Five9 Explains How Human-in-the-Loop AI Drives Better CX Outcomes

In this CX Today interview, Nicole Willing speaks with Jonathan Rosenberg, Chief Technology Officer at Five9, about the role of human-in-the-loop design in AI-powered customer experience.

Rosenberg explains why mature AI strategies should not treat human intervention as a failure or last resort. Instead, humans play a critical role in continuous learning, governance, high-risk decisions, and customer interactions where emotion, complexity, or value requires a human touch. The conversation explores why businesses should avoid chasing 100 percent automation, how AI can still support agents when interactions are handed over, and why CX leaders need stronger auditing, monitoring, and performance evaluation as AI agents become more common.

Rosenberg also shares practical advice for organizations beginning their AI journey: start with simple, high-confidence use cases, prove value, build oversight structures, and then expand carefully.

Transcript

SPEAKER_01 0:08

Hello and welcome to CX Today. I'm Nicole Willing. One of the biggest challenges in AI-powered customer experience is knowing where containment automation should stop and where humans still matter most. Many enterprises are racing to automate customer interactions, but the real question is whether efficiency alone creates better outcomes. And where does human in the loop fit? So to discuss that, joining me is Jonathan Rosenberg, who is Chief Technology Officer at Five9. Thanks for being here, Jonathan.

SPEAKER_00 0:35

My pleasure, Nicole. Glad to be here.

SPEAKER_01 0:37

Great. So to start, you know, a lot of enterprises are still treating human intervention as you know like a fallback or an escalation path. But in a mature production environment, you know, what should Human in the Loop actually look like?

SPEAKER_00 0:49

Yeah, yeah, great question. So I think let me put go on the record of saying this. Humans matter. There is absolutely no way to build a customer experience platform that has zero people involved. That doesn't make any sense. So the question is, where do you need humans? What are the right points in the workflow for human touch? And there's and any AI agent system requires humans at some aspect of the system and through its continuous operation. So let me point out one, I'm going to start with one Nicole that's like really not obvious to people. And this is perhaps one of the most crucial ones, which is what we call continuous learning. And anyone who's ever run a connect center will tell you like things change all the time. New issues come up, new topics of conversation, there's a new product, there's a new service, there's a global world event or what have you. Uh, there's some new situation with the business. The business is constantly changing. How is it that you can make sure your automations, your AI agents, their conversations are able to keep up with that? The answer is continuous learning. And what you need is you always need to have, even if you can automate everything, you still need some amount of calls that are handling these new situations and exceptional cases to fall through to humans to use the data collected from this human experiences, this humantic data, as we call it, to continuously train and improve your AI agents. In fact, you want some percentage of your calls to sort of always go to humans anyway, just to benchmark and get a side-by-side comparison to test the relative performance of your AI agents relative to your human agents. So that's not about fallback or the AI fail, Nicole. That's just like a good way to constantly run your business to make sure your AI agents are doing well and keep doing well. So continuous learning, right? A second area that I'd call out is obviously cases where, all right, the user has a high value situation. This is a high network customer, this is an important customer, this someone who's explicitly asked to speak to a human being. There are there are lots of cases where it's actually in the best interest of your company to connect this person to a real human. Like a lot of people running sales through their connect centers will tell you listen, there's a bunch of cases where I want a human touch to go close a deal. Maybe not for that initial uh outreach for a VDR, I don't need that. But you know, when someone connects back, if they're a high prospect lead, a good value customer, I want them talking to a person. So you want those things to flow through. So there's important cases to the business where you want it to go there. And then a third one I'll give an example with is even when the call or the chat or whatever does go to a human, the job of the AI is not done. And I think it's a mistake in the industry. Many people think about this as a all or nothing. Like either the AI handles it, woohoo, contained, or it goes to the human, fail, 0% credit. It's a continuum. Because when these conversations do go to a human, we're still using the AI technology to improve the efficiency of the way the human handles the task by providing them guidance and information and transcripts and summaries that they're still more efficient. So it's really just a question, again, of like what percentage of your is the required amount of human time in these conversations in order to achieve a desired outcome. So th thanks for letting me rant about that. I feel really passionately about this, that humans have an important role, and it's not just the AI messed up.

