From Old AI to Agentic AI  ## Metadata - Author: [[The New Stack Podcast]] - Full Title: From Old AI to Agentic AI - Category: #podcasts - URL: https://share.snipd.com/episode/1e0d1ba9-eef3-457c-9802-73aad2c0bbff ## Document Notes ## Highlights - **Cricket Sparked A Data Mindset** - Rahul describes growing up in India and following cricket through newspapers, radio, and conversation which trained him to synthesize scattered data. - He used statistics and contextual storytelling to fill gaps and create meaningful narratives from varied sources. Transcript: Rahul Auradkar It's interesting. If I think about it, if I go back to, thanks for the background, by the way, yes, I did grow up in India. And growing up in India, I was an ardent, even now I am, an ardent sports fan. I grew up playing cricket, a game that's very popular in India and Britain and Australia and other places. I also followed the game quite a bit as well. And growing up, thinking about data, right? Growing up, I followed the game through. I was a voracious reader of newspapers and magazines and also a listener of radio and television. And that's how I followed the game. And I did a lot of conversations with people. But then the foundation to all of that was I used to have all of these conversations with people about the game. And the players, their averages, their scores, it's a very, very highly statistically intensive game. And always I found myself answering questions which were grounded in some data, but not all data. Grounded in some facts, but not all facts. So it used to be very subjective. So over time, I started to look at data that comes from various different sources. The key here is not about whether the data is right. When you talk about good or bad data, bad data really comes in many different forms, right? Is it complete? Is it of high quality, high fidelity? Is it been summarized in a manner in which it's meaningful for the topic at hand right so if i sit here today and um 35 years 40 years after that and i reflect back on those days i look at it As in saying i was doing that all of that and we were doing all of that mentally as in filling in the gaps unlocking data that is sitting in newspapers in radio on what we heard on radio what You see and saw on matches and games then kind of bringing context to it what does context bringing context mean you really want to have some normalized schemas we talk about in the in Today's parlance we talk about normalized schemas but the idea is to make sense out of it through some normalized ways in which you can connect it to conversations that you and i can have So um, and if I look back today, if I look at it today from if you look back, the amount of effort I spent in bringing together all of those different facets of data, whether it is newspapers That's in today's form, we call it unstructured data, but in other forms in real stats, etc. I was a voracious consumer of those numbers, but I told stories using that. ([Time 0:02:29](https://share.snipd.com/snip/db4f9c8e-2713-4e64-bca5-bb3773f4f933)) - **Automation Amplifies Bad Data Risks** - Rahul frames AI as intelligent automation at scale that amplifies the harm from low-fidelity data. - He warns that automation magnifies errors and trust violations when data lacks accuracy and relevance. Transcript: Other Speaker I think you answered that question a bit when you talked about context. Me, bad data is very much tied with the experience. And if the data is inaccurate, that ruins your experience. You're less likely, I think, to rely on that data. I think that's one reason why I love to read the sports page as a kid, because I could read the box scores. I was a baseball fan. And I could look at the batting averages of the top hitters and on and on and on. But if that data was incorrect just once, just once, I would be done. And today, it seems like with AI, we face a whole nother realm of challenges with bad data, changing the experience. And it's a different kind of experience. Now it's like, wait a second. We're reading often about all these weird hallucinations that are happening. Databases being deleted. And I just wonder about that bad data story and how you see it today. Rahul Auradkar It's interesting you bring up the idea about bad data and in the context of AI. What you've all said and done at the end of the day, what AI is doing is it's driving intelligent automation and doing that at scale. Now, if you just replace an AI, whatever the AI piece of software is or an agent is, if you replace that with a human, right, you and I go through a lot of pieces of data. For example, a customer service rep in a CRM instance or an account executive in a sales instance is looking at some aspects of data that is informing their business decisions they're Making on behalf of their customers. Or you and I are looking at our credit card statement and we're looking up the bank's website and we're making some decisions. Now, when we, you and I do that, there's a little bit of latitude because there are humans involved. You make some decisions. You probably take a pause. But when you start doing what AI does, which is intelligent automation and intelligent automation at scale, then what gets lost along the way is that if there is high, the fidelity of The data is not high, it's not accurate, then it gets lost in the shuffle. Then the downstream implications of the automation are profound if you get it wrong not only profound in terms of uh in terms of uh accuracy getting it wrong also in some cases you might Get it wrong with trust being violated as well right it's it's about accuracy it's about relevance it's about trust and those things really become all the more important when you're Trying to automate at scale that you don't control as a human. ([Time 0:05:21](https://share.snipd.com/snip/8a2acbb3-b272-418c-80d7-7ddb977e6d48)) - **Data Evolution Created Trapped Silos** - Rahul traces data