Sponsored by: Salesforce | Elevate Agentforce experience with Databricks Agents and Zero Copy

Summary

Zero copy integration: Agentforce connects with Databricks without data movement, enabling real-time data access for agents.
Cross-platform agents: Databricks agents automatically delegate tasks, eliminating workflow switching and delivering seamless user experiences.
Unified workspace: See how Agentforce and Databricks create an intelligent workspace where data, decisions, and actions converge.

FAQs

What is zero copy integration between Agentforce and Databricks?

This video explains that zero copy integration allows Agentforce to access real-time Databricks data without moving or duplicating it. The integration is seamless and efficient, enabling agents to query and act on data instantly while maintaining security and governance.

How do Agentforce and Databricks agents work together?

This video demonstrates that Agentforce automatically delegates tasks to a Databricks agent when data access or complex processing is needed. This delegation happens transparently, enabling end-to-end execution without users having to switch between applications—eliminating the 'swivel-chair effect.'

What is the 'swivel-chair effect' and why is it a problem?

This video describes the swivel-chair effect as the frustration users experience when forced to switch between multiple applications to complete a task. Zero copy integration and cross-platform agent collaboration eliminate this friction, enabling users to stay in their workflow while agents handle data access and complex operations.

What business outcomes does this integration enable?

This video shows that unified Agentforce and Databricks integration drives faster decisions and smarter outcomes. Users get instant access to data, agents autonomously handle complex processes, and organizations benefit from consistent governance and compliance across all interactions.

