Turning Intelligence Into Impact | Dreamforce 2026

Thank you everyone to be here. I'm Silvio Savarese, Chief Scientist. I am leading the AI research organizations and incubation at Salesforce. I joined Salesforce 5 years ago from Stanford University where I was a tenured professor of computer science. And today we are focused on what's happening next, about how AI is going to shape enterprise. And it's really a pleasure to introduce Dr. Uri Yerushalmi, Chief AI Officer and co-founder of Fetch.ai. Thank you, Uri, for joining us.

Thank you for having me here.

And Uri holds a PhD in computational neuroscience. His work really sits at really at the intersection of biological intelligence, adaptive decision-making, and software systems. Uh, before turning your attention to enterprise AI, uh, Uri spent years on Wall Street, and here he served as CEO and head of AI at a major algorithmic trading firm. At your peak, the algorithms of the Uris team really were responsible for a significant share of the NASDAQ trading volume. So which is very exciting to see, to read. So Uri, welcome on stage at Dreamforce. Thank you for joining us. And today we'll organize our conversation into a number of chapters. So let's jump into it. So chapter 1, Ambient intelligence. So as I wrote in a recent blog, AI is evolving from being reactive to be proactive, and we'll see an intelligent layer where AI will be under in the background, will be aware of the context of a certain workflow, and will be really ready to act to help users to provide relevant insights. So Uri, in many years, in many ways, Fletcher really operates in a similar space, in a space of ambient intelligence. And tell us what you have built at Fletcher and what role AI plays behind the scenes.

So when it comes to ambient intelligence, I think nature gives us a great example of how powerful ambient intelligence is, because we see that resemblance between creatures that need to make decisions in a given environment and businesses that also need to make decisions in order to maximize their KPIs. And nature has evolved ambient intelligence to be an intelligence that is always there, that listens to the environment and takes actions accordingly. It's not something that someone needs to trigger or to prompt. It's always there. But we in Fetch.ai, we distinguish between 2 types of ambient intelligence. One that is instinctive, that is reactive by definition because it just listens to what's happening and reacts accordingly. And we in Fetch.ai, we focus more on proactive type of intelligence. And for that, one would need to have awareness regarding the environment. And by awareness, I mean having an internal representation of the market around you and allowing that internal representation to be actually a model that can be run in order to simulate possible actions that are made. Now, once the model is there, it can be used as a digital twin to run simulations and examine all of the simulations and possible decision-making policies in order to choose the best decision policy out there. That's why in Fetcher we have built a large market model that doesn't just react to what just happens. It anticipates what are the consequences of decisions being made in different scenarios and in different circumstances and acts accordingly.

I guess in this kind of scenarios, the challenge is to be able to find the right balance between potential false alarms and false positives, because the thing is you don't want to create too many alarms which are not true. At the same time, you don't want to miss something important, right? So finding this balance, it can be tricky, it can be complicated. Yeah.

Yes, the balance is based on right training of the model. Once you train the model of actual native data of the environment, then give the right We in machine learning like to call it loss function, right? So using the right loss function here is the key because the loss function eventually reflects the reward that the enterprise is trying to get in that sense.

Right, right. Okay, that's great. So let's now move into chapter 2 of our conversation, beyond LLMs. So everyone knows that LLMs have been trained on text data. So it's been fed with billions of tokens, and that we have been training these LLMs to do text predictions, which eventually create these emerging capabilities that we have seen coming up through different frontier models. At Salesforce, we build a different type of models, which, which are built on time series and tabular data. And these are actually different because instead of operates on quantized, on kind of quantized data, on words or letters, it actually operates on continuous signals. So temporal signals that evolves in time or numbers that sits on those tabular structure. So, and but this kind of data has been very, very valuable for really helping us operate at different business level, in particular when it comes to operational intelligence. So these kind of models can help us making predictions, understanding the root cause of certain, you know, certain incidents, or finding even remediations in a temporal scale. So you've been working on also different type of LLMs, the so-called market models, and those has been really built to optimize commercial and operational decisions at scale. So tell us more about this type of models that sounds very, very relevant for many application use cases.

