Author transcript
Sean Nolan: Hello there, it's Sean Nolan from CX Today. I'm really excited to be discussing AI agents, which play such a crucial role in how CX teams operate today. And who better to join me for this discussion than Brian Donohue, VP of Product at Fin. Hi Brian, thank you so much for joining me on CX Today. So in this discussion, we're talking about how best to optimize AI agents for customer service. And when it comes to AI agents, I want to start off by asking you, what would you advise in terms of managing an ecosystem of agents within your business?
Brian Donohue: Yeah, I think unquestionably we're in a space where there's going to be an increasing proliferation of agents. We're still in the early stages here, and there's just going to be more in multiple places as companies are using these across the business, internally, externally facing, and as they interact with other companies. So this will be a bumpy path forward as we're in this proliferation of agents. But I think what's important is for companies to realize and think about, hey, where do I need to really protect this and control the experience, and where is the right place to encourage lots of experimentation? And I think when it's customer-facing, it seems pretty clear that's where you're on the hook for the quality of that experience, and so therefore if you're in a space where there's multiple agents, there's just higher risk and more unpredictability. You're really multiplying that because AI is unpredictable fundamentally, and then multiple AIs intersecting increases that dramatically. So I think it's for companies to choose where to take risk and be experimental to push the boundary, and choose where to be rigorous. For the agentic experiences you're putting to your customers, I think it's critical that companies say, hey, we need rigor here, we need to treat this like a product, not like an experiment. That would sound obvious, you'd say of course people would think that, but I think over the last couple of years, we've been surprised by companies being more experimental here than you would expect. So I think that's the starting place where I'd frame that.
Chapter 2: Avoiding Data Silos and Disconnected Customer Journeys
Sean Nolan: Yeah, absolutely, thank you so much. And you mentioned there about risks. One risk that we are always thinking about at CX Today is how actually having this proliferation of AI agents can create data silos and disconnected journeys, which can then harm the customer experience. So what's your perspective on that? Is there a way to manage that risk?
Brian Donohue: Yeah, I think we're still early here. We've framed it as our customer agent vision, and we don't want to recreate the current reality, what existed before, where every part of the org who had a touchpoint with customers, anytime a customer traversed that, it was a mess. It was broken, there was no handover, there was loss of context, and the overall experience was sucky. And really this is the promise of AI, that we're no longer constrained by the fact that one person can't possibly traverse all those cases across sales and service and be a deep product expert and get a customer to success. That wasn't possible before, with AI this is possible, and therefore as a starting point we should be removing those boundaries. So fundamentally, it's hard to almost argue with that vision, and then it just comes down to pragmatic or practical constraints, where people say, I'll optimize for that instead of the overall experience.
I think there's another interesting space here, which is something that Des, one of our co-founders, has said almost since ChatGPT launched: AI is a convergent force. And I really think this is time and again proven to be true. Any walls or barriers or boundaries you put up, usually AI knocks those down, because it inherently doesn't see those boundaries. But we still, to progress, need to put up some boundaries, some scope. Here's one example of where we're working with this now: right now we sell, build, and offer three agents. One is an agent that our customers use to talk to, for their customer communication, that's Fin, what most people think of as what we sell. We have a co-pilot, which is for the team working in the inbox, where humans are talking to customers and have an agent working with them to be more efficient. And then we've got a third agent, which we call Operator, that's for the team who manages the support operation, helping them understand what's going on, how to improve, and actually doing those improvements for them.
I think in a future world, maybe a year away, maybe further, these will ultimately converge, and there will be one agent doing that. But you want to step towards that. For example, the co-pilot talking with your teammates in the inbox, you want the same context of your customers, the same knowledge informing all of this, the same guidance, here's how we write and respond and talk to our customers. You need that shared across both. And then it's interesting where the boundaries are, like what should just be for a human, or why can we not expose this to our agent for customers. So the boundaries often don't hold up well, but I think right now they're still important. So the state we're in now, we feel like these three agents make sense, in time they too will converge. But fundamentally you want to be operating from a space where it's easy to remove, to keep the silos to a minimum. So the context, the history, and things like memory can actually be shared. That's what we're aiming for.
Sean Nolan: Yeah, absolutely. I think as you say, the shared context there is absolutely so critical for creating that seamless journey for the customer and making sure their experience is really smooth and that agents are operating effectively. And you mentioned there about the work between human agents and AI agents.
Chapter 3: The Rise of the "AI Colleague" and Human Agent Training
Sean Nolan: So I wanted to ask about a trend that we're seeing, the rise of the AI colleague. What does this AI colleague's existence actually mean for the skills and training of the human agents that you have?
