AI Imperatives in Financial Services

Hello everyone. My name is Erin Agareos, and here with me today is Babar Khan. Thank you so much for joining.

Thank you.

I think we're gonna have a great discussion. We talk a lot with customers, and I think what is the kind of number one thing that you hear from customers in terms of autonomous agents?

Yeah, yeah, that's a good one, right? I mean, it has to be security, compliance, and trust. I mean, every single time.

Every single time. And I think one of the things our customers have been asking for is just more education, more education on this topic. What is Salesforce doing and how do we think about that? And so that, that's what we're gonna talk about today. And maybe I'll start off with kind of at the highest level. We get a lot of questions around like, is Salesforce a large language model? Are we a wrapper around these large language models? Maybe you could describe for us what Salesforce is actually doing. What are we thinking about in terms of the agent force layer and how does that work with our large language model?

Yeah, that's a great question, right? And it definitely is 'cause customers, sometimes they're confused as to what they're buying into, right? So Salesforce is a platform, right? We are not an LLM. What that means is that we are leveraging the LLM capabilities, but we are a platform, which means we come with a lot of security that's already built into the platform, right? And as you kind of build, go above a layer, right? You got Data360 and then you have the Agentic layer sitting on top of that. Sitting on top of the platform. And within all of this, for the GenAI and the GenTech use cases, we are using the Agentforce Trust Layer, which does all of this great work around dynamic grounding, making sure the sensitive data is masked, no hallucinations when we are making callouts to the OpenAI, right? Things which are outside of the Salesforce's trust boundary. Mm-hmm. None of that data is retained by those LLM models, right? So we have contractual contracts and things like those around zero data retention with those LLMs. And then as a response comes back up to the platform, right, we are then de-masking any of the sensitive information that we winded up masking. And then we are auditing all of that activity and keeping it within Data360 for the auditors and the regulators, right? Right. So, I know I talked about a lot of things, but the key thing is that we are not an LLM, we are a platform, which means everything that our customers are used to seeing with Salesforce, right, from day one, right? Genetec layer is leveraging and inheriting all of that security and the compliance capabilities within the platform.

And you did, you talked about a few things there that I think is worth just double-clicking on. Why don't we start with this idea, and I think we hear this a lot from our customers and our CISOs is, You know, I've got amazing data, I've got golden data, the bank's data, insurance company's data about my customers. I don't want that to train the large language model and, you know, give a maybe leg up to my competition. You, you talked about a little bit about that, but maybe let's just, what does that actually mean and how are we preventing that data from going to the LLMs?

Yeah, yeah. And we always tell our customers that your data is not our product.

Yeah.

Right? That's, that's Salesforce's stance. On top of it, when we are making callouts to the external LLMs, we have these zero data retention policies. What that simply means is that when we make a callout, that prompt is never kept on that LLM side. So it's like writing on a chalkboard. You write something, you do the equation, you figure out something, and then you erase it.

Right.

It's gone. So that's a very important point. On top of it, your sensitive information that you are used to bringing onto the platform, We give you the ability to do many things with it. For example, classify that data, right? Make it NPI, PCI, whatever might be the case. You can mask that information as well depending on the use cases. So none of that data is being leveraged for training the models, retaining it for any purpose whatsoever, right? So we are taking care of all of that as part of that Aginforce Trust Layer. And those zero data retention policies are then verified through the technical controls that we have to ensure that the external LLMs are indeed not keeping any of that data. And for our customers, we also provide additional details through our, obviously, the SOC 2 reports as well.

Okay, what about the agents and the access controls that they have or the access controls that we provide? I mean, do the agents get access to all the data?

