AI Handoffs Are Breaking Trust, Concentrix Warns

AI agents are handling more customer conversations than ever, but what happens when automation reaches its limit?
In this CX Today interview, Nicole Willing speaks with Craig Cotton, Global VP of Product Management at Concentrix, about one of the most important moments in the modern customer journey, the handoff from AI to a human advisor.
Cotton explains why brands risk damaging customer trust when customers have to repeat information after a failed bot interaction, and why even a small failure rate can create millions of poor experiences at scale. Cotton also explores the operational gaps behind broken handoffs, from telephony and CCaaS integrations to missing AI summaries, poor queue visibility, and a lack of ongoing quality assurance.
The discussion covers what a well-executed handoff should look like, why AI agents need the same onboarding, coaching, QA and brand training as human advisors, and how real-time sentiment analysis can help companies recover a negative customer experience before it becomes a loyalty problem.

Transcript

SPEAKER_00 0:09

Welcome to CX today. AI is now handling millions of customer interactions every day, but many brands are discovering that automation alone doesn't guarantee a better customer experience. One issue that keeps coming up is what happens when AI reaches its limit and a human agent needs to step in. Customers often spend several minutes explaining their issue to a bot and need to get transferred and repeat everything from the beginning. So to discuss that, I'm joined by Craig Cotton, Global VP of Product Management at Concentric. We're going to talk about why this battery handoff moment is becoming one of the biggest trust failures and what companies need to do operatingly to fix it. Welcome, Craig. Thanks for joining me.

SPEAKER_01 0:46

Thanks, Nicole. Happy to be here.

SPEAKER_00 0:48

Okay, so to kind of start, you know, a lot of the AI conversation still focuses on containment rates or automation success. But why do you think the handoff moment is getting less attention than it deserves?

SPEAKER_01 1:01

I think there's intense pressure to reduce costs, take some cost out of the business, increase automation. And I think the focus is there, and a little bit less of the focus of when these conversations increasingly get hired uh handled by AI or by automation, that uh we're going to have an increasing number of customers that are unhappy with those interactions.

SPEAKER_00 1:26

Is it the AI itself or it's the kind of design around the transition point that's the issue?

SPEAKER_01 1:33

Well, I I think it's actually just a numbers thing. The way we've been thinking about it at Concentrics is as the number of uh conversations that get automated using AI tools with voice bots and chatbots and other various um automation methods, as that goes from the thousands to the millions to the tens of millions to the hundreds of millions, and and and quickly, if not already, to the billions of customer interactions. Just pick a very small number. What if one percent, one out of a hundred of those interactions was not a successful interaction? The clients that we meet with think it's going to be probably higher than one percent. But uh, you can almost picture in your head the graph at 1% of millions to tens of millions to billions of unhappy customers. And we at concentrics, we think about that as a moment that matters. You've got a client, their sentiment is low. These systems are capable of detecting um sentiment analysis in real time and taking actions based on that. So we we can talk about that as a part of the solution. But once we detect that and the customer's not getting what they need, well, we've all been there with the older systems that weren't as intelligent. Um, human, can I speak to a representative, agent, advisor? Like, what's the word that's gonna get me to somebody live? And um, that even if that stays fixed at 1%, that's gonna be millions of these interactions. And if the advisor that then takes that that conversation says, Hi, this is Craig, what can I do for you? You could have a customer that just spent five, six, seven minutes with an AI bot trying to to uh get their problem handled. And when they're unable to, and then they have to start their journey over again, there's nothing more frustrating. We've all done it, we've all experienced that. Like, what do you mean you don't have my contact information? The the worst is when you get asked for your account number or a phone number or something, or the in uh in the US here it's the last four digits maybe of your social security number to identify you to the bot to authenticate, and then the human asks you the first thing over again. It's quite frustrating.

SPEAKER_00 3:48

Yeah, definitely. Um, so what are the the biggest operational gaps then that are causing these broken handoffs?