evolution from relational databases to warehouses, lakes, and lakehouses and shows how this created protocol and SQL fragmentation. - He identifies those divergent formats and silos as the core source of trapped data debt enterprises must solve. Transcript: Rahul Auradkar So I wouldn't say engineering mistakes. Let me just take a step back in terms of, if you look at it from the lenses of the evolution of where we are today, and it continues to evolve. As a matter of fact, it evolves much faster now than it ever did. Just from a data standpoint, if you harken back maybe 20, 30 years back, you have structured applications that use structured data, form some data type structured applications where You really access some data, you did something with it, you stored it back, and that's how automation was being driven. And what was really powering it in the back end was a data repository called as a database. And that was structured data. It was essentially a relational store. And then over time, people start saying, wait, that relational store has so much data associated with the transactional set of things that I was doing using my applications. I need to take that out and analyze it. And that's how things like warehouses were born, right? Warehouses were analyzing structured data on which you drove analytics. Analytics over time started get fed started being fed back into business processes processes to say hey my descriptive analytics or my analytics that i discovered if you may while Looking at the data here through analytics that i was running in warehouses helped me improve performance of business in a certain region or a certain aspect of my business and so i went In and changed my business processes. So that cycle was analytics. Data was an exhaust. And that exhaust was being used to analyze that analysis was being used to help improve businesses. Over time, then as you evolved, then they said, wait, I can take a snapshot of that data and also, and start training these machine learning models that allow you to do predictions allow You to do some sort of extrapolation if you may outside of the data set that exists in there that's how machine learning or what i refer to as the as the old ai or the boring ai or the predictive Ai was born right now what has happened over a period of time is that whole paradigm has now accelerated to the point where we're saying, wait, those applications in themselves don't Need to treat the data that they create as exhaust, but instead that exhaust itself becomes the way in which we define the applications. That's where all of this generative AI came in. As in, I know something about Rahul. I know a lot about what Rahul has done in my system. I know about his footprint on what websites he has gone. It's my first party data, as in I'm a business. And Rahul has come in through mobile applications, variety of, I know his footprint. Using a footprint, can I really tailor, this is also known as personalization, but can I tailor the business processes along the way? The reason I'm thinking that we should look at it from the lenses of that is along the way, what was born? What was born was databases, warehouses, data lakes, lake houses. All of these were born along the way. Lake houses were born because you wanted to get the best of warehouses and data lakes, which really took unstructured data into account. Now, all of those ended up having their own storage protocols, their own table constructs, their own file constructs, their own query protocols. Even SQL as a language has its own SQL variant, whether it is done by Oracle or by Microsoft. So that's how the debt got created, if you may. The silos got created, and the silos have a lot of trapped data in there. And how do you untrap it seamlessly so that it can be used for meaningful interactions for businesses in their business applications? That problem remains unsolved. ([Time 0:08:39](https://share.snipd.com/snip/c0928a8f-7ad3-408e-8827-4666fd203879)) - **Harmonize Data For Agentic Workflows** - Harmonize and unify data schemas and identities so agents can combine structured and unstructured sources in real time. - Deliver unified, real-time profiles and insights to agents to drive precise recommendations and faster human decisions. Transcript: Rahul Auradkar Yeah. So unified data, harmonized data, those are the phrases that are used interchangeably. Do mean different things, harmonization and unification. Let me explain a little bit about it. Previously, I talked about warehouses, lake houses, lakes, databases. There is data sitting in all of these silos. When we talk about what we do with data cloud and agent force at Salesforce, we're not talking about replacing the silos. The silos have a lot of gravity in them, data gravity, and that's what many of these vendors really are driving. What we're talking about is how do you bring that data to life in a fluid manner, data fluidity, if you may, that's harmonized and unified. Now why is it why is that necessary? As an example, if you go to help.salesforce now, that is agent first experience that we are driving. If you go ask a question in there as in does Data Cloud have an SAP connector, you will find that answer there much richer than what you would get in Google or ChatGPT. If you ask a question as in, can you show me the setup for the data cloud SAP connector, you'll get a fairly detailed answer there. Why is that possible? It's possible because we have hundreds of thousands of unstructured textual data that we have ingested, we have vectorized, and made it available when your agent is asking that question, Or you're asking that question that the agent is helping you with. Now, connect that with structured data. Now, let me use a structured data example. An example of that would be, you have a customer service representative, a human, working alongside an agent. The human has their, in this case, the customer service