They do not, right? And this is why I'm going to keep on going back to the platform, right? Because platform always had the access controls and the security model built in around who sees what, right? Who gets to do what within the platform. So when you bring in the Agentic experience into the mix within the platform, agents can only see what they're supposed to be or what they're able to see based on the type of agent that it is. So it can never see more than what it needs to see. On top of that, there are additional security controls as well where you can see things like, you know, who has access to what type scenarios, right, for your additional audit and the compliance and the verification purposes as well. But inheriting the platform security is the key, and agent is operating within the realms of what the platform provides from that access control standpoint, plus the role-based access control side of the house as well.

Okay, so is it fair to say if a human can't see it, their agent can't see it either?

That is true.

Okay. Now, another topic that comes up quite regularly is related to the ability to audit the decisions that these agents are making, right? The access to data, we've kind of talked about that. So they have walled access to data potentially. But how do we think about the ability to give our customers the view, I guess, as we've seen almost like the glass box into how those agents are making decisions. How are we thinking about that?

Yeah, so data lineage is one of the very common theme that we hear all the time in our conversations with our customers, right? And what that simply means is that they want to see from point A to point Z, everything that happened in between as the data was traversing from where it originated and wherever it's going and who touched it. They want that full visibility. On that front. So we got that covered, right? So with Data360, you know, especially within the Agentforce trust layer, we are capturing the audit trail of how that, all of that activity is taking place. Within Salesforce in itself, we, as the data is coming onto the platform through Data360, through different streams, it might be connected onto the external systems, the MDMs, the Databricks, the Snowflakes, Redshifts of the world, right? We are capturing all of that as part of that audit, so, and making it available for the reporting purposes as well. So, through that data lineage and through that auditing capabilities that is built on the platform, customers have the ability to see all of that information and provide it to their auditors and regulators if there is ever a need of it, 'cause they can actually see all of that information, right?

Okay.

So, I'll give you a quick example, right? So, let's say that, There's a prompt that was used, and within that prompt, you know, an email generation took place for the VIP client list, for example, for one of the West managers, for example. As that prompt was traversing through the Enforce Trust layer, as we masked the information, as we made the call out to the external LLM, as the response came back, all of this information is kept in the audit trail, right? What the prompt was, before masking, after masking, and what the response came back. All of this is kept and available for the reporting purposes.

Oh, that's fantastic. I think that's a huge relief, especially for those listening that are sitting in the compliance teams or are responsible for that. One of the other things that you mentioned as you were describing that trust layer was this concept of grounding and dynamic grounding. And I think, you know, as you and I have talked about, the value I think CRM or Salesforce delivers, again, not just wrapping the LLM and being able to bring context, could you talk a little bit about that data grounding dynamic grounding piece and why that is valuable and why it is part of the discussions as we think about security, risk, and trust?

Yeah, yeah, that's a great question, right? So we always say that your GenAI strategy is only as good as your data strategy.

Fair, yeah.

So when we talk about dynamic grounding, what we are saying is that data with a context is gonna matter big time, right? Because when it comes to agent force and the autonomous layer, right, it's not just about dealing with data out there 'cause, you know, through the public, channels and whatnot. It's the meaningful data that your business relies on that you as a customer have invested in and consider your crown jewels at the end of the day. So through Data360, we provide the abilities for our customers to bring in those data components onto the platform and then let the agent leverage that information. So as part of that dynamic grounding, That's exactly what we are doing, right? So if we need to look at some information, if the Bowers information is what's being used as the grounded information, the chances are the realistic response will come in. Otherwise, the chances of hallucination will increase because LLMs are LLMs. They're gonna, you know, they have a tendency of going out in the wild and provide you with a lot of information, but not a whole lot of meaningful information, right? And we always say that it's not about creative writing, it's about fact-checking. It's about getting the right information back, especially in the regulatory world, because if you are providing the wrong information without any context, that could be a compliance breach. That can really put you in the hot water, right? So this is why the dynamic grounding is so critical, and the way to do it is having the data with the context.

Right.

And we are providing that through Data360 and Informatica and other means.

