Name: What breaks when agents start doing the work
Uploaded: 2026-10-13T17:46:15.122Z
Duration: 25 min 4 s
Description: What breaks when agents start doing the work

Transcript for "What breaks when agents start doing the work": Alright. Kevin, I think we're live. Live in this... Live in the studio, sitting awkwardly close together, like close friends. Nice. Well, hi, everybody. Thanks for joining us. I'm Diane, the CEO of Front, and I'm joined here by Kevin who leads all of our AI efforts. Kevin, it's great to sit down with you and, excited to have this webinar. Hopefully, we'll make it mildly entertaining for everybody. There's live q and a along the way. So if you've got questions, please feel free to chime in. Here at Front, we're building the platform for all of your customer conversations to come together, and whether that's with people, across both people and AI agents, making sure they work... Get to work side by side to get things done. And over, one of the things I think we've been thinking about a lot lately is just how quickly the role of AI agents has been changing. You know, a year or two, I think we're all pretty excited, that an AI agent could go and answer a customer, a question that comes up, could go pull information from q and a or, and do that pretty easily. And I think today, like, what we see now is agents can actually start to do real work. And so we're moving aggressively to seeing software can actually take actions within systems, can start to do very complex things. And I think that's fundamentally a very different job of how we initially thought agents might appear in the world, and it's moving incredibly fast. So Kevin and I are really thrilled to spend a little bit of time here today to talk about really three things. Like, one, what agents should be allowed to do, two, how they know, the right thing to do, and then the third, what happens when they need to work with people or other agents? And so, Kevin, maybe to jump in, let's say, much like many of the people I'm sure that are joining us today, I'm running a support team. I've got an AI agent that's already answering questions. And now I wanna start issuing refunds or modifying customer accounts and actually start taking, actions within the systems that I have set up. You know, what fundamentally changes, and what should I, as the head of support, potentially be concerned about and worried about and thinking about? So, you know, staying the obvious, there's a huge difference between, potentially giving the wrong answer or not answering a question versus taking incorrect action. Yeah. Right? So, you know, we just had our customer advisory board talking to customers, in Chicago a couple weeks ago. You know, folks would talk about how it is fine to... For an agent to give a $20 credit. Sure. If it's a $5,000 refund, that matters much more. Right? And so I think for companies, the question is how do you get the agents to actually, do the right thing? And, you know, the key the key here is to have access to the systems that the, the humans would need Sure. In order to make this type of decision and also for all those... The decision making to really be, embedded into the agent. Yeah. I think it's really interesting. You know, a year ago, I think back, you know, we've been in the space for a number of years, and, I always look at billboards. You know, we're here in San Francisco, so you see billboards all the time talking about AI agents. And I think, you know, a year ago, I think everyone was really focused on deflection rates and automation rates. And that was almost like the holy grail. And I think what gets missed in that is we're not talking about the customer experience, right, of actually are we providing a better customer experience as opposed to just saying, like, we can go automate or deflect something. When we think about, I think, you know, one of the questions I constantly think about even for our our own support organization is, how do we know when an agent is ready to take on more work and give it more autonomy? And I think we all wholeheartedly see, like, agents are pretty good at going to answer the q and a questions that come in, the basic, the basic questions. And so if you're running support, like, how do you know when to give the agent more autonomy to go take complex actions across these systems? Like, what should be... People be thinking about measuring? What are the things that they should be thinking about taking into consideration? And I think part of this also brings into, like, should you be focused on not just the mod... The best model, but also, like, how do you provide the right context for the agent? Yeah. So the progression that we see most companies go through is that they always have this Front of trust. Right? So there's gonna be certain things that you know the agent is good at. For many companies, this is simple question and answer. And then what they'll do is they'll say, alright, what's the next tier of complexity? And then the agent that's able to kind of tee up Sure. Let's say, refund requests or credit requests that require human approval. And then once you get enough, reps, right, like, 20, 50, a 100 examples where the agent has done the right thing, humans have approved it, then you actually reconfigure the agent to say, alright, go ahead and automatically send those. But then the next tier of complexity, let's now tee that up for human review. And, basically, folks go on these progressive stages, and then over time, are able to drive, the the automation rate right up with confidence at each step of the way that the agent's doing a good job. Right? So that's usually the the progression that we see. Yeah. We talked to... You and I both talked to a lot of customers. You know, I think one of the other things we would recommend is, you know, pick and choose, like, the number of workflows that you're looking to automate. I think a lot of times in my experience talking with customers is, like, we wanna immediately go and say, like, hey. Let's go solve all 40 of these different requests that come up. And things are pretty complex. As I said, like, we're not talking about the basic q and a. You're talking about integrations into different systems. You're talking about having the appropriate context to go and take action. Do you have, like, some general guidance, to people that are listening and thinking about this as to, like, where do they start? Should they focus on just a few things? Can you actually go take an agent and say, actually, let's go and