SPEAKER_01 4:06

No, absolutely. And you mentioned there it's a continuum, and you know, I think one of the things that keeps coming up in this discussion is you know, deciding when AI should step in, and it might be, I guess, at different points depending on the process. So, how do you determine that moment when the automation you know should hand off to a human agent when it's interaction?

SPEAKER_00 4:25

That's that's a great question. And again, I think this is another misunderstood thing where a lot of people are like, oh, the only time is if someone says human agent, right? Obviously, that's an easy one. But uh, I'll give you some other cases where we actually traditionally set these things up to roll through. One is like when we're seeing that things aren't going well, like things are on a loop, like the AI agent is consistently misunderstanding or misspelling a person's name, right? We've had this many cases, especially in connecting, for example, PAI. I mean, a simple case like this, like even human beings sometimes will struggle to understand the spelling and and the pronoun of names and what have you. And if you if you're in a loop, you know, you actually want to fall through. So those are cases where the AI is struggling and you want to hand it off, right? There's a sentiment dimension to it too. Where in some cases, if the user's visibly angry, they're they're saying nasty words, you know, uh, you probably want to say, okay, great, I'll I'll hand you off to a human. Um, and there's a lot of cases that are actually business specific. So we find, for example, certain businesses will say, listen, I don't want to have loan approvals. That use case, I want that to get handed off to my humans. So there's like very business-specific use cases where they they desire those kind of things to get handed off. And that's in addition to all the use cases I've just described, where you just want a percentage to just be flowing through to humans anyway, for A, B, and performance comparison. Even that aside, there are these other cases where you still want humans in the next.

SPEAKER_01 5:54

Yeah, exactly. Because, you know, like you mentioned, there are some interactions where you probably don't want the automation in there at all. So what are the interaction types that you would say should never be fully automated? And then how would you define maybe those high-risk interactions?

SPEAKER_00 6:10

Yeah, yeah, again, this is like very, very industry and business specific, which is why, you know, it again, there's another aspect where's the role of the people? Well, the people are involved in the call in like setting these things up, configuring them, understanding the role of the business and keeping them up to date. That's another role for humans, it never goes away. And in a and for example, in the financial services industry, you know, complex financial transactions. So loan approval is is a clock example where like right now, you know, you literally have to be like a certified loan professional to issue uh a loan uh rate to a person. So that requires that literally requires a human. Now the AI can do some pre-work, it can do data collection, it can do discuss the different options. Uh but when it comes time to actually doing loan issuance, you know, you have to hand it off to a certified loan professional. Um, you know, other things like uh you know, healthcare is a great example too, where when something is gets into recommending treatment, uh, you know, you probably want to have a person involved, or where there's actual health and safety risk, you know, um, where life uh is on the line, like obviously, I mean come on, you know, put these things through to people, uh, highly emotional issues, right? Uh users are angry, someone's someone's crying on the phone, like, you know, I think you probably want to put them through to a person, right? So there are these all these kind of use cases where you you wanna you wanna have a human involved in those situations.

SPEAKER_01 7:38

Yeah, absolutely. Um, and then you know, alongside the the kind of operational performance side of it, the governance and accountability are also becoming you know priorities here. So, you know, how should um CX and compliance teams think about auditing AI decisions? And then when you think about you know regulations like the EU AI Act, what are the practical implications there?

