How Agentforce Emulates Human Decision-Making | Salesforce CIO Corner

All tech leaders are asking themselves like, is it worth investing in the complexity of AI? Do we want to go to this agentic ecosystem? And it really is only appropriate when this is what you're solving for. You're trying to replicate that human creativity.

Welcome to CIO Corner. Today I'm joined by Aaron De Cesare, the CTO at EasyCator, the number one food tech platform for workplaces in the US that will soon be using agents to handle complex tasks. Aaron, thanks for joining me today.

Thanks for having me.

One of the things that I would like to start with is if you could just tell the audience a little bit about your company, give us some more details as to what you guys do and how you support businesses around.

The U.S. yes, I'm glad you asked that question because I imagine many people are not familiar with what is a food tech platform, and yet probably all of us are very familiar with placing online orders for food delivery. So the difference is really in the nature of the use case. When you think about ordering off of these popular consumer apps, you're typically ordering for yourself or for a small group. So you know people's preferences or if you don't, it's very easy to individually get in touch with people. But when you think about the use cases in the business, it's very different. You're typically ordering for a very large group of people. In some cases, this is a use of cafeteria replacement. So, so it could potentially be several hundred or thousands of orders where certainly the person placing this order does not know everyone's individual preferences or dietary restrictions. And it's super important that the food is delivered exactly on time. This is impacting business productivity. Meetings need to start on time. So I'll describe EasyCater with the term I've been hearing a lot recently and I'm really loving it. It's service as a service. So it's not just software as a service. We're offering technology, but more importantly, we're offering human expertise. We have employees that fully understand the ontology that goes into placing these complex orders, as well as dealing with all the complexity of the logistics throughout the fulfillment cycle. When you think of things like restaurants running out of an ingredient while they're preparing these orders, drivers wrestling with weather and traffic, or on site services dealing with the fact that maybe the serving utensils are missing or there's not enough space for the food, and in all of these cases you have humans in using this external context to make real time decisions. And because that's the core of our business model. We are so ripe to understanding that AI agents are the answer to how we continue to scale.

I love that. And it's really interesting, right, because in your world you have EasyCater, you have your customers and then of course you have all the different groups and restaurants and companies that you work for to support something that by the way, you can have the greatest meeting in the world. But if you don't have the right meals, if you don't have the right food, the meeting is not good anymore. Right? That's the way.

100%. It's such an important part of our day.

Ezcator is soon going to be using Salesforce's agent force. This is our very first suite of AI agents. For many, many of your operations, you'll be using these types of agents. Can you give us just a sense as to how you've gone through this AI journey, why you decided to make an investment in your company or in AI agents, Especially in a place where you have to maintain that connection with the customers. Right. That's so critical to the success of your company.

That confidence that you have a human behind all these orders that are going to make it right is so necessary for businesses to trust working with us. And so as we thought about moving down this road of adopting AI as a means to scale and allowing our human employees to take on more higher value tasks, we really wanted to think about what is the platform that's going to allow us to seamlessly connect the benefits of AI technology, but then also trusting that if necessary, we can bring that human back into the loop. And this is where Salesforce ended up being really an excellent choice for us because this is going to be the platform that our human agents are actively using to understand how to service the customer needs. So as we think about mapping out that conversational UI of what is this interaction that we're having with the customer? In any place that we feel the fallback that the AI agent is incapable of finding what the customer needs, we can hand that back to a human agent and they can seamlessly pick back up because they'll have all the context of what has been this interaction with the customer. And this is giving us the confidence across the organization, including with our human agents, to feel that we are really empowering them in a better way to service the customer.

Are you more excited about one use case than another?

Yeah. So something I'll share that I think is really helpful as companies are thinking about going down this AI journey to really think about it as you have all these Humans performing these tasks and mapping out that ecosystem because that's really what an agent is, it's replicating a human making a decision. And so we've started to list out what are these core simple tasks that we have humans do and how did those combination of tasks. Tasks start to create this more sophisticated ecosystem of AI agents? And starting with what are these places that we're more confident that the AI agent can credibly come up with a reasonable human based decision. So think about the nature of placing an order. So today typically the customer decides what restaurant they want to order from in our ML technology. So machine learning based algorithms based on a few simple prompts like what cuisine type, what the location and head count, we can auto create an order that's going to make sense for that customer and it can be grounded in their order history, what makes sense for that particular event. But currently the sort of that legacy machine learning automation technology where it's not great at, is what I'll call the discovery phase that where the customer might not know exactly what is the right restaurant that they want or order from. And so today they talk to human agents to explore that. And as I mentioned, our human agents has this ontology of food discovery in their heads. So when a customer uses terms like seasonal or healthy, the human understands what that means. Or it's difficult for legacy software to make sense of all those terms without sophisticated labeling and metadata. And so this is where we're starting our use case journey of using AI because this is what it's great at is really understanding that human context of language that as customers are using this unstructured set of prompts, it can narrow down what caterer is best suited for that particular order and turn that into structured inputs to our APIs that then using rag we can validate. Is this actually a caterer that's on our platform? And then seamlessly pull that customer into the order journey.