So one lesson that we learn also by looking at biological systems is that in order to make right decisions in a given environment, it is important that the internal architecture reflects the complexities of that environment. That's why, for example, in our brains, we all have a visual cortex that is very good in processing visual data. So the type of data of the environment, in our case it's the market environment, the type of data is very important. Just like what you said, when you are trying to predict time series, it's important to use time series related data. So In our case, we, we train our large market models on actual data of, of business, like transactions data, like shopping data, inventory data, and so on. Now, if we compare it to what happens in agentic deployments that are based on language models, Language models lack that internal understanding, the quantitative understanding of what happens in, in the market. There are some first-class citizens in the markets, some concepts relating to price, relating to competition, relating to inventory. And these concepts need to be quantified. That's why it's so important for us to train on actual quantitative data in order to make the right decision. So in some sense, when discussing, for example, price, it's— language models would be quite good in talking about price, but in order to to understand and predict what would be the quantitative consequence of changing the price, you need a market model. Now, these market models can be compared to time series models or to, to models used in robotics. Robots, for example, in order to understand what they need to do, they need to have a not a language model. They need to have a world model that knows the physical world around them. It can be compared to visual models. Visual models are looking at the environment in a quite specific way. They look at images in visual data. Images are 3-dimensional tensors. They have height, width, and color depth dimension. And in every such composition of these 3 dimensions, there is a light intensity value. Large market models are much more multidimensional. When we are looking at markets, markets have time dimensions, the time of a transaction being closed, time of transaction being delivered, There are certain encodings of products, there are price dimensions, and in each such composition, we call it voxels. In each such voxels, we're not talking about intensity of light, we're talking about probability of something that would happen, like probability of meeting a demand or probability of some transaction to be closed. So we see these models as models that are complementary to large language models, where the main power of the large market model is to make the right quantitative decision in the market.

So in a way, these models are intrinsically multidimensional, to your point, and the correlation across dimensions is captured by the model. So it's the pattern that eventually the model discovers.

Exactly.

So in our, in the scope of LLM and also the time series models, augmenting data with the synthetic data, it's actually turned out to be very useful because you can actually populate the kind of training set, the training corpora with much more data. This is something also you have been experiencing to be useful for your domain?

Yeah, the usage of synthetic data is a key here. But generate— it's very important to know how to generate the right synthetic data.

Right.

Because if you generate synthetic data wrongly, then you are learning the wrong dynamics in the market. So it's very important in our case maybe to use synthetic data as some prior, but The action of fine-tuning on the actual data of the given enterprise is key here in order to learn the market dynamics of the said enterprise.

Wonderful. So now let's move on to the next chapter, which is the hot topic of the week, which is trust. So trust is a key value for Salesforce, as we all know. This morning at the keynote, Mark was talking about trust, the importance of building the right guardrails, the right boundaries for enabling, ensuring that AI is safe to use. It works in an environment which is trustworthy. As you know, I also sit on a UN panel, the Independent International Panel for the United Nations, and we have been advising and giving recommendations on how to build an AI that is ethical and safe. So, Yuriy, recent MIT reports highlighted how Fletcher bakes explainability and transparency into every decision. So you start with this concept of maximal guardrails, human in the loop on everything, and this really resonates with our approach. Tell us more about what's your vision about this.

Yeah, we see trust as something that needs to be earned. in an enterprise. And by earned, we mean that there is this loop in the enterprise where we expect the AI continuously to generate evidence that it is trustworthy and only then to gain that space to act inside the business. That's why when we deploy our systems in enterprises, we start with the very narrow action space.