deploy it across a broad swath of things because perhaps we get more insight and then can figure out word and error from there? Like, what are your general recommendations and guidance on that? So, for customer operations, like, we've done a lot of analysis of the data and you see that there is, like, a higher law distribution where, the most common, topics are... Come up far more often than the long tail. Right? So I would say that when it comes to, specific processes that you have to, you know, teach the agent, Certainly, focusing on the top five or top three will give, you know, disproportionate bang for the buck. Right? So so I think that that's true. Now that being said, I would say that when it comes to, just general, question and answer, you do not need to actually split them out by topic. Right? That Yeah. Agents should actually be able to just retrieve the right information, and then you should consider that entire class of of, question and answer as just a single single, yeah, class of problems that the agents can handle. Sure. And then, you know, a lot of times, we pay attention to the latest model. We see this all the time almost every week. Right? The the next Front Lab, releases a new model. We see the public benchmarks around it. How should people be be be thinking about balancing, right, access to the to the the best model to go and take action and relying on that versus, providing the right context, to the agent. At the end of it, yeah, how do we think about, like, that balance? So I think the way you think about whether a model is doing a good job is is it following your instructions? Right? And as... This used to be a big problem that models would, like, not do what you told it to do. These days, by and large, even the frontier models or one generation back are able to do that. So then the question is, what is the right context required to make a decision? Right? And so, you know, just give a concrete example, we have lots of customers that, handle, distribution of products. And, one customer was telling us this example of if a shipment is still in the warehouse, it's way less expensive to be able to cancel it at the the order. Whereas once it's gone out, right, the the cost associated are much higher. And so, when they're making a decision about whether to permit a counselor or what to do, like, knowing where exactly it is in the distribution chain is really important. And so, in that case, you need to make sure that the agent has access to that system and knows to look up that that piece of information. Right? And, like, that actually becomes where, agent tech performance really really shows up, whether you can provide the context that's specific to your business that matters. Yeah. I think it's a really important point. Right? Like, the context, becomes, really the the most important thing to pay attention to. Right? The access to the systems to help empower the agents. Like, obviously, you don't want agents powered by suboptimal, you know, models. But as I said, it really all comes down to the context and understanding access to systems, access to tools, that you're providing. And I think this this also creates, like, another interesting dynamic, which, you know, something that that we wholeheartedly believe upfront and made announcements on just last week was, really our open platform where, our fundamental belief that, look, more agents are gonna show up, in the workplace, specifically around customer facing work. And, we're almost now creating, like, the the same dynamic that we had many years ago of we got a lot of people working on customer problems. How do we make sure, like, everyone has the context and understanding of the customer? And now we've... Are now gonna throw a bunch of agents into that mix as well. And so how... You know, what advice would you give to somebody in terms of thinking of building these systems of, you know, should they be thinking about just having one agent? Should they be thinking about having multiple agents? How do they think about actually coordinating the alignment of these agents with people? So I would start from the customer experience that people wanna deliver. Right? What you want as... Is for your end customer to feel like regardless of whatever is happening behind the scenes, right, however many agents, however teammates, they get one unified experience and the problem is just solved. Yeah. And so, what this means is that you need to make it to the extent that there are multiple agents or humans that the handoffs are clean. Yep. Alright? And what this looks like, you know, we... We've actually learned this from working with human to human handoffs for for many years. A great handoff is what did I do? Why did I do it? What did I look at, right, to intermediate information, that and, we have spent a lot of time making sure that agents provide this context for other agents and also for the humans that they collaborate with, Right? And so that you don't have the experience of the end user having to explain themselves multiple times, right, to whether, agent or a human. Yeah. It's my favorite part of, customer support when I have to repeat myself multiple times when I when I get frustrated. But I think that highlights, like, a really important part and, you know, one of the bets that we've obviously made as a business is that we think, like, people are gonna have multiple agents, servicing customers. They're gonna pick fewer platforms to actually coordinate or orchestrate that work across systems. And I think the biggest friction point that's gonna show up in customer experience is around those handoffs. And so a lot of times when we think, you know, even internally selling against our our competition and and positioning our platform, it's all about, hey, go show what the handoff looks like. And I think to your point, the handoff needs to be really seamless. That needs to be really seamless across channels. It needs to be really seamless across, a person to person handoff, a person to agent handoff, even an agent to agent handoff. And so these all create like very new dynamics, within, as I said, within within the space. And then perhaps, you know, as you think about, how people should be thinking about, unifying the experience in these handoffs, like, any additional recommendations that you would give, for somebody to think about? You know, should I go and align or pick a platform that can do all of this in one as opposed to saying, Hey, should I take an open approach and pick one platform that can go