Tristan Baker, Salesforce | Neo4j GraphTalk 2026

Welcome back to the cube. I'm John Fur, your host in San Francisco for Neoforj's graph talk. Of course, there'll be a graph talk in New York City in September. I'll be there as well. This is where the practitioners and the leading technologists get together to talk about the future, the role of data and AI, specifically around graph databases, the role of knowledge graphs, ontologies. As the new data structures of AI emerge, a significant change is happening. We're doing our part to bring the experts to the table to discuss it. Tristan Baker here, senior director, head of data architecture at Salesforce. Welcome to the cube. Maybe soon to be called agentforce or AG&T, new ticker. I told Mark Beni off, welcome to the cube. >> Oh, thank you very much. I'm happy to be here. >> I told Mark Beni off said Dreamforce last year they should change the ticker from CRM to AGnt. >> Yeah, it might happen. >> Might happen. I [laughter] said, I hear it's in the works. Um Salesforce obviously pro agent that's really the narrative for Salesforce but you know in the technical community and the graph community >> everyone kind of knows the benefits and now it went cloud native cloud native is going AI native starting to see kind of a kind of a melting pot of largecale system architecture for business building out these AI visions and you have the top down management mandate put AI in the business and then you have the bottom up adoption on users and you have the practitioners building the systems to stand up knowledge graphs and put work AI workloads in production. That's kind of the big picture. But there's a lot going on. There's a lot of data. You got Oracle databases over there. You got a data bricks over there. I mean it's a tons of data. We talked to uh regulated industry in bio and pharma. They have tons of nuance data. I mean it's everywhere. >> Yeah. >> And but graphs are a nice solution. What's your take on the graph knowledge graph wave because it seems to offer up and and illuminate >> yes new relationships >> and makes the recursion of computer science work really well. >> So you I I I've been thinking about this problem for a while and you know going back maybe 10 years or so before AI as we know it today was as popular as it is now. um the the tools that you had at your disposal that are similar to what we're dealing with now looked like uh RDF owl and inference engines built on top of this extremely structured ontological language. So the push and it in my opinion it was still mostly academic at the time I was looking at it was take all your data and you can get the inference out of it if you represent your data correctly using this kind of precise cool uh representation language called RDF and owl and if you bring in these specialized interpretation engines you can infer interesting new facts from uh the data that an enterprise might have at its disposal. Um, and then kind of AI comes along and does it need highly structured information? No. You can just like feed it a document and it'll kind of read English and tell you out what it thinks. Do you need a hyper specialized prologue like theorem proving engine? No. You can kind of also give that to LLM and it'll reason through its uh neural network and come out with something else. Right? So um the the new question I think is the uh what is the right way to represent your data um so that AI can make the most sense of it. Uh in the old world you needed highly structured data structures and super specialized databases and machines. Now those things aren't needed as much anymore. But nobody is getting away from the issue of accurately describing your context, accurately representing enterprises information in a way that something can make uh sense of it. And I think the question on everybody's mind is of the 700 different ways people are currently experimenting with ways to do this. You've got graph rag that came out a little while ago, LLM wiki is a really interesting approach for representing knowledge. It has its limitations. Uh vector search AI or um semantic search over vector databases. These are all kind of like kind of u beating around the same question. Um, but I do think that, you know, in some of my own usage of it and some of the other folks that I think you spoke with today, I used to work with them in prior lives, so I know the problems they're working on, but everybody seems to be converging around, hey, maybe a graph is the right way to represent the facts and terminology and nodes and relationships that maybe you used to have to do in maybe what you might now call archaic old languages. Um, but now maybe there's a better way of doing it. And maybe we can demonstrate that if you follow these kinds of approaches, you're keeping your token usage down and all kinds of other things. >> You know, it's interesting. I love I love that explanation because it frames kind of where how we where we came from, where we got here. >> Yeah. >> It's funny. RDF um had a little triggered flashback [laughter] to the '9s, that one. But, you know, that was a metadata reasoning kind of, you know, issue. Yeah. >> Um and metadata was really more about describing data of data. So, okay, that's database stuff, but that's not reasoning. No. the way AI does it. I was just yesterday live streaming at the AMD advancing AI event and you know one of the things that came out of that event was is that inference and small compute pods could actually