SPEAKER_00 7:59

Yeah, so that's a great question, too. And this in many ways is another role for people, right? People have to oversee these AIs as they're performing. You can't just like set and forget, like, oh, I guess my AI is saying something like consumer customers, I don't know what it is. Really? So, what you want to do is uh great products in this space provide lots of data. Um, and so that that includes the traditional boring containment reports and call volumes, and and those are metrics you have to look at if you see big changes in transfer rates or hang-ups or durations of calls, these are signals for a problem. But what we're seeing is there's a whole new category of um auditability that's that's arising is this idea of evals. The idea with an eval is you use an LLM as a judge of how the performance of the AI agent did, and you can actually get really specific questions. Like you can ask questions, for example, like, you know, did the AI properly do its job in only answering questions related to these five topics? Or, you know, uh, did the AI successfully answer all the questions a customer asked? Like, or if the human asked to be transferred to a human, did the call actually transfer to a human? So in natural language, you can actually describe these criteria and you'll get an analytic output that you can see graphs and analytics on. So you need a team of people who are constantly monitoring these eval results, who are crafting these new things as time progresses to track what's important to your business. And you also want to go and you want to sample. You want a good platform will show you'll be able to click all the way down to an individual conversation and see what the consumer said, what the user said, what the agent said, what the user said, what the agent call, what were the tool calls and APIs that were made, go do a little bit of sampling every once in a while, and especially on problematic ones, uh, and assess the situation. So there's this sort of oversight supervisory role. Again, that's a human role. Um, that's a that's a key part of it. And so the the auditability of the platform is its ability uh to do that. And and that's really in addition to uh the sort of regulatory. And I think one of the mistakes people make is they think about like, oh, the EU Act, it's like a tax. Like I have to, I guess I have to have governance and audit and human oversight. Like, what do you mean you have like the government is making you have it? You you want to you're running a business here. You need to understand how this thing is performing for your users. Like, if you don't care about auditability, if you don't care about oversight, if you don't care about governance, like you're not doing your job. Exactly. Forget what the EU act says. So I think these are are really critical components of these systems. And in many ways, I think we're seeing the over time, the capability of these platforms is going to be increasingly complicated and focused on the design and construction and training of these models, and then the auditability, the oversight, the management of these models, and less than the actual running of the phone call itself. The hard parts are those two bookends where again the human touch is required to make sure both of those things go well.

SPEAKER_01 10:56

Absolutely. And I think you hear it there, you know, for many CX leaders, the challenge is figuring out how all of that falls into place without disrupting their existing operations. Um, so you know, if you were advising a CX leader who's looking at this today, um what's the sensible phased approach that you would advise them, you know, to move from maybe like an automation-led approach to this outcome-led AI without disrupting their operations?

SPEAKER_00 11:23

So, what I would say on this, and it's the same advice we've been giving for a long time, Nicole, on this, which is start simple. Okay? Don't a lot of people read, you know, you go on Twitter and you read like, I fired all my humans, I have a 100x improvement in performance, and you feel like you get this FOMO. Right? We see a lot of this FOMO. Oh my gosh, if I don't automate everything, if I'm not at a you know, at 100%, I'm missing out, stop. Okay, it's not stop. Okay, don't run your business by what you read on Twitter. Um the the formula for success is to start simple and expand. So what we see is take a bunch of high, you know, simple, high confidence, easy intents that are high cost to the business, easy, low complexity kind of use cases, and automate them. And my favorite example, still in 2026, is password reset. It is it is mind-numbing, mind-numbing the volume of calls to and chats to conic centers that are still on such mundane topics. And we still have a long way to go in automating those. So if you're not there yet, why don't you start with that? Right, make that successful, and then expand to simple use cases like in financial services, like um, you know, account balance lookup or simple transfers, these things that are not complicated, and grow and expand using this combination of you know, volume of the business, risk to customer experience, and brand damage if things get messed up, and lower complexity, and then just continuously expand the bubble. So that's the general process for it. And as you do that, build these structures for human-in-the-loop oversight that I've been talking about. The the upfront design and creation based on using the experiences of people uh who are handling these calls study to train your agent and adding in the governance and oversight to make sure you can audit and manage it. And once that's in place, just continually expand. That is the that's the that's the formula for success. And then as you go, expand your metric capabilities too. So, again, these new AI agent platforms are really good at this with this new agency LM as judge. You can get better and better metrics on performance as the as the system itself grows, uh, and and you can start to shift from over time simple metrics like containment, which is really what you care about for something like password reset, to more as you build more complicated flows, your evaluations can be more outcome-centric.

SPEAKER_01 13:48

Sure. Absolutely. Well, I think those are really important points for our audience to consider. Uh, thank you for your insights, Jonathan.

SPEAKER_00 13:56

My pleasure. My favorite topic here, Nicole, to talk about as you can talk.

SPEAKER_01 13:59

I know we could go on all afternoon talking about this.

SPEAKER_00 14:03

I apologize.

SPEAKER_01 14:04

But it it's clear that the future of AI and customer experience is about designing systems that know when human judgment matters the most. So for our audience, um, you know, thank you for joining us. And to learn more about Five9, you can go to their website, five9.com. And for more interviews and articles on the challenges facing the CX leaders, visit cxtoday.com, subscribe to our newsletter, and join us on LinkedIn. So thanks again, Jonathan.

SPEAKER_00 14:30

My pleasure. Thanks to call.

SPEAKER_01 14:31

And thanks to everyone for watching, and we'll see you next time.