Brian Donohue: It's pretty huge impact. Franco, one of our directors of customer support, actually just posted something on LinkedIn a couple of days ago that was super relevant to this. The starting point is the job itself for customer support has changed so much. For companies who are maybe on the frontier, an AI agent is already doing the majority of their customer interactions, we're talking 70 to 80%. So of all the inbound volume, 80% is handled by the AI agent, and the remainder is necessarily the hardest, most complex, or most frustrated customers. It's a whole different category of work. And so what's interesting is, how do you train new people up to start with the hardest stuff? This is relevant for customer support and also generally for white-collar work. Franco's like, hey, I'm trying to figure this out now. What's interesting is, the humans need to train on what the AI agent is already doing, you read those conversations and learn from this. Effectively what it means is the humans have no ramp-up of starting with the easier stuff to build confidence and then getting to the hard stuff, and the challenge is you still need the knowledge of that easier stuff to be able to competently answer the harder stuff. So this is a generalizable challenge for white-collar industry, and where support teams are getting those high impact rates from AI agents already today. I think that's one of the interesting things, you start by reviewing what your AI agent is actually doing.
The other part here is how much our view has changed on where there's collaboration between AIs and humans. One example is where previously, if Fin couldn't answer a question, it would just hand over to the human and summarize the situation, it was a very one-way door, goes over to the human. What we're building now is way more of a genuine loop, where there's genuine collaboration, and instead of being a one-way door, it can be a two-way door. So Fin will go, and you can actually set a confidence threshold, and Fin will say, hey, I'm not sure about this, here's why I'm not sure, here's what I think I need from you so I could answer this. Or, I think I can answer it, I just need your approval here. Really the same way a human would. So it's like Fin acting exactly as a human would if they're six months into the job and dealing with an awkward situation, tapping their manager on the shoulder and saying, hey, I actually need some help, I think I should do this. That kind of collaboration is what we're now building into the system, and I think is a real unlock, it almost changes your mental model of how you think about those AI agents, feeling a little more human-like, not such a binary one-way door handover where agents do this work and humans do that work. This is what's interesting, we're in a space where it's just way blurrier of how humans and AI work together. That's one example, and the other place is the support operations job, white-collar work all over the place is blurring in this way now.
It's the same way for humans, teammates who respond to your customers, AI is doing increasingly more of that work. AI is increasingly doing more of the work that your management team was doing too, the same way across all of our jobs. But again, in a more collaborative fashion. I think that collaboration is the new unlock for us, to think through how to build human confidence on those loops, like okay, now I'm confident that we can automate this, or how much do we need to keep building up that confidence loop. That's taken us a while to work through and figure out how to build into the system. But it's a really interesting dynamic that's shifted, it's just way more genuinely collaborative, and I think this is where you even think of how people are framing how you should use AI today, as a thought partner. I think that's the right mental model, because the AI will sharpen your thinking, but it still needs to be your thinking, a lot of it needs to be your thinking. So it's not a tool, but a thought partner, that's a shift, and I think that's part of that collaboration. It's a juicy and fun space.
Sean Nolan: Yeah, absolutely. I think it's really interesting what you mentioned there about the handover from an AI agent to a human agent. As you say, sometimes it doesn't need to be, you have to read this from scratch and tell me everything that's happened, you have to diagnose the situation from scratch. Actually, as you say, it can be as simple as, can you have a quick look over this and give me your approval. That actually lets the human agent know what to do and helps them save time, and makes their interaction with the agent a lot more straightforward. So I can definitely see the value in that kind of change.
Brian Donohue: Yeah, exactly. Or it may just be, there's still a challenge of exposing data, there's so much data locked up in systems that there's no API for, or it can't get built to expose into the system. So then the agent can also say, hey, I don't have this, can you get this data, and then I'm able to do the rest of this work. So reducing it down to exactly what you're saying, we can shrink the amount of effort the human needs to spend on this, to just be more dialed into it rather than requiring all the context and getting up to speed for what turned out to be one minute of work. So it's unquestionably way more efficient for that human spend of time.
Sean Nolan: Absolutely, yeah, absolutely spot on, thank you.
Chapter 4: Advice for CX Leaders Deploying AI Agents
Sean Nolan: I guess I wanted to ask next, based on your experience working with Fin customers, what's one key piece of advice you have for CX leaders who are looking to deploy AI agents now for their customer experience?
Brian Donohue: I think there's two thoughts on this. The first one, where I'd anchor initially, is having genuine ambition, and recognizing how in most places generally there's a massive technology overhang, from the capability of the technology to the actual value we're getting out of it. That capability is there, and I think we still see where teams can get to good success fast, but do they have the ambition to get to 80-90% of all their volume across all their channels? You know, customers might say, maybe not email, it just doesn't feel right on email, I'm just not quite comfortable on email. Sometimes there's genuine capability constraints, but often there's not, and folks really haven't pushed to find where that boundary is. So I think you need sufficient ambition to not have a massive technology overhang where you're underutilizing what can be done today, because the state of technology today is jaw-dropping. We take a lot of it for granted, but it is jaw-dropping what we all see every day. So I think that's number one, and that'd be the single most important point from my point of view.