do a lot of good things and context was a discussion in their world. So you know when you look at >> metadata around databases that was just to make the database retrieval better. Yeah. >> It wasn't getting reasoning and the best answer. Yeah, I >> mean search gets stuff fast, gets the first thing first and you get it. That's not AI is not really a search paradigm. You have to have a search discovery mindset. >> Yes. >> So how do you think about the architecture now when you look at the neural network if you will being graph based because you know metadata in the graph is relationships. >> Yep. >> That has power. So there's a lot of benefits in computer science there now. So what what about AI do you like about the graph? Because the community is raving about converging. Wow, I got great search. Well, I'm done with that. What do I do next? And so there seems to be a progression of advancement. >> Yes. >> Where oh my god, when does this stop getting better? >> Yeah. Yeah. >> So how do you lay that out without getting over your skis, you know, on the architecture side? How do you zoom out and say, okay, I'm going to lay out super AI infrastructure. >> And I said yesterday that AI is now infrastructure. So you have AI infrastructure. Nvidia, AMD, Intel, Cerebrris, they're all pumping and clouds are getting fast. You got Neocloud, Salesforce, the original cloud. So the the infrastructure is getting better but AI itself is infrastructure if you think about it. So it's not metadata for the database. >> No >> the graph is an infrastructure component. >> It is it's becoming like that critical piece that helps tie the end agentic experience that you want to deliver. It's almost like the glue that stitches you stitches what the customer or the person is asking to the data that in the context that is needed in order to answer that question. Maybe to better answer your question, I can do it through the example of a use case and it's one that's top of my mind now and it's probably similar to a lot a lot of what people are thinking about which is um you know my company and all my peers working at all their companies all the CEOs of all these companies just say like where's my chatbot where I can just ask a question and I get my c [snorts] my QBR like report kind of spat back at me and then I can drill 17 layers deep on that and ask uh why do I see customer drop off over here? Where's my churn happening over there? Uh what is my financial outlook going to look like based on the uh issues that I'm seeing here? Where should I make the next product investments? Um those are the kinds of questions that um are top of mind for any leader. Um and >> by the way, they want those answers in like seconds. >> Yeah. In seconds. Mills. >> And they need to be right. [laughter] Right. Uh so it's one thing to just spit out garbage. It's another thing to spit out something that I actually believe. >> Yeah. They want a graph with that too, by the way. >> Yeah. Yeah. So not don't just give me the number. Tell me how you figured it out and give me all the things it's connected to. So this is I don't know I think this is kind of like the penultimate epitome of where >> at least my colleagues and a lot of us are attempting to go but um the the mechanism >> that's the user agency piece that's been verified. People want that. Oh absolutely. >> There's high demand for that use case. >> Yeah. And at least unless somebody's withholding information from me. I don't know anybody that's actually like totally solved this one. >> Yeah. No one's got this. But now they get it to the data. So we're backing into that. Okay. What does it take to do that? Okay. You got access to the databases. Where's the data? Because you want the best reports. You want to have access to everything >> and you got to get it retrieved, organized, reasoned in less than a second. >> Yeah, exactly. >> Or maybe mill. >> Conversational speed, which is seconds, you know, not much longer than that. Um, so, so if you unpack that problem a little bit, I like to think about it in a couple of different layers. At the maybe we'll go bottom to top. At the bottom bottom layer, it looks just like the database infrastructure that you're familiar with, right? I think that you're going to continue to need lakeouses or analytical data stores. I think you might continue to need um I think maybe what's colloially referred to as uh customer profile stores. And I I think you're going to continue to need your operational databases. Um so relational databases, NoSQL databases, um >> time series databas time series databases. These things exist because at the fundamental le level they implement different data structures that are optimized for different kinds of retrieval. um you you have an indexed based data store and those tend to be the things that power your um uh front-end applications because you got one customer right now and they need to look up just their record. They don't care about all your other millions of customers. And then you have uh columner based file formats and lakehouses and they're organized that way and they don't have indexes because nobody wants one customer. People want the aggregate sum of something across all 10 billion customers. So, as long as you're going to have different kinds of um questions, you're going to have different kinds of retrieval data structures and the engines that power them. So, that's always going to be varied. I don't think we're heading to a world yet where