Learn Conversation Design to Create Reliable Agents | Dreamforce 2025

So every single day, customers come to you to get something done. And the reason they come back is because of trust. Trust in what you deliver. So today the question that we'll explore together is what would it take to create AI agents that are truly worthy of your customer's trust? So every single decision that an agent makes, how it remembers things, how it reasons how it responds, and if it escalates to human, that is a choice. A choice to reduce frustration or to create it. A choice to build trust or a choice to break it. Right. So the way that we design conversation is so much more critical now than it ever was. Today I will show you the playbook for creating human centered, reliable agents. Step one is mechanics, and this is how an agent behaves so that people don't get stuck. Step two is metrics. So how do we know what good looks like? Well, we have to measure it, and that's measuring in human terms, meaning did the person feel understood? Did the job get done? And third, guardrails. This is the protection that we put in place so that people feel in control and safe. So show of hands here. Has anyone ever talked to an agent that just didn't get it? Oh my gosh, almost everyone. Okay, so I'm here to posit that trust isn't broken by AI, it's broken by bad design. What I mean by that is if we don't design for memory, for decision making, and for the logic that carries context forward, then experiences get broken. Do we have any parents in the room? I have a seven year old myself. Her name is Echo. Yes. Okay, so imagine the first time you try to apply for childcare. Say you're interacting with an agent for the organization. And the agent loops, it repeats questions over and over again. And then you get to the end of the flow, right? The agent doesn't submit your form. You find out about this like two weeks later, by the way. Right. And so it's not that you're thinking, oh, AI is broken. You are thinking, wow, this organization does not care about me. Has anyone ever lost their wallet before? Happened to my friend, like last week. We are racing around to try to figure things out. So you have. Yeah, he did. I'm sorry. Okay, so imagine you've lost that wallet. You're racing to freeze that credit card, and then the agent, like loops, asks you all these questions, repeats the questions, doesn't even escalate. You're not thinking, oh, AI sucks. You're thinking, someone is spending my money, honey, and it's not me. So that's not good. Right? That's a frustration that can definitely be mitigated by conversation design. Even a more serious high stakes situation. Someone who is dealing with a heart condition calls into an agent to report chest pain. The agent misfires and states that this isn't an urgent situation. That's not just an AI failure or, or a design failure. That's a human failure. But we've learned so much with the failure of conversation design in the past. We took it from scripts which were conditional logic, if X, then Y to intents. Really mapping words to types of intents and then decisions. But let's zoom out to look at the history of human to system communication and how we can't keep designing how we used to design. So we started with telephone menus. A friend likes to lovingly call these phone jails. Think press one for sales. Dos para espanol. And how many people have just got stuck yelling operator into the phone? I certainly have. So customers really hated being trapped by these menus. But then we moved on to something called scripted chatbots. Very familiar with them. Again based and grounded in conditional logic. But bots tended to follow canned scripts. So if there's a word that triggered the bot that it didn't really recognize or was familiar with, then you're lost and you're stuck. We moved on to still smarter intent based assistants. They did feel smarter, but they were based on being trained by limited data sets. And if the agent really didn't know what you wanted, then it also got stuck. Right. And then who remembers? I think it was like end ish of 2022 when ChatGPT became a thing and everybody was pacing around saying, oh my God, I'm going to lose my job. This is so smart. It's amazing. But then we found as we dug into it that these types of LLMs were prone to mistakes. So we're here now in the era of agent behavior design. With AgentForce, we're not just designing words, we're designing actions, we're designing decisions, we're designing moments to moments that are life changing for people. This also comes with the responsibility of being human centered in all of the decisions that we make to shape trust. This brings us to agent Forsh, which is the foundation of reliable agents. So these are three key components. Memory, tasks and rules. So memory is what the agent remembers and what it reuses to be able to create a seamless experience with you. A task is what the agent does. So people should come and leave with what they came for. So the agent needs to call on tools to act on your behalf. Then finally, rules again, making sure that people feel safe and that their data is protected. There are different levels of rules. We'll talk about guardrails, which are the ethical seat belts, the values that we hold as humans that we expect agents to follow. If we think of agent force as the raw power of AI, the skeleton, then conversation design gives it muscle recovery, ability to adapt, be agile, be able to predict needs and anticipate. This takes us to the first part of the roadmap, which is step one, the mechanics of behavior. Let's start with clear goals. I like to think of this as really being crisp and concise with the outcome in mind. An example, schedule a service call for Tuesday 2pm and send confirmation. So we've got the outcome in mind. The agent needs to know where it's headed to be able to fulfill the outcome. But then we've got context. Right? So it's really this idea of remembering history. I see you called with the same issue with your AC in July. Would you like to book the same technician? Very simple. Really understanding the history and the interaction and really using that, leveraging it to create a seamless interaction. And then we've got the tools and actions. So this means that the agent needs to call the booking API, the