Full transcript

Okay. Hello. Hello. Can you all hear me? Cool. Awesome. Uh. I'm a technical architect with Salesforce and today we're going to be talking about um our Salesforce experience inside uh rather inside of Salesforce. So how you can hydrate and make that experience better by connecting data bricks and uh by zero copy and then also connecting to data bricks agents. If you haven't already seen this, our booth is right behind us and you can see this live if you wanted to.
So first for those of you who are customers, thank you very much. Uh thank you for being Salesforce customers and thank you for coming to the conference. This is pretty awesome. This is my first data data and AI conference and I'm loving it so far. So, thank you all. Um, and we may mention things or I may mention things that are forward-looking. So, only make Salesforce purchasing decisions based on this beautiful slide that our legal team provides to us and is totally necessary.
Okay. So, when we're talking about data, a lot of you have already asked this question at our booth actually, which was pretty cool. Um, but in this age of data, you have to have access to all of it everywhere. So, we have a a product called data cloud, and you can see the little connections there in the screen, and it's showing you where in your different sources, um, those can be coming from and where they're going to.
Um, it's when you have like orders or you have product telemetry or something that exists in a data warehouse, uh, or lakehouse or somewhere in your data bricks environment, you want to be able to share that. So we give you an easy way to do that through data cloud and then serving it up where your customers need it. So giving it that example on the top there when you're seeing the top layer. Um so when we're talking about digital labor so digital labor being the agents.
So many of you may have already created agents and are using agents at your company today. But where those exist and how those exist uh can be a little bit different. Some may exist on an AWS, some may exist in data bricks, some may exist in Salesforce. So connecting those through to uh and having a singular point is very important. So they can make intelligent actions based off of that data, based off of the directions that you give when we're talking about AI.
So we're talking about this third wave of AI agents. So these agents need to be smart. They need to be available all the time. They need to be proactive talking to our customers. and they need to be scalable and accurate. I'm sure many of you may have taken a Whimo and the Whimo is powered by is uh an advanced AI agent powering and being able to drive you around and you're able to talk to that agent at the same time, but that agent has a series of instructions.
That agent has a role that it needs to play. It has data and it has different access as to what it can and can't do. Obviously, um it will get stopped by a stop sign or things like that. Those are hard topics uh and instructions that it can take. So moving on from this third wave here, let's talk about the Salesforce strategy, which is our applications, data plus agents. Underpinning all of this is the Salesforce platform.
So those of you who are Salesforce customers, you probably built flows, you probably may have integrations, you may have different apps, metadata. All that rich metadata can then permeate up into data cloud. So you can share all of that rich information uh throughout data cloud and that data that data from data cloud can thereby be activated. If you see in the graphic there's these little arrows from data cloud. So when we're talking about zerocopying data in from data bricks and you're bringing over say a specific table on product telemetry or something like that can then get moved into the partial into the experience layer what we call our customer 360.
So you can see the different areas here such as Slack, your sales service, any of these areas you can start to bring in that data cloud uh data in. So again that product data, something to help with marketing, anything like that. And ultimately we have that servicing to agents. They can be different types. So we have examples here such as a CX agent, an analytics agent. How how are my sales teams doing today? How are my service teams doing today? or something along uh legal and HR, how to answer some of those questions which are a bit harder.
So all of this powers what we're doing with agent force. Now I mentioned zero copy. So this is going to show us zero copy in a few different ways. Um as this has come up quite a bit at our booth. So um it's really great to be showing this to you all. So when we're talking about zero copy, we're talking about those lakeouses and data lakes that we're so familiar with. Obviously, we're at a big conference about data bricks, right?
We see the logo here, but we see the other the um uh other lighouses there as well. So, you have that zero copy federation. You're just doing a quick connection. You can sign in with your client ID or you can do a username and password and that's going to share the tables that it has access to. From there, you're going to start to share that information in Salesforce creating in DLOS's and DMO uh DMO inside of um uh data cloud.
And then ultimately back to those agents that we were talking about before or those experiences, that's where you're sharing that information. So if you needed to share it with an AI agent, if it's specifically in uh an application, if you're just showing orders, product telemetry, something like that, you can create a quick related list. Very quick and easy. Um, and then for those on ad platforms or BI platforms, analytics platforms, you can share that out to Tableau, you can share that out as you need it.
And then it can also get served out to all your different ad platforms. So things like CDP, marketing, any of that can be shared out. So any of this data that you already have, this rich data inside of any of your lighouses, it's very, very easy to quick and spin this up. Uh, and again, if you need to see any of this live directly behind me, you can see it. Um, and you can see folks just do a light setup there for you.
So again with our zero copy let's talk about the benefits. What does it do? Um traditionally if you're bringing in uh using Salesforce connect or you're using some sort of a legacy format API calls you may be charged on API calls uh back and forth bringing in that data. Um it could be errorprone it could be a lot slow you you might have to wait for a once a day refresh or something like that. With zero copy it's very costefficient to be able to bring this information over um in the way that you need it.
So you can accelerate that data. You can make it faster in real time as you need it. Uh it's accurate. It's also compliant with all the rules and logic that you've already set up with your tables and your warehouses. And all of this passes through our secure Salesforce trust boundary. There's also data governance behind this. And you can also create different spaces and how you want to partition this data and share it out to different folks throughout the organization such as I'm in sales and someone's in service. they may need to see some of different types of data and then based off of that data you can share that out.
So let's talk about how that works uh in more of the technical format. So on the top we see our query federation of creating those data lake objects that I talked about before and once you connect to your data your data bricks environment asking that connection you can then click on those tables you can look at that schema and then you can refer it over and that's ultimately you're going to be creating your canonical model and you build that uh canonical model and then you can utilize it wherever you want to.
So going back to that initial slide that I showed where it's going to be shared, you have that data model and that can go out. But now you can also do the file federation aspect. So you can make it faster and for those of you uh on a serverless environment, you can share that information much much quicker. Um and this is all built on the iceberg format. And if we want to talk a little bit more about this, we can talk at the end, but I want to make sure that we get through all of our content.
Um, and at the end of the sharing, so we're sharing that information back to data bricks if you need to go get it. Okay, I talked a lot about uh the data side. So we're going to talk a little bit about the agents. I referenced some of this before, but in the most basic format, we're going to be looking at essentially this type of slide in the next three slides. So just to give you an idea, the attributes of what an agent can do.
So I'm sorry about whatever double drunken words are going on in the screen. I can't I don't know how that happened. That's I'll just blame it on PowerPoint. But the agent needs to have a role, right? So I can be agent force and it has to have specific instructions as to what it's going to do. Agent Force Alex is speaking at a conference. He is talking about data data and AI but we have to power him with data. So we have to give those access to those links.