one database is going to do everything. >> I know because it's I mean, there was an old expression I always used to use. So, I'm when I I do cub interviews or meet people, I'll have to tap that section of my brain. Yes. >> In a way, you can look at a time series database and say that that's a section of the brain. So if you're building >> things in a different way than other sections >> and if I'm a company and I was had the you know the tech team I'd be like okay I want a brain for the company. >> Yeah. >> And I want this brain to be powering everything. Yeah. So if you could build your company brain what would that look like? Yeah. Exactly. It's a thought exercise. Okay. I want my time series over here. I want my my art art some creative side here. I want my math. So you can actually get and and then okay how do I tap it? Then the how do you tap it piece is the next layer up which is okay you've got essentially the same fact about a customer maybe represented in six different databases 60 different databases so now immediately you have this problem of well where's the truth about my customer because now I'm copying it 60 different times so the next fact then becomes uh or the next layer you might need on top of that then starts to look a little bit like this metadata layer that says where are six different copies of this data maintained what's the source of truth what's the uh system of record, what's the eventual consistent copy and when should I use what for what kind of question? Those become uh interesting metadata descriptions that you need. Then uh uh you also probably need what does it mean? Like now you're starting to get to the point where you're going to map the way that a a natural language thinking person mentions a metric like uh like a KPI like uh average order value or average contract value or uh a annual recurring revenue for product XYZ. That's how people talk about it, but that's not the name of the column in your database. So now you need something that maps your business terminology to where this stuff exists. You also need something that maybe keeps track of what kind of math is the person implying that I need to run right now. And if it's this kind of operation, now I'm getting specialized into this data store. If it's another operation, I'll go to this one. So now you're kind of packing in context and knowledge about what you would normally use a data engineering team for or a database design or an engineering team for. You have to kind of uh bake in that engineering uh context as well so that you can meet that requirement that you mentioned which was yes, I want an answer, but no, I don't want it a day from I want it in like 5 seconds >> and I need a graph to be the I guess spine or backbone or you know the connection where you can hang the relationships off. So you have customer once but it can be mentioned zillion places >> and that'll have that will that have the relation graph of graphs people think about >> at least the way now I'm not pretending to say that I know exactly the right answer because if I did then I would you wouldn't be here going public. >> Yeah I'd be I would have built it already. So now I'm talking I'm speculative maybe more assertive. So uh then then you go okay well what is what is the role of that graph? So in my architecture now the role of this graph database is to manage the relationships and the context that are at more of that metadata level that say where is the data how should I query it and how can I connect things together because potentially my user is not asking for a metric that's already calculated they might be asking for a relationship or correlative maybe they're exploring a correlative causal hypothesis between two things in which case I might need to at the same time fire a query over here. I need to fire one over here and bring the two results back together and then uh let the user determine if there's actually like gold in. >> And by the way, the user is treating this as a generative query basically because you never know. You can't anticipate >> the what the relationship ask is. >> Yeah. Yeah. >> You just I mean that's has to be done >> in the graph itself. >> What makes that happen? Is it is it the ability to recurse through that? Is it the science behind it? What makes it so good? >> Uh yeah. So the you mean the the power of the natural sorry the power of a of a graph database is in its ability I mean you have kind of a declarative language of saying I think these two things are related find for me the connections that prove that that is the case >> um and now I can take advantage of those connections. So the now we're getting into the like the mechanics of how a graph database works which honestly I'm not the right person to tell you exactly why it's better than a relational >> database but you lay out the big picture. I mean it's like laying out a system architecture for any system. Yes. I mean data architecture now isn't >> database specific. No, you got to look holistically and say okay we're building a data system. Exactly. That's what the mindset I'm seeing people who are succeeding in production with AI because they have to handle you know scale deterministic agents agents and robotics are really pointing to deterministic >> behavior because agents behave and they touch everything storage memory >> um and like microservices if they go off the rails >> you know >> I mean the only good is what do you give them >> yeah exactly >> like you