We—

and in the narrow action space, we— I mean that we do not expect the enterprise to trust the AI in any sense. All of the rules that are there should remain. And then we start having the AI make decision— decisions in the enterprise and to show some trustworthiness evidence. And by trustworthiness, I mean that first, to show some initial results that compared to control group using scientific A/B testing, it is right and can make the right decisions. And also to all of the decisions in our systems are explainable. Meaning we always show the factors that influence the decisions. We always show the counterfactual cases of what would happen if the decisions would have been different. And that is also a factor that adds to the explainability, to the trust that the humans in the system gain in that sense. And once the humans in the system understand the reasoning behind all of the decisions that are made, once they also trust the system because they learned from it, they learned something new about the market, though they are working decades in a certain market, the AI shows them some dynamics that they never saw before. And that leads them to trust the system more and to give it more space to act. And by space of acting, that means to gradually peel off some of the constraints that they initially put on the system and also giving the AI system more higher-level instructions like just maximize my revenue without digging into how exactly to do that. So eventually, once that trust is gained, the degrees of freedom that are given to the, to the AI are gradually getting higher. The humans in the loop feel more educated regarding the market. And in the bottom line, the KPIs are showing a lot of uplift.

Right, right. So in AgentForge, when we build agents, we try— so one focus, one way to incorporate trust is by injecting some deterministic constraints. So the AI, the agent, can come up with their own way of operating, but then it has to operate within certain boundaries. However, this is actually pretty challenging to incorporate from the, you know, from the technical standpoint. So have you also been thinking about finding a way to incorporate deterministic constraints in the decision factors?

The way we approach it is that we differentiate between the deep learning neural network system that is predictive in its nature. That neural network predicts what would happen. And these predictions are not constraints because the predictions are passive. They are just saying what I think, what I as the neural network think would happen in different circumstances. And then there is the optimization phase, which is in some sense similar to reinforcement learning type of system. And then when making the actual decisions, then the constraints are set. Therefore, it's because of the nature of the system that is very simulation-oriented, it is easy for us to put these constraints.

Okay, great. Okay, so now let's move to, uh, our next chapter, which is on multi-agent ecosystems. Uh, so as, uh, plenty of businesses already been deploying agents, uh, in a way that, uh, can not only implement a specific agent for a certain task, but actually assigning multiple specific tasks, multiple tasks to different agents. Then there is this concept of agent orchestrator, which kind of allows to manage this fleet of agents and be able to coordinate the actions. This also requires communication across these agents such that they can actually really exchange the right information to achieve a task. Well, as we— so this is true within a certain organization, but as we move across organizations, let's say from Corporate A to Corporate B, now agents have to communicate across orgs. And here we face a number of very interesting challenges related to how to ensure that agents communicate in the right way within the right guardrails. But also as they negotiate, they also come up with a win-win scenario for both parties. So what is your take on multi-agent systems and how do you see them being implemented for enterprise?

Yeah, we see that convergence into models of multi-agentic orchestration systems both inside enterprises and outside of enterprises. But, but the way we see it is that these, these Agents are basically defined by some authority limitations. They have their guardrails, they have the data that they can use, and so on. But overall, there is a holistic concept in every enterprise, which would be how does the enterprise see the market? And that's where we see the market models come into play. We see market models as being a serving centralized tool where the agents, either agents inside the enterprise or agents that are working externally, they get advice from these market models in order to understand what's the best way of doing actions. For example, an agent that eventually is oriented towards negotiating with some external, external collaborative agent, that agent needs to get the right advice from a centralized market model that knows what are the expected dynamics in the market and what type of, of negotiations are the ones that are expected to gain the maximal reward for the given enterprise.

