now, it will help customers understand the security posture in natural language. They can ask things which previously would have taken them reach out to a security expert. Now the security expert is available to them 24 by 7. Next slide, please. So imagine I'm a platform admin and I would like to understand my security posture. Trust Center gives me a great overview of things that I should address, but I'm very interested in figuring out what are my top risk or my CISO asked me to look at something. or my security team asked me to go investigate a particular risk on the platform. In the past, I would have looked for menus, I would have looked for SQL statements, but now I have CoCo to help me. Let's start the demo. One of the things that you will see in this demo which is running and I'll be talking over it is that customers can come and when they see Cortex Code running or Citra Center running, besides the UI, they can ask the natural language question. Here what I'm asking is, tell me the five top users I should be concerned about as it pertains to finding. CoCo quickly goes in, figures out through all my findings, which are the most five users I should be concerned about. Then I ask it, I'm concerned about the privilege, tell me about what's my privilege issues that I have in my account. It gives me that summary. Now I go back and look at my critical risk and say, okay, let's go back and start fixing my critical risk. So I identify a critical risk. I invoke CoCo through it saying, let's go and resolve this critical risk. This particular risk, which I clicked on, the most critical risk I had in the account was with network policy. It came back and said that the policy is not configured, so I go and say, okay, help me configure the network policy. As part of that, CoCo runs through it and not only examines the risk, but give me the exact statements that I can use to fix that as well. I just tell CoCo, yes, help me identify it and also remediate this. It goes and identifies which all IP addresses I've been logging in in the last 30 days, creates a network slash firewall rule out of it. In my case, this began demo account, there were just two IP addresses. I select those two IP addresses. Yes, that looks right. It's a demo account. I understand. Let's go and start creating a network policy for me. Previously, creating this would have required me to spend minutes, hours, days even. I look over the command by CoCo. Yes, that seems like the right command, and I allow it to be run. Look at this copilot way for me to examine what CoCo is telling me, decide I want to implement it, and I say, yes, go if you file it. I'm actively remediating a critical violation in my account. Obviously, CoCo is smart. It will go back and check what are the required ways to enable this control. It comes back and checks. There has to be a database to which this object has to be attached. It identifies via database. I decide to choose that particular database, which is great, and now I'm creating that object. But this is a network policy. If not done right, I could be logged out of my own account from logging in remotely. So now I'm asking CoCo, OK, go step by step. Show me everything you're doing. And I'm looking at each and every command as it is being proposed. I like what I'm seeing, I'm rolling with this. And here it tells me, look, I want to make sure this is a critical step that my current IP address through which I'm logged in is covered by this network policy. It verifies it, I says, yes, go ahead and activate the network policy. And at this point, the network policy rule is being implemented as we speak. All of this done without me trying to ask somebody. CoCo is there as my co-pilot. I run the exact statement, network policy rule has taken place, and now this violation which was reported in my account should be addressed. But the violation is typically addressed when I rerun the scanner or when the scanner runs on its own. In this case, the scanner is configured to run every day. Once the scanner runs, this violation I'm showing on my screen should go away. But you know me, I just can't wait. So I'm just going to come back and try and run the scanner again. But it takes a while after the scanner run for the violation to be resolved or to be reflected. It can take from minutes to sometimes an hour. So you have to be a little patient. While this is going on, I decide to make an audit trail so that when my auditors come and I have to generate a report for it, I can tell them how I remediated