Yeah, you've given the audience multiple places to actually begin their journey as they continue to move into AI. And, and I really like the way that you described, hey look, look at the entire set of tasks that the humans work on and then start identifying those initial tasks where you can implement AI and then over time keep adding more and more and more to build out a more comprehensive AI agent that supports the customers. I think that's a huge takeaway for our audience to really think about as they try to tackle this problem and use agents in the enterprise.

And one, as I've shared with you, our journey of going down this Path I would love to hear your opinion on as companies are going down this, what are the most critical metrics that they should evaluate and understanding? Are the agents actually having a positive impact on business performance?

I think that's a great question and I really believe that as professionals in it, we are going to have to come back to our businesses demonstrating how these agents are creating value for our customers. Right. So the areas that I think about the most are, for example, back to your examples. Can we demonstrate that these agents are improving the time to respond to a customer's need? Are these agents helping us create a better customer experience? Are we ultimately creating value by reducing cost, Improving, you know, the capabilities that the company has by being able to scale more rapidly as the company continues to grow or as your customers demands continue to grow?

I was going to ask, so I had mentioned, also part of the thing that we're paying attention to is that interaction between the AI and a human. So I'd love to hear more your thoughts on that process of how you can ensure that seamless transition.

You know, I think that's really, really important. And you also alluded to this before because one of the things that we want to make sure is that we don't create choppy experiences. Right. So if you have a customer that it's on EZCaders website and you're interacting with an agent, but it gets to a point where there has to be a human engaged. The level of complexity of the interaction reaches a point where now it's best for you to have one of your human experts support the customer. It's very important that that transition from the AI agent over to the human, it's smooth, it's effective, it's one that doesn't create this very choppy experience where the customer will be more frustrated at the end. But I actually think, Erin, that in time, the agents, as we continue to evolve, are going to get even better at actually doing that. But for that there will be some need for customization. I believe not every agent in every company will be the same.

Yes.

So that's why, as you were asking me that question, I was thinking, hey, how are you thinking about customization of your agents for your customer use cases? And how can IT leaders listening to this conversation understand that not every agent in every company is applicable to everybody else? How do you think about that?

Yes. So I think a term that's really helpful to ground us all in is this notion of being context awareness. Because in such a literal way, it's the simplest way to Think about this, that the agent is the equivalent of a human. And a human can only make as good of a decision as what information that they're working with. I'll transition a little into another key part that we haven't talked about yet, which is the data ecosystem that's powering AI agents. Now we've been talking about this for years in the machine learning space that we know are only as good as the data we have access to. And that's probably even more the case is in the case of generative AI, where the difference here though is how do you be certain that you have access to real time context in the same way that a human needs real time external information to understand what is the right decision. And this is a critical part, getting back to your notion of customizing, because to effectively customize the agent, it needs to have the appropriate context for that particular customer, for that particular company and for that particular moment in time. So I mentioned the example of things like a delivery driver dealing with weather or traffic. These are all prompts that an AI agent would need to understand in that moment. What is the weather condition, what is the traffic pattern? And we're really excited about the possibility of where data cloud is a solution to this. How do we allow the context both at EasyCater and the companies that we service and external data to bring it all into one place where, where the agent has that real time awareness of all of this 360 information to then appropriately reason of what is the right response in that case. And I think whether it's data cloud or I imagine there'll be other solutions. But I think this is the critical part that all technologies leaders need to be aware of is are you creating a data ecosystem such that you can pull that real time context into one place to make it available to that agent?

You know, it's really interesting that you mentioned that Aaron, and of course I knew that this was going to happen in the conversation today because every AI discussion always goes into data. And it's one of those most critical things that as CIOs, as CTOs, we're going to have to really address in the company. You did trigger another question that I need to ask you. How is Ezcator thinking about data and how are you really making sure that data is maintained as a strategic asset for the company that can support the use of agents in the future?

Yes. So as I mentioned, we're in the process of really connecting the fact that our human agents are using service cloud as the way that they are performing their tasks. So it's making a sort of an obvious choice that Data Cloud is likely the solution that allows us to bring our internal customer data transactional order history all into the same ecosystem. So that whether we're having the agent perform the task or then transitioning it back to the human agent, they have that same context. The other exciting thing that we're realizing with Data Cloud is that we can also share information with some of our key clients in a secure way if they're using Salesforce. And this is going to be another thing that technology leaders are having to become more aware of. I've heard the term standardization come up a lot as leaders are talking about where we're going with AI and how do we successfully pass secure information back and forth so that agents can work across systems. And so that shared data eco secure system is a critical part of how we're thinking about pulling all those data elements together.

I mean, we're excited about Data Cloud and you know, it's, it's one of those areas that for many years we've been trying to address. Right. How do we bring data from multiple sources, structure data, unstructured data, make it available to the business to create more insights? Well, I think Data Cloud is starting to get us really close to that answer, which is definitely needed for AI and for agents in our agent force to really work effectively. But there's one thing I didn't mention and that is how do you, you know, how do you make sure that you have the right people in place, the right resources, the right expertise to be able to support, well, one agents in the enterprise, but also people that can take advantage of AI effectively as they do their work. What's your perspective on that, Erin? I'm really curious.