The second one is, hey, we're going for it, we're going to swing big, is recognizing that you need to view this, you now need to operate like a software engineer. You have software as your product that's delivering your customer experience, and you need to apply the same rigor to building and iterating. You need to think, I'm shipping this AI agent to my customers, this genuinely is software, and I need to be thinking like engineers do. You don't need to be able to write the code, but this means doing things like regression tests, I need to ensure I'm not adding functionality that's accidentally making other parts break. So you need a robust monitoring system in place. Teams don't have this, they have QA for humans in place, but that's different, that's a different setup than a robust real-time live monitoring of the quality of what your agent is delivering. And as you climb that hill to get to that impact, every step needs to be increasing your overall efficiency and CX score, not accidentally degrading it. So you need regression tests and safe testing environments to actually do this, all the things engineers need. That's what we've been leaning into, because to match that ambition, you need the rigor and capability of the system to essentially ship this critical software and ensure it's doing what you need it to do with confidence. So CX teams need to think and act like software engineers, but they don't need to write the code.
Sean Nolan: Yeah, no, absolutely. I think it's really refreshing to hear that you're encouraging that ambition from CX teams and businesses, but obviously with that strong foundation underneath, and I think those two elements working together is a recipe for success. So great, thank you.
Brian Donohue: If I can just jump in, there's one other thought that I think goes to the ambition, but it's recognizing it's not a project, it's an ongoing investment to climb that hill to get to that impact, not thinking, hey, we can spend whatever that initial timeframe is and then we're good, and then we're in the new way of working. No, it's your new spend of time, continually improving and optimizing the system, and you need folks who are durably on this to do that hill climbing. So that's another place, that's probably my third, you said one, I took three, so I was a little greedy there.
Sean Nolan: No, I think that all sounds like good advice, I appreciate it, thank you.
Chapter 5: Predictions: Trust and the Future of AI in Customer Service
Sean Nolan: For my final question, I'm curious to know, looking ahead, do you have any predictions for how AI agents will continue to redefine customer service experiences? And maybe you could touch on trust as well, a topic we haven't really talked about much yet. Where do you see those pieces fitting in?
Brian Donohue: Yeah, there's two interesting angles on that. First one, we actually just released some public research on the state of trust in AI, and it's still mixed, particularly in the context of customer service. What's interesting is it's shifted, but where there's still, where trust is lower, it's channel specific. On chat, people are more comfortable with AI, they're more likely to trust the answer, compared to voice, where there's just stronger distrust of that. I think it's hard to know the why, it's speculative, but because they're using AI in most chat contexts, that feels grounded, like okay, this is where I expect AI to be, and folks are less accustomed to it in the voice setup, and maybe more burned by crappy voice experiences of older technology. So trust is still a challenge where on voice, people want to escalate immediately because they think they're going to get a crappy IVR. And we still hear from our customers a lot, a big challenge they have is Fin gave the correct answer, but the customer didn't trust it, and so they're just having to say, yeah, Fin was right, let me know if I can help with anything else. People have macros built for, "yes, Fin was right." So this is a hard one to solve, we've tried a bunch of detailed design to navigate that. But I do think this is one where at a macro level it just evolves as people adapt to new technology more broadly.
Another interesting angle on this is, well, when does customer service move to agent-to-agent, right? When are people actually using a consumer agent that's doing the navigating? Then you've got a whole different trust category there, they're not going to be distrustful of AI. So either the general public, as much as you can meaningfully say that, has higher levels of trust based on good experiences when enough of those are out there, or as more of it shifts into an agent-to-agent setup, that trust kind of dissolves away, because that's no longer the blocker anymore. So it's interesting to see how this will evolve, and I think time frames are hard to assess on these, because sometimes this stuff moves way faster than you expect, and other times way slower, it's usually that kind of slow-then-fast. But we'll see how it continues to shift.
Sean Nolan: No, absolutely, I think it's really interesting research you've done there, and as you say, maybe on a macro level trust will change, and maybe the channels that people trust AI agents with will change, and I think that will have a knock-on effect for a lot of CX teams, and there's always that process of constant evolution that comes with that. But yeah, thank you, really fascinating insights, and thank you so much for your time today, I really appreciate it.
Brian Donohue: All right, thanks for having me, Sean.
Sean Nolan: And thank you for watching, if you want more content like this, go and create an account on the CX Today website so you never miss another insightful conversation with a CX leader.