scheduling API, and understand if the slot for that particular technician is free or if they're filled. But it also needs to understand boundaries. So if there are no slots available, does it decide to hand off to an agent, a human agent, or does it do something else? So these are the four levers that turn raw power of AI into reliable behavior. And these are the four levers that conversation design leverage to create reliable agents. But how do we know what's working right? We've got to measure it. And again, we're measuring in human terms, this doesn't mean uptime or latency. This really means how did the human on the other side feel? The first thing is turns to resolution. And this metric is about time. How many turns does it take for the agent to solve the issue? Did it take 20 seconds or 20 minutes? I like to think of it as. Teams generally sit across a spectrum of performance when it comes to AI agents. So at the functional level, it works, but it takes probably as slow as a human chat to be able to complete. That's eight to ten turns. That means back and forth eight to ten times. Agent says something, customer says something. That's as slow as human chat. So it's not really driving efficiency. But then at the fluent level We've got a concise but polite interaction. The turns are reduced in half and the exchange is resolved faster. The conversation is smoother, much more natural. But at the trusted level, this is where we want to be, right? The agent is able to anticipate needs. It resolves in minimum turns, about five or less, and it feels frictionless. The outcome here that we're aiming towards is that time saved equals trust earned. Moving on to containment now, this means is the agent able to handle an issue on its own, or does it have to hand off to a human? This measures independence and fallback behavior. At the functional level, there's frequent drops without a lot of context, and one and two issues don't get resolved by the agent. At the fluent level, handoffs are smoother but clunky, and we've got up to 75% containment. Better, but better still is trusted. There's rare escalations, seamless handoffs, and we've got almost up to 100% containment. What does that mean for the human on the other side? So less bouncing around means we have more confidence in the system. Hallucinations. Everybody knows what this is because maybe you've interacted with AI that hallucinates. This is a measure of accuracy. Are the responses grounded in data and facts, or are they just made up? At the functional level, we'll generally see that the agent is probably thought of as unreliable, and that's because about 20% of the answers are fabricated. At the fluent level, we're mostly accurate, but there are some guardrails that are inconsistent. What do I mean by that? Maybe there is that same agent that's booking. It's allowed to book an appointment, but it's not allowed to renew a contract, but it goes ahead and does that. So that's a moment where a guardrail is inconsistent. And then at the trusted level, there are rare errors. This is where we want to be. It's grounded in verified data. But the one thing that I do want to point out is that when it's unsure, the agent clarifies, asks clarifying questions to get to the resolution that's needed. The outcome here, accuracy protects credibility. Task completion. Did the job get done? Did the flight get booked? Did the appointment get booked? Functional. So there's a lot of retries and failures. And so what happens is about half of these tasks are not completed. At the fluent level, we've got 70 to 80% completion. And most common tasks succeed, though there are some edge cases that may fail. And at the level of trusted, we're reliable. We've got 90% completion, tasks are validated and confirmed. And the outcome here is that reliability builds adoption. So moving on to arguably the best, most important critical one, user trust. Did the interaction feel reliable and professional? At the functional level, you'll see that humans are unsure, so they're double checking with other humans. Fluency means that the users do accept some answers, but they're still skeptical. They're second guessing. At the trusted level, customers reuse use and recommend the system. The outcome that we want is that first time use turns into habit. This is the cheat sheet of what good looks like for metrics. The outcome that we all want is that clarity, speed and confidence then equals human trust. And so now I'd like to move on to the guardrails again, the ethical seat belts, the value system that we hold as humans and how to align that so that an agent knows what are the rules for ethical behavior. Then first off is clarity. What this means is the difference between it looks like your renewal is coming up versus March 1st is your renewal date. How would you like to proceed? This clarity means that the user is not guessing. It knows exactly what it needs to do. And the agent is acknowledging the historical data that they have on that interaction. Here, context not only is recalling and reusing information naturally, but importantly, it's that the agent knows what data is sensitive, what data not to expose, and has the discretion of being able to make sure that it doesn't expose that data. Transparency is that users see when the tools are used and why. So it's this idea of the agent showing its work. There's no black box. The user knows exactly how the agent is going to achieve the goal. In this frame, trust is really the idea of designing recoverability into an interaction. This means that when the agent fails, and it will, we build in the measures to make sure that the agent knows how to recover from errors. Then finally, safety responses that honor user intent and escalate to humans smoothly. I really like to think of these guardrails as the promises we keep to every single customer on the other end of the experience. So now you've got the playbook for designing trustworthy and reliable agents. And I really like to think of it as putting humans at the center of AI. You've got the mechanics to stop, frustration, metrics to prove human value, human human impact, and guardrails to protect people. Because reliability isn't a feature, it's the shape of relationships that we get to build one conversation at a time. Thank you so much.