something. I looked at the remediation steps, I copy and paste them, and I put them in my in a very cheat sheet or notes which I have in my own Trust Center account. I copy and paste it from here, and I go to my Trust Center, Findings Lifecycle Management, and I take a note. So next time somebody asks me about this finding, I know what I did. I can put any statement. I can put a JIRA link. I can put a ServiceNow ticket number. It's all taken care of. And I can even mute modifications because I've taken care of this finding. Now, as I'm waiting for it, I read on the scanner, I come back, and I notice, yes, I had 22 findings before, now I have 21. That particular finding has been addressed. That finding is taken care of. But you know me, I'm like Dory. I sometimes forget what I did. So I go back and ask CoCo, hey, CoCo, I know there was a finding here we were remediating. What happened to it? So I go back to CoCo, it has all the notes. It comes back and then tells me, you remediated it, Sameer. Here's the steps you took. Here's how you remediated it. So that's a very great example of how we have infused AI, interest-centered, helping our customers improve their security posture. Now you might say, and this is the next slide, please, that Sameer, you made it work for an account, a demo account with just two IP addresses in it. Will it work for a large account? And I will say, absolutely. So this next slide, and if you can run the demo for this next slide. And I'm not seeing the slides, the demo move, but yes, there you go. This is a big account where hundreds of users log in. This is a real account, hundreds of users logging in. Poco runs through it, finds 10,000 IP addresses. It goes back and slices and dices them, puts them in the right CIDR range, comes back and tells me, hey, this is a very complex thing. Here's maybe you need a hybrid policy, maybe you need more than one policy, and gives me the right information for me to go tackle this real life 10,000 IP address logging in accounts. So great, right? It works for simple case, it works for difficult case. But is this the only thing you can do with CoCo or security skills? No, you can do a lot more things. So let's take an example of a Cortex Code scale, which we call as the accessible shooter in this next slide. And here we show a demo of that you guys may run into a whole lot that you are accessing something in Snowflake and you don't have the right access to it and you want to go investigate it further and you start doing it. So as soon as you type your natural language command, the right skill for that, which is the access troubleshooter skill, that gets fired up. And for any kind of errors that you might have in the account, you could either remediate it or you could understand it. So it untangles the complex web of complex roles, hierarchies, and it comes back and tells you the right role that you need to access something. In this case, there are three things I need to do to get the access that I want to do. Think about any complex scenario you're running into or even simple scenarios where you want to know, I want to access this object, what needs to happen? Or I need to give Joe access to this object, what least privileges can I give to Joe? All of that is taken care by access troubleshooter skill. I gave you just two or three examples of how you can use security skills with Trust Center or without Trust Center. Now, I also promised you something about the teaser, about what's coming at Summit as it pertains to using AI or using security of AI. So this next slide, please. I'm not going to talk a whole lot about this, because you have to come to the Snowflake Summit to hear about all of these great things. But at a very high level, if you're making agents on Snowflake, now you can go on them with these privileges. You can ensure your agents don't have more privileges than the user that they're running on behalf of. You can have data movement policies to control that agents are moving large amounts of data. You can put restrictions on it. You can proactively stop them or you can proactively monitor them. Obviously, you will have a great view interest center as to look at your entire AI state. All of those great things are coming at summit. You'll have to wait for the summit announcement and I'll be happy to be back here talking about all those great things with you. So thank you for listening in. I would love to pass on to you the next speaker, Shankar Narayanan, our data superhero to talk about Microsoft. Shankar.