representative has their, his or her own agent. You're asking a question of that agent saying, tell me Rahul's credit score, because I'm asking a question of that, of another, my, another customer facing agent. I'm asking, saying, hey, I want to, I want to get a credit raise. So the customer service rep is in turn asking a question to his or her own agent which is internal to the bank employee should this person's a bank employee asking a question saying can You give me rahul's credit score can we and you get the credit score then you say can you give me the criteria by which i need to raise the credit limit now it's going into unstructured data And getting that criteria. Why is the unified data matters here? It's harmonized and unified. What's harmonization and unification here? You have data of all the transactions Rahul has done on this credit card over the last that many whatever period of time that and it's sitting in a warehouse somewhere. You're bringing that data in real time in a fluid manner and combining it with Rahul's profile. And then you're connecting it with some criteria that the agent is, the AI agent is given, giving to the human agent as in saying, this is a criteria if Rahul's credit score exceeds X, Which is maybe 700 or 750, whatever the number is. And then if this is what the transaction history looks like, here is the criteria by which you would raise this credit limit. Now, that harmonized unified data is coming from, perhaps it might be coming from S3 in AWS, the unstructured data. It might be coming from Google files, all of the information that is related to some criteria that is stored locally, maybe, or it's coming from Snowflake, all your transaction history. You're bringing all of that. You're harmonizing it, meaning you're creating a consistent schema of what does a credit card mean? What does Rahul as an individual mean? What does a contact mean? What does a credit card mean? You're harmonizing it to that schema. And then you're unifying it because Rahul might be using his cell number to come in through one mode. Through another modality, Rahul is coming in through his email address. There are two different modes now. You're unifying the same Rahul now. So it's harmonization is different from unification. Then you're deriving an insight. An insight could be, hey, the last 10 transactions that happened with Rahul is paid up in time. There's no problem with this history at all. Then I start looking at 100 transactions. He's missed 14 of them. There's a problem, right? So you're driving insights. Now, using those insights, that system is now giving some guidance to, should I or should I not raise credit limit? Then the human is making the decision at the last minute to give the credit card raise. So bringing it back to what you were saying, if, for example, that swivel chair exercise, this human would have had to go and log into another system, take a look at all the transactions, Log into another system, take a look at the criteria by which I raise credit limit, log to another system, take a look at my credit score. The human now becomes that much more productive because data is so much richer. It is fuller. It's complete. It's harmonized. It's unified. It's available to the agent. It's available to the human. The agent is then making recommendations. ([Time 0:13:20](https://share.snipd.com/snip/98bff1f6-375d-44fd-80c6-f2d1bf2dc269)) - **Automate Governance And Runtime Controls** - Tag data automatically for sensitivity (PII, PCI, etc.) and enforce policies with attribute-based access control at runtime. - Use automated governance so agents access only the data your policies allow and maintain trusted AI. Transcript: Rahul Auradkar So first off, you refer to identity management. I'd elevate it to access. I'd elevate it to trust. I'd elevate it to governance to security so when you start looking at agents by definition some of these powerful llms are really set up in in which they can get access to as much data as They possibly can get access to that's by design that's not that's a feature that's not about but from a standpoint of it being relevant to enterprises who really care about their brand Name, they care about security, they care about providing the right experience to customers, they care about data governance. Unified first party data that is reliable, but the more important thing is it's trusted, that in a meaningful manner in which the agent will get access to only the data that the policy That is defined by the business. And how do you go and force that data governance? Right? Recently, we just at Salesforce, weed a data governance offering inside of data cloud the way that works is that you automatically tag a data whether it's structured or unstructured Within chunks you can tag them whether it's pii or pci or it has a social security number or it's external or internal etc you tag them and you and data is happening at 10x, 100x, 1,000x More scale now than it ever did a couple of decades back, we're doing that in an automated fashion. Once you finish tagging, then we define policies. The policies also could be default policies, but they also be human-driven policies, as in saying anything that is PII-driven, nobody gets access except by exception. And then you enforce that during runtime with this thing known as attribute-based access control. You enforce those policies for an agent or for a human. So that now is available across structured and unstructured data. I like to say that we went slow to go fast because we wanted to make sure that we deliver trusted data for trusted agents. ([Time 0:19:00](https://share.snipd.com/snip/c3d0261e-e673-497b-8c2e-b2b86e3bc41f)) - **Design From Outcomes, Not Tech First** - Start from desired outcomes and work backward to define data fluidity, normalized protocols, and unified schemas. - Break silos and design architecture and teams around delivering trusted, real-time