Right, and not only provides, you know, the ability for these agents to not hallucinate, which is super important, but I think also provides a tremendous amount of value, right? Hooking into either the CRM data that's in Salesforce platform or even external systems. I mean, we see customers bringing in core banking transaction data into the platform to do things like, you know, submit a transaction dispute as an example. That's so good.

Yeah. And this is where the beauty of Data360 really comes in, right? Through those streams, I mean, you can connect the stream to anything out there. Right, and this is exactly what it is because as you're bringing your external data, we're going to harmonize it, we're going to unify it, and we're going to do the identity resolution, and we'll know exactly what the Aaron's hierarchy looks like. And that is then available for the agent force experience as part of that dynamic grounding for your needs around those use cases.

Okay, there's two more topics I want to hit. The first one is, and I think related to this context, is that, you know, we've been reading a lot in the news around SaaS-pocalypse or, you know, is SaaS dead? I think some of the things that you've described in terms of the rich data on our platform and the ability to enrich the large language models to do, you know, things like, you know, have autonomous agents run the platform. Maybe you could just talk a little bit about what you're hearing and what you're talking about with customers in relation to that.

Yeah, yeah. And we definitely hear that Now that, you know, with the vibes in automation and, you know, CRM can be written through by the agents and you can kind of like produce things and whatnot. So again, key thing to remember, right? It's not about dealing with a lot of data through LLM and plugging it to your data lake, right? Try telling that to your CISOs and see how that kind of works out, right? Because security guardrails, meaningful data, knowing the context and leveraging all of that in a meaningful way is the key. And the reality is the Salesforce CRM holds the most amount of data when it comes to the, you know, customer's kind of, you know, customer-related information, right?

Yeah, yeah.

So meaningful data, context, all those things were still coming to play. And the reality is having just LLM connected onto a data lake is never going to be an option, especially for the customers who are concerned with having the right information coming back. Because again, the information that is not right, right, from a compliance standpoint, that could be a big tone for many of our—

Huge. Yeah, huge. And then the last thing I wanted to touch on, because it does come up a lot, is this idea of data sovereignty or even the ability for data to stay within the country potentially. We spend a lot of time in the EU. Maybe you could talk a little bit about what we're doing there and what you're kind of advising customers on in that regard.

Yeah, so Hyperforce really kind of allowed us to get very innovative around a lot of things, right? So data residency was always a big topic for us, right? And then with the EU operating zones, right, now we have this data sovereignty style use cases handled through that as well because What we are doing is through the EU operating zones, through the operational isolation, right, within that country, you know, the requirements can be covered, right? So for example, for Germany, right, just using that as an example, the EU operating zone operating there with the operational folks kind of working from that location and covering the end-to-end really kind of gives that peace of mind to our customers, right? Now Hyperforce is the infrastructure layer, that is essentially exposing the Salesforce platform, on top of it Data360, on top of it the Agentic layer, right? So the underlying layer through Hyperforce, through the EU operating zones is allowing us to satisfy these data sovereignty requirements. Now that's not, it's just the beginning, right? So as we kind of move ahead with these things, right, more and more innovation will come around this topic and will allow us to even do things for the, other regions as well outside of the EU, perhaps, you know, other areas on a global scale.

That's incredible. That is very reassuring for our customers. I want to thank you for your time. I think we've covered a lot of topics here. Was there any last words that you wanted to kind of hit on before we close out?

No, I think, you know, the only thing I'll just kind of do a quick plug here. I know you and I, we talked about a lot of things my team does. A lot of these kind of deep dives, workshops, and these sessions. So, I think it would be good for our customers to kind of know about these and they can work through their account teams and we'll be happy to kind of double-click into everything that we talked about here and not just go through the slides, actually do some hands-on work and really kind of test drive everything.

That is really important. That is perfect. I know you spend a lot of time thinking about these topics. Trust is our number one value, so it's important that we continue to educate our customers. So thank you for your time, Babar.

Thank you. Thank you, Cami.