Santiago Giraldo: Thanks for your patience, everyone. Having some technical difficulties. We'll get Shankar on in just one moment.

Santiago Giraldo: I'm going to hand it off here to Shankar in just a minute to do a deep dive of the Snowflake and Microsoft implementation. If we go to the next slide, I think we can just dive right in. This is a great demo. Excited to pass it off. Of course, Shankar and myself here will be around for Q&A on all this.

Shankar Narayanan SGS: Thank you for having me in this. I'm Shankar. I lead the Snowflake ISV team here at Microsoft. Today we're going to talk about the innovations that Microsoft and Snowflake is doing across our AI ecosystem for our customers. Starting with our flagship integration with Copilot Studio. Copilot Studio now can natively connect through Cortex agent using the new MCP Snowflake and CP servers. With this, customers can make use of the data in Snowflake, and the Cortex agent to build complex multi-agent orchestration toolkits and systems that can be directly deployed to Teams, web apps, and all the other places where the business users can consume from. Extending that is going to be our MC65 and Teams integration. This brings the power of Cortex agent right within the teams, like a knowledge worker. So business users can actually start asking questions to the Cortex agent and get the response back in a text format, right? Supporting both unstructured and structured data right within Snowflake. And on the third segment, we are also powering the Snowflake Cortex using the state-of-the-art models using the Foundry. Customers can use the models in Foundry, build agents in Foundry powered by Cortex, as well as leverage the Azure OpenAI services directly within Snowflake Cortex for building advanced complicated agents. And today, we are going to see a short demo of how we are going to integrate the Copilot Studio to Snowflake using the MCP connector. To begin with, Copilot Studio is our flagship low-code, no-code agentic platform that customers can use to develop state-of-the-art agent models by using drag-and-drop interfaces. It supports both code model as well as the UI model where you can build the agent, test the agent, evaluate agent, and as well as deploy it in various different publishing work schemes. We have a support for integrating MCP services, and we are going to be leveraging the Snowflake MCP server in order to integrate this. And for today's demo, we are also going to talk about the US Cold Storage, which is one of our cool customers who have leveraged this particular integration in order to reduce the friction they had for their business users in consuming both inventory and operational data. They wanted to have a place to democratize data access and have all the non-technical users get instant access to data without needing to open a report, build or write a SQL query, or find other means to expose the data. They were able to consolidate all of the data within the Snowflake AI Data Cloud, build a bunch of Cortex agents on top of it, use the Copilot Studio to orchestrate the old agentic platform, and then implement and roll it out directly to Teams so that their end users can just open and chat with an agent just like how they are chatting with other employees around. They were able to showcase this completely and provide the democratized data access right to their business users so that they can start making insights directly without waiting for the entire report load or without waiting for the entire data to completely appear for them. So in today's demo, we are going to be talking about the Copilot Studio integration using the MCP server. We'll be building a small Cortex agent that is powered by both the Cortex Analyst and the Cortex Search, supporting both structured and unstructured data. And this is going to be exposed as an MCP server tool, which is then consumed by the Copilot Studio. So over to the demo. So if you see on my Snowflake, I have an agent, right? And I did use Snowflake CoCo Cortex Code in building this specific agent, right? This contains two artifacts, right? Two tools. One is the Cortex analyst tool, which is the sales metrics data and the Cortex search tool, which has the sales conversation that sales agent has done with specific customers. And on top of this, I have exposed it and created an MCP server adding this specific agent Agent as a tool and this MCP server is what we will be consuming and using it right within the copilot studio in order for consuming through auth and And if I go back to my Copilot Studio environment, I have already added the tool within the Copilot Studio for leveraging the Snowflake MCP server. And it is straightforward. You just add the information, provide the MCP server, provide the OAuth details. And the details step is provided on the blog that is attached with this particular session. And I'm going in, and I'm going to start asking a specific question. Let me clear the session. Now I'm gonna start with the unstructured data. I want to find out about what one of my customers call went by, right? So I'm asking a specific question. What it does in turn is it's going in, finding out the MCP tool, which is the sales intelligent agent that we have in Snowflake, calling that specific MCP tool, and then let Cortex agent do its magic in pulling that unstructured data. And then it's gonna return a structured response for us to observe and act on. So it's right now processing the data. And if you see, you can get the data. It's returning the data back out, right? It's giving me the detailed highlights, the detailed specific instruction, and the list of action items that have happened from that particular call between the agent and the customer. And it also suggested me the follow-up questions that I can ask around this in order to find more business-related information. Now switching context, let's go and ask a question that leverages the Cortex analyst using the semantic views. I'm going to put a question that asks about the average sales by customer. This in turn is going to get translated by the Cortex agent into a text to SQL, and that SQL query is going to get generated in Snowflake, and the response is going to be returned back for me to directly see and consume that data from. It's again going through the same process of calling in the MCP server and using that MCP server in order to process the data, get the response from the Cortex agent, and then give me the response back into my interface. If you see it showing me the data, which has the largest one and which has the smallest one, it also gives me the key insights as well as followed by questions that I can ask specific to this. If you see this is very powerful in integrating this agent of Copilot Studio within Cortex, consuming the data from Snowflake within Cortex and orchestrating the entire workflow directly within Copilot Studio. You can also extend this for all the other sources with your operational data along with the data in Snowflake, right? As well as you can go in and also build the Copilot Studio skill, right, that is now available. You can use the skill along with the GitHub Copilot and the Snowflake CoCo in actually building the Copilot Studio workflows, right, using your AI agent, right? With this, I thank you for the opportunity and as well as I'm going to be sticking around for the live Q&A as well.

Santiago Giraldo: Thank you, Shankar. Up next, I'm going to be handing it off to Jesse and Neeraj, who will be taking us through the latest innovations in AI functions. Jesse, Neeraj.