Right. So in a way, um, when it comes to agent-agent communications and negotiation, uh, the problem of trust becomes even more significant because, uh, any turn the agent can go, it can go into some direction that is not favorable. Then as you you know, kind of look at, accumulate the interactions, the probability of something goes wrong becomes higher and higher, right? So do you feel that in this kind of scenario it would be useful to have maybe a third-party agent, a third-party entity, or that third-party layer that actually can help administering, can do the right administrator for having agents to communicate in a proper way? Uh, this is something we've been exploring at Salesforce also with some partners, some collaborators. And, uh, you know, there are 2 camps. One say, no, agents themselves should incorporate their own guardrails for ensuring the right protocols. Others say, no, we should have actually a third party, should be more like a, um, yeah, a negotiator or a judge model that allows to ensure that the communication is It's properly done.

Yeah.

What's your take?

You know, we, we in Fetch.ai are very focused on having responsible AI, and having guardrails in the right place in the system is an essential factor in all of our architectures. Therefore, I, I, I think it is very important to always have guardrails in a centralized place. You may call it a third party that makes sure that all of the actions that are made are working according to the guardrails that were set. Of course, as you gain more and more evidence of trustworthiness and evidence that certain guardrails are enough and you don't need additional guardrails, then you can start peeling off some of them. But having a central place of steering what's going on inside the enterprise, that's essential the way we see it.

Okay, great. All right, so I think we are coming to our final chapter, which is the AI-powered enterprise. And as we look ahead, how do you envision the AI-powered enterprise of the future? What will be— how different will it be from the traditional way of thinking about enterprise AI? And maybe most importantly, what should leaders do to really prepare themselves, prepare their own organizations to this kind of new wave of agentic AI?

Yeah, so, you know, we have a lot of experience deploying this AI system in enterprises, and it always comes to the point that it's not just having a traditional enterprise and adding on it some AI magic. Eventually, the success comes from some very deep changes in the DNA of the enterprise, in the way that it embeds into it AI. First, the AI being embedded needs to be AI that is ambient, As we spoken before, an AI that has the capabilities to be proactive, and in order to be proactive, it needs to be an AI that models the environment and understands what would be the consequences of certain actions that are made. First, and another point here that is very important in the way of embedding the system. Is that the AI alone is not enough because the AI can do a lot of micro decisions and make sure that all of the micro decisions are well synchronized one with the other and the decisions are not done in a silo manner. But there should be some top-down steering of humans that are Guiding the the targets to be set that are setting the guardrails that are deciding how much do they decide how how much do they trust the AI and for that they need to have the right tools to to understand how much do they trust the system they need to have the right governance of the system governance I mean they cannot humans cannot really verify every decision that is made by AI because of the amount of the decisions and because of their speed. So humans need to have that abstraction layer that allows them to aggregate all of the decisions that the AI makes.

So eventually human has to be the helm of the final— all the critical decisions, has to be still in the loop of making the critical decisions.

Yes, the humans are not going anywhere. They are just moving up In the decision-making process and just setting the targets and the guardrails.

In a way, something that is actually very important in this kind of system is also ensure that AI manifests their own uncertainty. The fact that actually— so if there is something that AI cannot do or doesn't have the right information, the right knowledge, the right skills, this should be manifested to the humans and say, okay, look, I cannot do this. Yeah, I need more information, I need more help. And often they show this system that actually they are very confident in making mistakes, right? So I think one aspect which is, uh, is very important to ensure that AI knows what it doesn't know, right? And, and, and this is to me a foundation for any trustworthy system.

Yes, one aspect of such AI systems is not only predicting what would happen but also having that confidence interval within them. The probabilistic notion of, of AI is important. And when it comes to interactions between these humans that are steering the AI and, and the AI that needs to make the actual micro decisions, they are not thinking the same way. AI is very probabilistic and humans like clarity. Either something is allowed or prohibited. So serving that governance tool is, is, is a key here to allow the humans to to deterministically set the boundaries in which the AI makes the intelligence— serves its intelligence in a probabilistic manner. That's part of the key here.

All right. Well, thank you for your attention. Thank you, everyone. And thank you, Uri, for these great insights.

Thank you. I really enjoyed it. Thank you so much.