do the orchestration and actually have multiple different agents? Maybe you're going to have a voice agent, maybe you're going to have a chat agent, maybe you're going to leave people on the e mail side. What recommendations or advice or just simple perspective do you have on that? Yeah. Yeah. So I would say a few things. First of all, all of you are the subject matter experts about your business. Right? And so one thing that I see many of our customers doing is kind of building their own prototype agent. And I actually highly recommend that because the... There, you're able to prove to yourself, like, on on a single day or afternoon, whether or not an agent can follow your business instructions, right? And, but then the second part is like, okay, how do you actually get this into production? And then all of a sudden, you start running into, like, scalability issues. Like, what if you have five requests at the same time? How do you have the infrastructure or a 100 requests at the same time? How do you have the infrastructure to be able to field that? How do you monitor customer experience at scale? Right? Like, can you actually see whether or not having agents involved increases customer satisfaction or not? Yep. And so I think those are the places where you want a platform to do that because, whereas encoding your business logic is something that is a really good use of time for every individual customer, building the infrastructure that every company needs for their agents is less use... Good use of time. Right? Yep. And so, you know, we seek to make it easy to do, on Front, but we are also totally open to you, hosting that, you know, somewhere else and then bringing it into your front end as well. Our view is that, you know, the industry is moving so quickly that we want to be a platform where those handoffs are really seamless, regardless of whether or not your agents are Front agents or, agents that you're hosting through a different service. Yeah. You mentioned something I think is really interesting and important. Right? We talked about, I think, a lot of focus initially on automation and deflection rates. I'm... You know, I alluded to the fact that, like, look, you actually wanna be focused on the customer experience. I wholeheartedly agree. You know, at the end of the day, you want your customers to experience one company, one brand, regardless of how they've reached out to you, regardless of whether they're interacting with people or agents. I I think that's, like, the ideal experience. As we think about measuring the impact of agents, you know, how do you think about... What are tools that people can leverage to go and do that? You know, CSAT is, you know, traditionally, people would send surveys out. I think we've got some pretty interesting things now around, oversight and governance of agents so we can go infer customer satisfaction across every interaction that an agent has. So these are ideas and things that I would push people to think about leveraging the QA tools, so the quality assurance tools, and the inferred sentiment, and the inferred CSAT tools that are there because you can deploy that across the agentic experiences and suddenly see, you know, is my agent actually having a positive or a negative experience, on the customer? Any other ideas or suggestions that you would give to to people to consider as they deploy agents, not just to be focused on automation and deflection rates? Well, so I think just to talk a little bit about how we do CSAT, you know, one of the problems traditionally with CSAT is that low response rates, right? One in 10 people, on average, respond. They tend to be disproportionately five stars or one star. One of the things you can do with AI, is to actually kind of read the customer interaction and just infer, hey, what's the level of effort? How happy is this customer likely to be? And if you can do that for 100% of the conversations. And then, you know, what we found very effective is to then couple that with, the reasoning of the agent, that leads to lower CSAT scores versus very high CSAT scores. And that actually reveals the next frontier of improvement for the agents, right? And so, you know, you actually don't need customers to tell you whether they're happy or not. You can actually just tell from the interactions themselves. And then, like, the next level thing you can do is to actually go and look at the systems to see whether the underlying complaint that people had were solved. Right? So, someone's complaining about errors within your system, you can see, hey, did the errors go down? And so, I think that's really the next frontier. Yeah. Right? I think these are all some of the things that that we announced not to, you know, to to plug all of our own product, but I think these are, like, general guidance and recommendations. Right? It's thinking about, yeah, obviously, monitoring the deflection, the automation rates, to be thinking about inferred customer satisfaction to Kevin's point. You can go deploy that across every channel across all of the agentic experiences and suddenly see, like, did you have a positive or a negative customer experience? Where are you seeing gains? Where you might be seeing some degradation? And then the last part is all around the agentic experience, which is really having... Does the platform provide you with the oversight to say the agent followed to this process or didn't follow the process or didn't know what to do? And then do you have rein... The the reinforcement learnings, the simulation, and the training to go and make changes, and put them back in the production on the agentic side? Yeah. Anything else you would add there or or thoughts? No. I think you covered it well. Nice. Well, Kevin, there's obviously a lot more that, we can get into here, but wanted to open it up, to some live questions. I'm sure folks, we've seen a couple of questions come in, that are live. But, you know, one of the first ones, that immediately came in was, and this is from Julie. We're already using a couple AI agents, and we're working pretty well. How do we know it's safe to give them more autonomy without risking the customer experience? So I think this is going back to the, the Front trust thing I was talking about where you want to have the things that are on the frontier be undergoing human review. Sure. Right? But you do that temporarily so that you collect enough data to understand whether the agent is doing well enough. And then if so, then you flip the agent to saying, all right, we'll