Jessie Felix: Thank you, Santiago, and thanks everybody for being with us today. I'm a product manager with the Snowflake AI team, and I look after the capabilities that use AI to transform unstructured data. And today, we're going to walk you through some of the latest innovations with our AI functions in multimodal and document capabilities. Next slide. So at a very high level, I want to share some high-level trends that we're observing. So a very interesting trend is that the share of total data stored is increasingly dominated by visual and auditory data. And this is really driven by how different users engage with products and services through different content. So if you look at the slide just in the last 10 years, we see that video jump from accounting for 10% of total share of data to 35% of total share of data just in the last 10 years. And I think a big trend that I'm starting to see across every organization is that they're starting to adjust their analytics strategy to really account for this new reality. And the key thing here is like really trying to understand how the users are actually using and living with the products. So there's three broad level use cases that we're observing modern organizations focusing on. Next slide. So the first thing is all about social listening at scale. So this is really about businesses moving from understanding what people say in surveys to understanding how people actually use their products. And a key thing there is that by analyzing how users actually use their products, you're able to identify trends before they really hit mainstream and be able to really craft campaigns that resonate more with how users actually use different products. And the next thing is really around brand protection at scale. So for organizations where moat is really defined by the brand equity, protecting that brand equity is everything for your organization. And you really want to really understand what is the sentiment around your brand, how is it actually being used across different content creators, so that you're able to intervene and actually protect how your brand is viewed across different users. And the other trend is really about extracting trends and generating statistics. And a top use case that resonates and I see more and more is stat generation in sports. So we all know if you're watching football or basketball that we love stats and stats drive fan engagement. And different organizations know this, and there is a lot of focus on creating more interesting statistics, but not just in sports. It's really like extracting different metadata from different content that you're able to join with your structured data in Snowflake and get different insights that help you build better predictive models for engagement, for example. And given all of these new use cases that we're observing, we wanted to make that easy to do inside of Snowflake. So in the next slide, what you will see is that in the same way that we made it easy for customers to be able to process text, We're now enabling customers to also be able to natively process video inside of Snowflake using our AI complete function, which is one of our AI functions in Snowflake. So whether you need to analyze brand sentiments, summarize advertising trends, or extract insights how customers actually use your product, you can actually do this all with a simple SQL statement. and process videos that are stored in an internal stage or an external stage. And when files are stored in an external stage, such as object storage, whether that be S3 or Azure Blob Storage or Google Storage, for example, you don't need to copy those files over to Snowflake. All of those are processed in place, so there's no data movement whatsoever. We simply process those files and all of that metadata is stored in a Snowflake table, which you can then use for downstream analytics all inside of Snowflake. So in the next slide, what we're going to get into is a visual demo of this. And I'll talk over the demo as it begins to play. So the demo is meant to showcase an end-to-end multimodal analytics pipeline in Snowflake, a fairly simple one. But some of the key things here that I walked through is creating a stage, In this case, I'm using an internal stage and using our file data type, which I'm using right now. For your context, the file data type provides a way to reference files that are stored in stages from documents, images, files, for example, and video files. It really makes it possible to use a file location and the metadata that we can actually operate on unstructured data just as easily as structured data. Now the workflow itself, which I'm showing here now, it's a demo which is, I have different advertisement files and what we're gonna do is that we're gonna try to understand what type of advertisements for different pet products are being generated. So I'm gonna generate a summary or very specifically generate structured summaries and thematic classification for each advertisement. And I'm using the Gemini function to do this. And here it already ran. And we see that we have a summary per each video file. And again, everything that is processed within the stage, so the AI operates on the files, the extraction happens, it's stored in Snowflake, and you really only have to run once. And then once you have the summaries, then you can very easily use this for other structured data analytics. And here I simply flattened the data so we can have a more tabular view of the insights. And the summary span from health of dog nutrition. So I having a dog myself, so I thought it was pretty interesting, like all the different trends around social media and animal care. And last but not least, what I'm doing here is using all our AI functions. Now that I have multiple summaries per video, I'm going to do a meta summary that is grouped by the different themes. So for example, we have health and nutrition, we have um healthy dogs etc and you can get a quick analysis of like what are the top themes so we see health nutrition premium ingredients are the top themes and then creating a summary of summaries that shows us what are the the main uh the main themes that are identified across all of those videos and for your context i processed in this query specifically roughly 500 videos using the Gemini model. And now that I have these insights, I can then generate different ideas about how to better improve a targeting campaign, like really aimed at pets. So for example, this last insight that was there was around reinforcing that your animals are like family and that messaging really resonating with different users. In the next slide, one key thing that I want to emphasize is that the AI complete function and when you're working with Snowflake in general and our AI capabilities, it's really a model garden and you have the ability to choose the best model for the tasks that you are performing. So what you have here effectively is the support for different modalities across different models. So whether you're looking to process text or images, documents, audio, now video, the aim is for you to have any model that you can choose and easily with a very simple SQL statement, update the model that best fits the task that you're trying to do, and very easily update a pipeline without you needing to manage multiple APIs, for example, with other providers.