So the way that we were talking about that zerocopy information and bringing in a table, we can bring in a table of information. We can also call any of that information via API or if we wanted to grab that from another agent, we can do that as well. So there's different ways that you can go about that. And then there's the guardrails and actions which we call topics and actions. So you have to make a topic for that agent.
So agent could be the topic could be FAQ. And then you have to give them a set of instructions to follow. So every time someone comes to your store, respond with what's your email address so I can look at your account. You have to give them specific instructions which are all natural language. So you can just write in the natural language. It takes what you're going to do. It infers that and it's going to make that query.
And then that actions piece that is the most critical part because based off of those actions it's going to take and what's the job to be done? It's going to do that. So if it needs to query a record, if it needs to go out to a data bricks table, if it needs to write or update a record, any of these actions you can make real by any of the things that you've already built such as flows or any of the API calls, Apex classes are powering these actions.
And it's very very simple to write these in and power that agent. And lastly, is the channel where it exists. So does this need to be a WhatsApp bot? Does it need to be existing inside of Salesforce? Is it a Slack app? Is it an API? And is it existing on another site? So that's completely up to you is how uh where how and where these need to exist. Uh in the case of our agent, so if you go to help. salforce. com and you wanted to test this live, you could you could ask that agent.
It exists on a web store or rather it exists on our help. salforce. com and you can ask it any sort of questions and it's powered by the knowledge base uh inside of Salesforce. So, anything you want to know about any of those help docs, dev docs, you can go right there. You can test it out. And it tells you how many folks are doing it and how many times it's already been done. So, let's look at that again one more time.
So, as the agent takes action with its skills, this is what I was mentioning before. Now that we've set up our agent or theoretically, we have to see what the utterance is. Where is the status of my order? Very basic, right? And then it's doing metadata matching. So it's looking the LLM is understanding what those instructions are. And then it's using an LM to be able to parse through the instructions and look, I'm going to escalate refunds above 500.
Always confirm a customer's email before making those changes. And never change a billing method without customer confirmation. So those are those topics, instructions, those are those guardrails that we talked about before. And then ultimately it's getting to that action or uh orchestration and it's saying hey I need a flow to look up that order and I'm also going to provide the shipping status. And this can be done via Apex.
This can be done via prompt. This can be done uh you can use Apex to call API. There's a number of different ways that you can do this. HA MCP a lot of different things. A lot of different actions can be taken here. Ultimately, what the user sees is just this response at the end that your order has been shipped and will be arriving in 7 days. And then obviously you can include rich information, there can be links, there can be citations to whatever they're going to do.
And this is all underpinned by our trust layer. So, if there is sensitive information in here, say there is an email address that shouldn't be shown, say there's an account number, uh, credit card, anything like that is redacted and that is not also not shared with the LLMs that you're using because any of this information here can also, uh, the prompts that are calling it can you utilize a different LLM. So, if you wanted to use uh, Amazon or you wanted to use OpenAI or Anthropic, you have the choice to power your agent in different ways.
So when we're talking about this like what are we solving for? What are what are those things that an agent can help do and how can it do it with a different area of data. So if you have different metrics that you're after and you're trying to call into a semantic layer, you can grab different metric thresholds. You can grab something from different devices. You can grab them from different tables. Lots of people will have disparate knowledge sources.
So we have something called enterprise knowledge which goes across all different sources such as Confluence, Jira, uh Google Docs, Microsoft, any of those and it can call those different sources and pull it together and you can have something called an agent force data library. That library then powers that agent and gives it more. But when you need an agent to be able to look at how many people are coming to the site because there's an issue, a lot of those escalations are coming to a person and that human doesn't have enough time.
The agent is really good at those repeatable actions, being able to escalate that and pull it back and make sure that it answers all those questions on the front end, which is why I mentioned to go to help. salforce. com because if you're looking for a specific article or if you're looking for a different way, you can test it out there and if it's really an issue, then it will go to a human. You can say, "Hey, escalate to a human.
This is too much." Um, so this next slide, I'm just going to quickly highlight this for just a moment. But this is what's being shown over on the other side. So this is more of an A2A call, but you're going to be for those of you who are interested and you go to the other side, um you would see a customer service rep having an issue and they need to grab information. So what I'd say more folks are interested in and how it's actually working in the high level architecture.
But in short, customer service rep has a call, has an issue, has to get information out, and they're calling into that data bricks agent. So on the left hand side, we're talking about agent force. So agent force had that query that we showed before. It's going through atlas our reasoning system and it's using utilizing a flow but it's also grabbing information from um from data cloud uh doing vector search and it's grabbing information from service cloud to make that action when it comes across in that API call it's going to the agent from the servicing endpoint um and then it's going through mosaic and it's ultimately calling that table inside of data bricks.
So that portion I tried to make it really quick and simple, but if you want to see it live, you can go ahead and see it live. You can play with it today. Um, and it's very cool. This one will say demo. We're not going to do the demo. I don't have enough time, unfortunately. Um, so we're going to skip to that portion. Um, so this is kind of our productivity cycle and what we've learned, right? So we can identify everything.
Simple, small, repeatable tasks can be done with agent instructions uh going through a secure AI platform and then ultimately the channel that exists on this is leading to less tech debt. Lots of people are adopting and utilizing this to take like I'll I'll take the most simple form of can you reset my password. that's something that you can identify and it takes away a huge amount of just very simple cases that can be done by an agent.
Um, and that agent has to be trusted, right? Because this is enterprise data. This is not just something on the open internet that you're calling and sharing information with. So, you're providing it those secure guard rails. You're allowing it to have access to that specific data, provide those citations from your business location, and also see the reasoning. So, if you wanted to see the reasoning and the logic of this agent, again, I would implore you to go to our booth because you can look at that inside of our agent builder and you can see exactly what it's thinking, how it's working, and what it's doing.
Uh, this will give you understanding as to what it builds, what it does. And as I'm talking about that on the right hand side, the production scale, so when we're talking about um our agent DevOps testing center, the analytics and the observability and those insights. So when you're building this agent, you want to be able to understand again just a very at a simple level when you're in the agent builder what it's doing.
What are those reasoning? But you also have to test. You can't just if you wanted to test at scale but you're expecting 10,000 customers to come to your site a day. You want to make sure that you're doing it uh testing at scale through our testing center. And you can observe the analytics how that's working and how it's responding to different prompts and ultimately what what is that and then change it. you need to be able to change and monitor and update that agent.
Um, so yeah, in a nutshell that is agents working together with data that you may have shared in your lakehouse or data bricks. So thank you very much. I appreciate your attention and hope this was useful. I hope you have a great conference.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.