automate this and then, you know, consult humans on the next level of complexity, right? And so, I think that is the methodology that is fairly proven through many customer deployments of how you you kind of build up to a large swath of agentic actions that have high customer experience scores. Great. And then we have another question. So I'm just reading that... Trying to read the the small names. And this is from Greg. We're evaluating AI agents from a few different vendors. Should we be trying to consolidate onto one platform, or is it realistic to expect agents from different companies to actually work together? So, I would say that one thing that you want is a unified customer experience, right? And so, you do I think you would prefer to have the same agent running on all of your different channels because you don't want someone to get a different answer or different, like, you know, refund policy if someone is emailing you versus texting you versus using WhatsApp, right, or chat. And so, I think, from that respect, I think it makes sense, to try to consolidate. Now, there will be times where there are... There's privileged access to certain systems and you want, or like, we have many customers that end up building their own agents to do, like, the core of their business, right? Like, you know, if you are if this is a business that, like, bids on contracts, the mechanism by which they choose what you bid is, like, core to their business and makes sense for that to be their agents. And so I think, in that case, like, it does make sense, maybe to, have have multiple systems, but you do have to make sure they talk very well to each other. Yeah. So that, again, the experience for the customer is seamless. Yeah. I think the important part, like, you know, obviously, we're a little bit biased talking about this and and respond to, like, a live question. Like, you know, I very much believe that, like, like, you're gonna pick... There are gonna be fewer platforms that show up in business that people consolidate to. And then you're gonna have, a lot more agents that show up. And so you want consistency across, those agentic experiences. So you want, you know, whether somebody's engaging with you over voice or chat or social, SMS, you want that agent to sound the same, is one of the things I would argue... That that I would argue. And so that really comes down to perhaps the con... The instructions, the tones, the systems prompts for those agents. And then I do believe the biggest friction area that we spoke about is, like, where are the handoffs? You know, what breaks in the handoffs? And I think that's really where suddenly the customer experience is gonna shine or fail For somebody. Is is somebody gonna feel this completely disjointed experience because you've got these agents, that perhaps, they need to focus on on, on the context and having the seamless, as I said, the seamless experience. And, we very much believe that, like, you're not gonna just go deploy one piece of software to go do all of those things. I think that's, like, a really big task. And so, again, you're gonna pick a platform that can help orchestrate and provide the governance and oversight, and then bring people and agents together. And so I'm super excited about some of the announcements that we had last week. If you missed them, please go check them out. You can find us at www.front.com. But our bring your own agent open platform is, something I think, we're incredibly excited about, and it's a great vision for the future. So I think, ultimately, you know, to to to wrap this up, Kevin, and bring us home, like, I think the biggest takeaway for me is that, you know, ultimately, agents are starting to do actual work, and we need to obviously rethink, how we're managing them. I think that's probably one of the biggest opportunities it needs it out there. It's not enough for an agent just to give the right answer anymore. I think we see that. We need to understand what it's allowed to do and make sure it has the right information around the customer, and ultimately give it a way to work with the rest of the team. Because, again, I don't think these are places where you're just gonna silo agentic work off. You're actually going to see that people are gonna interact with the agents to get work done. You know, we fundamentally, as I said, like, don't believe, like, one company is going to build, every AI, agent. But we do think that all those agents need to ultimately have a single platform and a surface area to work together alongside all of the people, that we have, right, in in in our jobs and in in our workspaces. And so that's a big part of... Ultimately, you know, you and I have spent a lot of time thinking about this strategy and execution is a big part of why we've opened up front. And you see me smile and laugh and super excited about it. I think it really is the the future of having a platform that can help you go orchestrate work across people and agents. And then in our next session, we're actually gonna invite our CFO, Meredith Finn. She's gonna come on and chat with me and really tackle the big question, which is, you know, should you be building your in... Agent or should you be buying one? We came out of our customer advisory board, last week. We had 24 customers that are there. Every one of them was experimenting with building agents internally. And so that was, I will say a surprising, and shocking, raise of hands that I saw. I'm not terribly surprised by it. But I think, again, it opens up the opportunities that these are, like, very real topics that are top of mind, I think, for everybody, not just taking customer operations, but, in every function of, like, do I go build an agent to go and do something? Should I go buy one? And then, ultimately, how do I plug it into the platform where my people are actually doing work? And so I think it's incredibly interesting times. Anyway, Kevin, thank you for being here today. I appreciate your insight. Pleasure. It's a joy to work on this problem and opportunity with you every day. And, hopefully, we shared some great learnings with the folks that are there. And again, you can find us at www.front.com. If you missed our announcements last week, please go check them about... Check them out about our open platform, and feel free to reach out. You can find both me and Kevin. We're pretty active on LinkedIn, so feel free to reach out. But that said, thank you for all of your time and listening, and, hopefully, we'll catch you at the next session.