Jessie Felix: All of this can happen natively in Snowflake, and we manage all of that infrastructure and that overhead for you.

Jessie Felix: So now passing it on to Neeraj, who's going to talk to us about the latest in our document intelligence space.

Neeraj Jain: Thanks Jesse for the overview of multimodal Cortex-AI functions. Hello everyone, my name is Neeraj and I will continue the journey with Document Intelligence use cases. So let me walk you through what we have shipped in Document Intelligence space and how our customers are using it. As you can see from here, these are the broad key themes we see our customers building using Cortex-AI functions. The first theme is around enterprise search, where customers who have a purpose of large set of documents want them searchable, whether it's for RAG, semantic search, or agentic-powered Q&A. Second broad use cases is around structured extraction, things like invoice processing, KYC, forms processing, or extracting important structured information from contracts and other documents. And third broad use case is around research and analytics, where customers typically want some kind of reasoning over the content of the document to interpret clauses and sections and compare sections within the document as well as across multiple documents. And as I'm going to show you, these use cases are going to be very easy to set up and production ready with Cortex-AI functions. Next slide, please. So now let's talk about the key AI functions that power these use cases. So we have these four document specific AI functions, right? And each of them plays a very distinct role. The first and the most important is AI pass document. It converts any document to a layer of a text preserving all the tables and sectional headings. The second function is AI Extract. It pulls structured specific fields and tables along with their confidence codes and citations. Third AI function is AI Classify. It is used for routing documents by type or any other attribute. And finally, we have AI Complete, which is used for open-ended reasoning by your chosen LLM. As we will see in the next slide, these functions compose together to form an end-to-end document intelligence platform. Next slide, please. As you can see, The documents to be processed by the AI functions, they typically resides in an external blob storages like S3 buckets or Azure blob storage, or it could be stored in Snowflake internal stage. The output of these AI functions is then fed downstream, which can be used in customer facing interfaces like Snowflake intelligence, Cortex agents, or any other BI applications. As you can see in the last section of the slide, you can compose or chain these functions together. As an example, AI parse document and AI complete is chained together for a research and analytics use case. Now what we can do is we can go deeper with individual AI functions. Let's start with AI parse document. In this foundational AI functions, what our customers get is the ability to bring large corpus of document and make it usable for their downstream processing. And very recently, we shipped three core capabilities to do it more efficiently and more powerfully. First is the support for extra long 2,000 page documents. Second is the support for native image extraction, which are present in the document. And finally, very important, significantly improved handwriting recognition, character recognition, and understanding of complex multi-column documents. So what this means is when a customer in just a large document like an annual report or a clinical record, which has, you can imagine text and tables and images and all of it, you can use a single AI SQL call to ingest all of this in a single go to your Cortex search or Cortex agent use case without doing any kind of pre-processing and you preserve the full fidelity of the document. Next slide, please. AI Classify, as you saw earlier, of all the use cases we talked about for document processing, typically what happens is, in a customer's environment, these documents are landed in common places in a mixed format. And before you can do anything, you need to know what document you are dealing with. So AI classifies is used as this common routing block, which sits in the front of every pipeline. And as you can see, it now directly supports document modality, where you can ingest the document and it will use the full multimodal context of the document to classify it. And based on the classification, you can route the document to the right pipeline. Next slide, please. Now, let's talk about the second use case of structured extraction. AI extract is used when customer wants to transform their unstructured document into a structured data, which they can further query and analyze on. AI extract recently shipped with three new capabilities. First, very important, is confidence scores. Now, every field extracted will also come with a confidence score. Second, every field extracted will also come with a bounding box, which is a citation, and it will point to the source of that information in the document. And together what we have is a human in the loop workflow. That means you can combine these two functions together to create a human review process where no confidence extraction can be sent to a human reviewer who can directly see where the extraction came from in the document. and make it an enterprise grade extraction experience. In addition, AI Extract, just like AI Parse, supports extra long documents of 1,000 pages. Next slide, please. For those customers who are dealing with domain specific use cases with respect to their document format or content like abbreviations and terminologies, we recently launched the ability to fine tune their extraction model, ArchiExtract, using their own label examples, all within the security and governance parameter of Snowflake. And finally, at the end, next slide, please. Here is a quick reference for what each and every function we just talked about does, what kind of information it outputs, and what are the use cases it's suitable for. And these are the building blocks we just discussed which customers can use for their document analytics use cases. I hand it back to Santiago.