data for agents. Transcript: Rahul Auradkar Yeah, I would say come at it from a standpoint of what outcomes you desire. Let me explain what I mean by that. Today we talk about data cloud as an example, but data cloud and agent force can be an intelligent activation engine where it's bringing the full value of your enterprise data to life. Now working backwards from them, backwards from it, right? How can data play a central role in the agentic enterprise? And the way it plays a central role in an agentic enterprise is to provide unified, trusted, and real-time data as a foundation that fuels agents. If you work, then you take the next step as in saying, for that, you need fluidity of data. For that, you need fluidity that the agents have what they need in the moment for the business use case they're trying to deliver value for. It means you need to start looking at data as in moving around in a fluid manner, in a trusted manner. So start to look at normalized protocols, start to look at normalized schema, start to look at unified data. To do that, start breaking silos, right? Think along those lines, and then all the decisions you make at the architectural level, at the systems level, then will start to fall in place with that objective in mind. ([Time 0:24:04](https://share.snipd.com/snip/34c5af96-a84a-41b1-97f4-882a7c49585b)) - **Databases Stay, Abstractions Enable Sharing** - Databases remain the backend for transactional apps while lakehouse-style platforms enable shared analytics and agent use. - Abstraction layers and zero-copy approaches let systems share normalized data without constant movement. Transcript: Rahul Auradkar A databases standpoint, the way I look at it is they continue to serve the purposes in general as being the backends for transaction systems. You have an application that will continue to use databases in the back end. That's what we're seeing right now. Those databases have improved over a period of time. They know how to use structured unstructured data. They know how to really create graphs. They become more speedy. They've become more scaled out, if you may. They're cloud-based, etc. All of that evolution is working through. Now, there is also data, in some sense, think about it as being databases like the way they're fitted into applications. You're finding other data platforms like Salesforce Data Cloud, which are providing what I refer to as the Lakehouse foundations and on which we're building all this tooling, just Like what other vendors are building as well, to make that data relevant for not just the not just AI agents, but also relevant to feed it back into the transactional systems. So if you refer to Databricks as an example, we are working closely with Databricks. We're working closely with Snowflake. They are partners of ours. One thing that is evolving over a period of time here is an interesting one is that we're finding abstraction layers that allow for data to be shared across systems. So that's an evolution that we're seeing quite dramatically. We went out and we created a zero copy network, meaning Salesforce went out and created a zero copy network. Why did we do it? Because our customers asked us the ability for us to leave the data where it is, but use it for the use that they want to so they don't have to move it around in some cases it's sensitive data In other cases it's for economic reasons they don't want to move it so uh databricks for example is now adherent to the as an example in the file layer with iceberg they even bought a company Called tabular to drive iceberg-based uh standards we drove uh i mean iceberg is an open standard, but we went out and pushed and using that for zero copy. So you'll start to see a lot more normalization. So data can be shared seamlessly across all systems while underlying the plumbing also changes quite a bit so that you have this concept of connectivity, the concept of graph, if you May, where you can find all touch points really driven through profiles, driven through what, whether the profile could be a human, the profile could be a lead, it could be an account, It could be a vehicle, it could be anything that people are really trying to attach it to. ([Time 0:25:54](https://share.snipd.com/snip/7a8f5c6a-c978-4e00-98c9-62cd962729a9)) - **Salesforce’s Deeply Unified Platform** - Rahul outlines Salesforce's deeply unified platform that combines Data Cloud, Flow, Tableau, and Slack to power agentic CRM. - He says industry apps and semantic models run on this foundation to deliver consistent enterprise semantics. Transcript: Rahul Auradkar We have talked about this concept of a deeply unified platform on which we're building our business app, CRM, and what CRM has become now. And we continue to invest heavily on the deeply unified platform, whether it is the agentic foundation built on top of the deeply unified platform, data cloud inside of the deeply unified Platform. Deterministic flow based automation, we have this product or a service known as flow inside of Salesforce that drives trillions and trillions of automations within our business Applications. We now have Tableau, we just got Tableau Next, which is running on a semantic model that is built inside data cloud that's part of the platform that allows you to have consistent semantics In an enterprise we have slack now that uses the same data for collaboration as well and we have industry applications that run on top of it on top of the platform that allow allows us to Understand the domain better but uses the underlying deeply unified platform that's the direction in which we are headed we have talked we it at last Dreamforce. We're going to talk quite a bit more about this upcoming Dreamforce on the value and the power of the deeply unified platform for being the fuel that really drives the agentic enterprise. ([Time 0:28:51](https://share.snipd.com/snip/f9734179-7ac8-49f9-80ee-854067b20695))