train the agent well why it's new and kind of slapping live, bring it live, you know, so there's a lot of that going on. >> Yeah. So, like most things, you know, many systems are garbage in, garbage out, right? So, you can have the best technology in the world, but if you're not careful about the content you're exposing it to, then you're going to end up with nothing very useful. >> What are some of the talks here at Graph Talk at Neo Forj because New York's coming up September. Um, a lot of these conferences when they're small community events, all the leaders are here, they're the pioneers, the frontier leaders. >> What are some of the hot issues uh people are talking about? What's the >> what is the top? So the one that I can talk about the one that that I that I think is the least discussed I'm again I'm just kind of doing a pulse check when I go around and talk to people about this stuff. I don't hear discussed very much um the the access control and governance aspects of um how are you going to protect the data from misuse or inappropriate access uh uh in the architecture that I just described for the use case that I just described. So again we can go back to the my example use case and let's say that you >> well if you're mapping data yeah >> you need to know certain things. So if you if you have a CEO that's conversing with an agent, maybe you can make the argument that the CEO more or less has access to be able to see everything. But then, you know, the CEO has divided his sweet seat or her sweet suite up into well my there's my marketing person and my uh legal compliance officer, my security officer, etc., etc. And then they have organizations and they have and so on. And eventually you get to portions, you don't have to go very far before you get to portions of the organization. Well, maybe you don't need to and shouldn't see the same data sets that this portion of. >> You don't need to see payroll. Sorry. >> Yeah. So a lot of people don't need to see payroll at that level of detail. Um so we talk a lot about okay maybe there's a semantic representation of um all the data at the company. Um but again maybe it lives in very low-level more specific databases. Traditionally the access control is managed at those database levels. So I write a grant statement and that's what protects my um uh my rows and columns and things in my relational database. I set an S3 bucket policy and that's what uh prevents people from accessing buckets of un unstructured data. But if we're pushing all of the metadata up at a higher level to help tie that conversational analytic use case, then I guess in addition to needing a semantic description of the data, you start to also maybe need a semantic description of your access policy because your legal team is going to come to you and they're not going to say here's how data should be protected in Postgress. They're going to say people like this should not be able to access data like that. >> They're going to speak in like normal >> speak they're going to speak in normal language language. >> So at the same time your CEO and seuite and everybody is asking natural language questions of your data. >> Yeah. your legal and compliance team is reading a bunch of contracts and agreements that they've negotiated with their clients and they're saying, "Yeah, sure, but don't let him do this and he shouldn't be able to do this." And they're also describing it in you. You're bringing up you're bring up the top conversation that I was talking to the Neo4j CTO about, which is, you know, Neo4j with the graph. That's a great solution. They're doing great with it. The timing is perfect with AI, but system of intelligence isn't just about graphs. >> It's an ecosystem. You mentioned um different database. You got data bricks and snowflake dominated the data analytics business but now they're trying to go in data bricks has come in and say no no no we're the AI company with open tables take that snowflake um and then so they're they're $188 billion company they're10 billion less than IBM yeah >> but then they're private company don't really trade so it's really not a right market value but you know think about data bricks is like literally $10 billion less than IBM >> so it shows you that that the data lakes are key piece but pipelines come up a lot too. You got to have good data pipelines. Yeah. >> Um and you got to feed the agents. So this is not just about graphs. Graphs I like I think from my standpoint I look at it's very valuable for the neural network piece. But you have an ecosystem of data. >> Yeah. >> And you got to get you got to rein it in. >> How do you think about that? How are people thinking about because then you start going outside the the religion of grass which is you know I love to talk about grass. I think it's super important people will know more about it but there's a bigger picture. >> Oh for sure. Yeah. that the the the map the domain map if you will of everything having to do with managing data is much much more broad than uh just the role that a graph database or even a semantic layer might play in it. Um uh so I I think the problems themselves are no different than they have been over the last 10 or 20 >> Yeah. >> or forever as far as I'm concerned, right? Um you've got your kind of governance compliance concerns about making sure data is appropriately used um and that you're not leaking it or or or losing track of it. Uh you've got your performance concerns like am I spending