Yes. So this year we actually launched a company wide training program to get our entire organization more comfortable with AI because as I mentioned, so much of our business is, is that our entire company, you know, employees across the ordering experience, the fulfillment experience that are performing tasks every day that we now want to enable and augment with AI. So it's super critical that they are open and excited to partner here because the AI needs to replicate all of the expertise that's in their heads. And so we want them to feel that this is something that's exciting and going to make the company and them better. And what's so interesting is that AI has this sort of mystified terminology that makes it seem so unaccessible. And so that was just simply a big part of the Training was getting people to understand, unlike in the past new technology advancements where you really had to understand this sophisticated syntax that you potentially really need to be able to code to engage with the technology. AI actually is very conversational and, and so much of the training is simply thinking about this software as talking to a human. And that prompting is as simple as asking it questions, seeing its answers, and then figuring out where to go next. And that is in fact a lot of what the training was. Just getting our employees comfortable with this notion of prompting, of having this two way dialogue with AI to understand where it can actually answer the question better than a human, and then getting our employees participating in guiding us of where AI agents really we can rely on versus where they need to be augmented with a human. And I think this is going to be the cultural shift that so many technology leaders are up against is just getting our employees comfortable, that this is something that can make them better employees, that can create opportunity and you don't need sophisticated technology background to engage with it.

I really like that and I think the way that you described it makes me really think about the key points that need to be in your training program to ensure success. Right at the end, there has to be enablement for people to effectively use AI. It's not just something that you throw out there and see what happens. Enablement is important. You actually articulated the importance of making sure you take the mystery out of this technology. And then of course, ensuring that people clearly understand how this complements what they will do, do the work that they actually support. I think those are really key points.

Yeah. And I would love to ask a follow up question here because as I mentioned, we've chosen Salesforce because so many of our employees have this direct interface with it. So I'd love to hear you speak to how AgentForce is actually ensuring that those outputs are interpretable and explainable for the users and decision makers of this technology.

Absolutely. So first of all, I think the user experience, the user interface that we have been building to support the use of agents, it's great. It actually fits in the flow of work. It fits in the way that you engage in your day to day activities. It's not something external. You don't have to swivel to another app to be able to take advantage of these agents. It's all embedded in where you're doing work within Salesforce. So I think that's a huge thing. The second thing that really comes to mind is the fact that, you know, Salesforce has spent a significant amount of time, effort, money in building a trust layer on all of our AI. And Erin, as you well know, you know, we have to be careful when we implement technologies in our business that they are not, you know, at the end of the day, that they're secure technologies that are not going to create output that can put the company in a bad position. The trust layer that Salesforce has built has really helped us ensure that first, your data is grounded or your AI is grounded on your data. Second, that your data is protected. Third, that there's an auditing process in place to make sure that you can go back and see how the outputs are being generated. And then of course, that there are multiple steps along the way to protect that data, to mask that data when appropriate. Now, one of the things that I hear from CIOs time and time again is, well, Juan, if I do this, if I build these agents on Agent Force, do I need to do anything with my IT infrastructure to be able to bring Agent Force into my company? You are at the forefront of implementing agents on Agent Force. What's your take on that, Eren?

Yes, so what's great is that, Baxter, I was mentioning that Data Cloud is allowing us to pull in our existing data sets. It's not causing us to have to completely re architect our data infrastructure. So today we're using Snowflake as our data warehouse. It has connectors to HubSpot, which is a current version of some of our CRM. We have transactional databases that are flowing into Snowflake and Data Cloud is really just sitting on top of that. That allows us to pull in what data we want available to these AI agents. This is really simplifying. We don't need to re architecture. We just need to be thoughtful of what is the data that we want to feed into these agents. And I will just emphasize again, as IT leaders are thinking about that data ecosystem, something that's been really useful as we're starting to map out what is the data that we want to feed up into Data Cloud is back to that notion of listing out what are these tasks that you're expecting agents to do and thinking about them like a human. What would be data or information you want the human to know real time to be able to provide that reasoning. And this is now a much more effective way of rationalizing what are the data requirements and then fundamentally then building the ecosystem that the data is really well fed into our agent ecosystem.

I like exactly what you said in terms of when you think about your agent strategy, think about your agents as if they were human. Think about how these agents would be. What would you humans do? Right.

Yeah. What would you want them to know? Yep.

I like that. I think people listening to this today are going to actually say, well, yeah, you know, these agents work in all kinds of companies. And here you have an incredibly interesting, important company that provides very important services to organizations. And they are going to be in a place where agents will simplify their work, will help them do their work more effectively, and ultimately create value on both sides for your customers as well as for your company. So thank you for doing this, Erin.

Thanks for having me. This was great.

Please check out the links in the description below to be sure to stay tuned for more episodes of CIO Corner.