In February, Madhav Thattai was elevated to executive vice president and general manager for Agentforce. Right before that, Informa TechTarget had a wide-ranging discussion about where the company was going with agentic AI, and what they're doing to help customers find their groove with this technology.
Editor's note: This Q&A has been edited for brevity and clarity.
You and your colleague, Srini Tallapragada -- president, chief engineering officer and chief customer success officer -- discuss "the four elements" that make agentic deployments successful. What are they?
Madhav Thattai: Let's start with context -- there's a lot of rich content created by companies: PDFs, documents, images, tables, workflows. What we did in the past was, Oh, let's take all that content, let's throw it in a vector database, and we'll ground these things, and we'll see what happens." Well, that's not enough. We have to be able to interpret and contextualize all this rich content and use it when we actually respond, so it drives not just the accuracy, which is really important, but also a richer experience.
Madhav Thattai
The second idea is control -- LLMs [large language models] are a profoundly impactful technology because they have changed how we communicate with technology and how we deploy it. [For automated workflows,] there is a structured process that has to be followed, and what we found as we work with our customers is that the more complicated your instruction set gets, the easier it is for an LLM to make mistakes. Now, in some industries, 90% might be okay. 95% might be okay. But for us, we have customers in healthcare, financial services, regulated industries and industries where process adherence is at 99.9%. And so that determinism is really key.
Observability is pretty important. Most of the work in an agentic enterprise isn't just in the build phase; it's actually in the operation. How am I running this enterprise? Am I hitting my KPIs? Am I making my improvements? Am I actually modifying the agent and my processes in a way that they're going to get better? This is a constant evolution. You're adding capabilities, improving the agent and improving your processes.
The last one is orchestration. Salesforce has always been a place where humans orchestrated with each other. Real work gets done on Salesforce. Collaboration happens on Slack. And so when we think of orchestration, we think of it as the orchestration of the system, where it is orchestrating the agents, but also orchestrating our human processes, these handoffs and so on.
Slackbot is the kind of capability where an employee is using it not just in Slack. They're using it to interact with other systems. Maybe their IT system, maybe their HR system, maybe their accounting system. And so you start to see the super-agent interfaces become really important [as customers begin building them later this year to orchestrate human-agent interactions].
Agents require users to evaluate their processes and, in some cases, bend current processes to accommodate agentic AI. That is a pretty big lift for a lot of companies, especially large ones.
Thattai: One of the reasons why people have felt the need to modify their processes is because they have been unable to control the agent behavior. It is our strong opinion that we want to marry deterministic execution with LLMs. That's really important because one of the reasons you have to modify it is because LLMs are unable to execute consistently.
A second thing is that there is a learning process of how these agents work. That does lead to process modification over time. For most customers, if you don't start with a clear outcome -- a clear goal in mind -- and you're only trying to take an existing process and copy it over to an agentic system, we have seen those pilots fail.
Nothing matters more than customers actually driving outcomes. At the end of the day, the technology is a means to that.
Madhav Thattai Agentforce EVP and GM, Salesforce
However, when a customer is aligned at the CEO level, when there is a clear objective, when you have customers that are really thinking about what the right outcomes are -- which we help them with -- then it really becomes more about how the agentic system collaborates with humans to deliver those outcomes. And yes, over time, they improve processes. Those are the customers we've seen be successful.
You and Srini have talked about "prompt doom loops." Unpack that.
Thattai: LLMs are being used at the communication layer to do things like summarize an account, send something to a customer, and help with customer live chats. [For example,] make me a marketing campaign that is hyper-personalized. They're being used at the task layer, but also they're being used at the reasoning layer: They are used to reason through a task and intent and break it down into a series of steps.
This is where the idea of the prompt doom loop really starts to happen. The more complicated the instruction set, the more the reasoning engine is unable to retain its memory -- and its deterministic execution. Srini's engineering team has built this new hybrid [deterministic and probabilistic] reasoning approach. We think that unlocks new, much more powerful use cases than "Hey, I have an agent that's just answering questions."
It's easy for you, as Salesforce, to think about AI -- and data -- at deep technical levels. How do you get customers excited about standing up and maintaining complex agentic AI automations when Salesforce is just one of many systems in their enterprise IT stack?
Thattai: Nothing matters more than customers actually driving outcomes. At the end of the day, the technology is a means to that.
And so, we talk to customers about where other customers in their industry are succeeding. How are they changing the outcomes that they're delivering for their customers? What goes into that? It isn't, "Hey, go build agents for the sake of building agents." It is "What way is your business going to transform?"
Also, the most critical part of a customer's business is their relationships with customers. That's why businesses exist. So when you think about your relationships with your customers, and how your competitors might accelerate the kinds of experiences they are building for their customers and employees, I think we have a very credible, validated point of view on how businesses are going to evolve.
We live on surfaces where people work, whether it's in our apps or in Slack, and we live on surfaces where businesses communicate with their customers. So we have a point of view at the operating platform layer for what it takes to actually run an agentic enterprise, not just experiment with technology. So that's really, that's really how we talk to our customers about it.
Don Fluckinger is a senior news writer for Informa TechTarget. He covers customer experience, digital experience management and end-user computing. Got a tip? Email him.
Madhav Thattai