more money than I need to for the amount of value that I'm getting out of this. I've been in many different projects called clean all this mess up because we're paying too much. >> Exactly. And those projects go really slow if you don't have great tracking of why this is here and what value it's providing. Uh plus not to mention it's its own compliance risk to have things lying around. So there's those kinds of concerns. Those aren't changing. >> Harmonization layers are emerging. >> Master data management still I think continues to be a critical piece of this because you you will continue to have multiple sources of opinion about who this customer is. And this is definitely true as you're talking about like your prospect to customer pipeline. your prospects, you're always like kind of sure who they are, but you're not 100% sure. And is this person the same person as the conference person I met over here? Are they related to the same leader at this organization? So, you're always harmonizing and mastering records. And so, some of these things we talked about how ontology is maybe what's old is new again. I think master data management as well is gaining popularity or I don't know mind >> governance is the number one issue on aentic >> master MDM and and governance go to hand to hand that like you get that right good things happen that seems to be the pattern. >> Yeah. So, so really it's just it's taking every one of these problems that and I don't know anybody who spent any time in this space has probably spent a lot of time thinking about and now you're thinking about great >> how do I identify it right like how do I make sure >> it sounds so easy >> it does doesn't it >> should throw [laughter] a well I mean Salesforce I you know we've been meeting a lot of your your SVPs in the tech side over the past couple years Dreamforce last year too you have so much data system of record dominates the system of record business engagement is evolving but this holy grail is the system of intelligence. Yes. >> Because now they have these AI infrastructure systems. AI is infrastructure. >> Yeah. >> It's a perfect scenario to actually deploy >> of course >> a systems of intelligence. Everyone's trying to figure it out. >> Yeah. Yeah. So the the investments that Salesforce has made in over the years continue to be relevant. Obviously you have the core CRM product, the um uh the marketing engine, the the marketing products, data cloud, now data 360 or maybe it's >> agent uh core. >> Yes. Uh so there's there's agent force >> agent force >> y >> agent core that's Amazon so confused >> right but the this pattern of that I had described earlier of like you have your your core product in this case a CRM product then you have your kind of activation layer which is built on top of your customer profile store that's like uh data cloud um and uh those those patterns aren't really changing right those patterns I think are present not only at companies that I've worked hard for before but it's also the core product that Salesforce sells And now they're talking about okay congratulations if you've been using Salesforce and using it properly and stuck with us over this period of time. This thing called agent force now gives you the opportunity to kind of identify not only your workflows but also get agentic based insights into your data. So that's >> I think and I think the you know the playing with data as a value proposition is not what end users sign up to do not usually. So if you can eliminate the toil of data entry and eliminate the hassle of getting at the data and doing the reports, that's the holy grail. And you know, I love the Slack integration. >> Yeah. >> Into Salesforce is another talk about neural network uh opportunity to build the company brain. You can chat. That's the the layer of consumption or the layer of engagement is is what Slack is described as. Um and we uh we practice what we preach as well. I've been, like I said, I've only been there for two months, but >> well, as senior director, head of data architecture, you are at the right place to get your hands on the keys to the kingdom. I mean, there's so much to to master. It's such a great domain of opportunity. Yep. >> Thanks for coming on the cube. Appreciate. >> Thank you very much, John. >> What do you think about the event here so far? >> Oh, it's been great. I I was able to attend half of it. Unfortunately, not all of it, but um uh I was in a panel just before this um answering some of the same questions that you were asking me. So uh what you're asking me is certainly relevant because there's 200 people out there that were asking versions of the same question. >> The whole world wants to know. I mean I think this g be very popular. I love the graphics. We have a real critical piece in the data stack. Yeah. >> With AI as well as the system intelligence system. Yep. >> And that's the operating system of the data operating system emerging. There's no doubt about it. And and I hope to catch up with you at Dreamforce couple weeks. So >> good to see you right down the street. Of course we'll be there. Thanks for coming on. >> Thank you very much. >> All right. Cool. I'm John F with the cube. We are here at the Neoforj. Graph Talk. It's soon to be a podcast because it sounds like a great podcasting name, but it's also an event here in San Francisco. The leaders get together, of course, in New York City in September. I'll be there as well with the Cube. Thanks for watching.