Santiago Giraldo: Excellent. Thank you so much. And with that, we have some time to move into Q&A. So I'm just going to go ahead and dive right in. I'll help moderate some of these questions. And folks on the panel here, please turn on your cameras and help dive into some of these questions live here. So the first question here is, when will AI Complete have more output tokens than 8,192? Felix, do you want to?

Jessie Felix: Yeah, so that's a... So different models support different optical token lengths. So, for example, some LAMA models, rightly so, I think the max context window length is one... um 8192 but some other models like our anthropic models can even go up to 124 000 tokens so it's really dependent on the in the model uh i think if there's a specific ask for a specific use case that we can better enable um passing that feedback through the account team and reaching out to the team here would be a good next step

Santiago Giraldo: Excellent. Awesome. Next question that we have here is, can we use our own models within Cortex with an enterprise version of Snowflake, of course? I'm happy to answer that one, unless somebody else wants to take it. But the short answer is yes. You can bring in pretty much any model directly into into the Cortex infrastructure. It can be used with Cortex Code. It can be deployed and used with Snowflake Intelligence or in apps or through the API. We have a full suite of machine learning products to help you manage that model, maintain that model, operate that model, fine tune it, and everything in between. Anything to add there, folks? OK, great. Next question, anyone else experiencing lag? No, sorry. When someone starts on Snowflake, should they just start in AI slash Cortex to learn about fundamentals or learn about fundamentals via training? Um, I would say it depends on your role. So snowflake is a near zero admin platform that really takes care of the data governance, the data security, the data management piece of it. Cortex then enables you to do the work on top of that. Um, things like Cortex Code make it very, very easy to be able to do. the full lifecycle of all of your data, soup to nuts from raw data curation to pipelines to AI-ready data to delivering results, whether that be through AI functions, intelligence, Cortex Code applications, down the line. So it is quite intuitive. I think that learning the entire stack is always good, but really you can do a lot of the work directly inside of the natural language interfaces with little setup and only some foundational knowledge. OK, let's see here. I have a question here. Is this activity being carried out in the production environment? If so, should there not be an established approval process in place? And I believe that this question came in when we were talking about AI security. So Samir, I'll hand this off to you. Samir? Okay, so I guess maybe some of you are having some audio problems. Sorry, I'm back.

Samir Kakkar: Yeah, no, I'm just curious. This question was about the production environment. Yeah, the demo, I had a couple of demos running there. But yes, two of the demos were from a production environment used internally. But yes, there's 10,000 IP addresses, network policy. That's a real production environment used internally.

Santiago Giraldo: Awesome, excellent. Then the question here is, is it able to provide graphical representations of data from Snowflake in Copilot? This is a Microsoft question. I will hand that off to Shankar if you're here.

Shankar Narayanan SGS: Yeah, I'm here. Can you guys hear me? Yes. Sorry, I'm going through a storm here, so I'm on and off with multiple connections. I'm back in. So yes, you can use it to generate charts, right? You can ask the co-pilot studio to generate charts or graphicals based on the data that Cortex agents sends back, right? So we do support it. It's an advanced parameter that you need to enable as part of the agent configuration. And that should allow you to basically have those agents respond back in a graphical chart rather than just a text or a table form.

Santiago Giraldo: Awesome. Good answer. Here we have another question. In one prompt, can Cortex find all the relevant information and provide a response based on data across multiple connected data tables? Short answer is yes. So that can be done through Cortex Code with one prompt in natural language. It's an incredibly powerful system that can rationalize all of your data, rationalize semantics that you either input or begin to generate semantics on its own. It can contextualize data and pretty much do whatever you need with it while taking multiple connections into account. OK. I had a question here. When should you use AI Complete instead of AI Extract? I think best practice question here.

Neeraj Jain: I can take this one. As we discussed earlier, AI Compute is better when your extraction is powered by reasoning. And AI Extract is more recommended solution when you want to extract directly based on the prompt you have inputted. Awesome.

Santiago Giraldo: OK. Can we use Cortex AI to connect to Confluence or SharePoint or ServiceNow for information retrieval? Anybody want to take that one?

Neeraj Jain: I can take this one as well. Snowflake has published a lot of Openflow connectors. For any third-party sources like Confluence, what you can do is you can use Openflow connectors to bring those documents to an external or internal stage. Post that, you can process those documents using any of the AI functions you choose.

Santiago Giraldo: Awesome. Okay. Here's a longer question. Recently during the trust review, we were advised to reduce the maximum programmatic access tokens, the PATs, expiration period to 90 days. Currently, all our service accounts are configured with 180-day PAT expiration that are included in the MFA override policy. What integration options or compensation controls can we implement to address this compliance violation? Felix, any thoughts on that?

Samir Kakkar: Take an attempt. Santiago, could you just repeat the question or just point me to it? I couldn't get the last part.

Santiago Giraldo: Sure. So the question here is around trust reviews, where this company was advised to reduce the Mexican programmatic access tokens, the PATs, the expiration period to 90 days. All the services are now configured with 180-day PAT expirations included in the MFA override. The question here is, what can we do to mitigate some of that and get it to compliance?

Samir Kakkar: Yeah, so you will definitely require to regenerate those PAT tokens or auto-rotate them every 90 days. But if you're looking for a monitoring control inside Trust Center, you could write your own extension and have a monitoring control running inside Trust Center, which will flag you those accounts. Or you could even run queries ad hoc to detect that. But if you're talking about customization, those are the two options I can think of right now.

Santiago Giraldo: Okay, awesome. So we have another one question here that I think is relevant to a lot of folks that may not have been on Snowflake for some time or maybe have stepped away. The question is, what's the best way for someone who has been away from Snowflake for the last few months to get up to speed with Snowflake Cortex-Coat? and the other AI suites, such as the new function capabilities, Snowflake Intelligence and beyond. Is there one-stop training available for AI? This is a bit of a two-parter. The best way to really start experimenting with Cortex Code is to get hands-on. We have a series of quick starts, lots of articles, lots of repositories that you can begin to play with. It's available out of the box as soon as you open up your Snowhouse instance. And it can be accessed through the CLI very easily just with a quick install. It's in our documentation. So we'll be happy to send out resources to get you going. The system is incredibly easy to use. So just diving in and asking questions is really the way to start. The other piece here is, is there a one-stop training available for AI? As I mentioned before, we've been really focused on expanding our training capabilities. Right now, we're focused on what we're calling Cortex training, which enables large language model building and fine-tuning and modification. We have a full notebook environment. uh powered by jupiter notebooks that helps you do any training in any way that you'd like and you can also build and train advanced and sophisticated models directly in natural language through Cortex Code so the short answer is yes snowflake is the one-stop shop for doing everything ai agents as well as any training that you need to do on your own models bespoke models or modifying existing or open source models that you might be using OK, I think let's take one more. And then I think we're pretty much out of time. Let's see here. One question is here, can multiple agents be used for a single request, or will the request always use only one agent? And feel free to chime in, folks. But in Cortex Code, we enable what basically is planning mode, which lets you run multi-agent workflows directly inside of the platform to achieve complex tasks in any way that is required, depending on the complexity of the prompt and the solution that's being that's being looked for. In addition to that, we recently released agent teams, which basically enables you to have sub-agents that execute tasks for you in a coherent way in order to achieve those complex tasks as well. Anything to add there, folks? Okay, so I think we'll pause there. Thank you, everybody, for your time. As a reminder, this is a monthly series where we'll always be talking about the latest and greatest being released regarding AI at Snowflake. Please tune in every third Thursday, and we'll see you next time.