SPEAKER_01 3:55

So the first thing is the AI, of course, is uh the AI systems are very capable of very quickly, almost in near real time, summarizing those conversations for a proper handoff. But then you've got some technology in the way, you've got uh telephony links and SIP SIP connections, um, you've got the the CCAS platforms that the advisors are using, and there's, as we all know, five or six major platforms out there. Um, there's the sort of the new AI uh agent tools like our iX Hero, uh, which sits on the desktop of the advisors and works with all of the different CCAS platforms. So there's a lot of different ways to then connect that call from the distressed customer back to a human advisor. And the key thing there is just providing that advisor uh an AI summary of the conversation. Here are the three or four things that the customer has had an issue with. And then um, here is the first thing that you should uh ask them or say to them, reassure them, hey, I it looks like you've just spent five and a half minutes with Amy, our AI assistant, and it looks like you did A, B, and C, but then you got a little bit stuck on this topic here, and I'm here to help.

SPEAKER_00 5:14

Yeah, that makes sense. So then um you touched on it there, but what does a well executed handoff look like in practice?

SPEAKER_01 5:22

In practice, it's um recognizing all of the uh anything at all you you would want to prompt and configure the AI bot first to um uh the first thing would be when you automatically detect the low sentiment to suggest, hey, I I I see that uh you're not getting what you need out of this interaction. Would you like me to get you to a human advisor, human agent uh to help you? Would be the first thing. Um, the other nuance of that is you want to have a view into the queue lengths. Yeah. Um, if there's a 10-minute queue and you've got an unhappy customer making them wait 10 minutes, so there's a situation where you'd like to say, um, you know, I'm not able to uh to help you, but I do have extended wait times. Can we call you back in approximately 10 minutes? Uh would be a better way to handle it versus throwing them into a queue. And let's say there's somebody available, then say, Hey, I have an advisor available right now. You don't, you're not gonna have to wait at all. Um, can I send you and I'll provide them an AI summary so you do not have to start your journey over. So give them the expectation right from the beginning that this is gonna be a better or different experience. You'll raise sentiment just a little bit immediately by doing that. And then you've got to hand it off, and we've got to then determine what is the system or the piece of tech. And again, there's various ones, there's different choices available depending on the the client and what their needs and their systems are, but um, many different ways. The good news is not there's just one to provide this AI summary, and that's got to be short enough that they can uh basically, in the first maybe 10 to 15 seconds of the call, kind of read that out like here is the summary that I got from AI, and and I'm here to help you. And it looks like we need to now uh uh do this. Do I have that right?

SPEAKER_00 7:13

Yeah, sure. That sounds like a sensible way to handle it. And you mentioned obviously the scaling of it's 1% if you've got millions of customers. Um, are companies kind of underestimating how complex that orchestration can become when you're dealing with that large number?

SPEAKER_01 7:28

Uh I don't I don't know if they're thinking too much about the orchestration, like we just talked about of how do I actually mechanically handle it? Um I don't I don't see a lot of clients spending too much time there. Uh for I think several years they've been thinking about well, automation is going to handle the easy stuff, and then the the complex interactions will still be handled by humans. And I think there's some truth to that, but honestly, AI is quite capable of handling complex transactions as well. So, like anything in life, um well, let's just go to human advisors. You've got um very new ones, maybe maybe that don't have proper training or the proper skills or a good enough QA helping them out. Um, and then you have to uh kind of do the things to upskill them and train them and do QA in them and give them coaching. AI bots are the same, like they're gonna be really well-functioning ones uh that just handle these things, including complex ones effectively. And then there are gonna be other bots, uh chat bots, voice bots, SMS bots, uh social interaction bots. There are gonna be ones that um uh aren't built well and are gonna need coaching, they're gonna need training, they're gonna need upskilling along the way. So we we um over the last really 12 to 18 months, we've really thought about the technology that a human advisor needs is we think identical to the technology that an AI uh bot would need. And we think the journey of onboarding and learning and training and QA and coaching and upskilling and all of that, uh we've got that journey that we've mapped out for our advisors. Uh, we call them game changers at Concentrics. Um, we've had that mapped out for years, and we've got we've got a large team that does all of these things. We copied and pasted that exact journey and said, well, how do we adapt it for an AI advisor? And there's very few changes. It's uh essentially the same kind of journey that's required. And uh one last example, um, an AI bot that's working great, and you don't have a um a lot of unhappy uh customers, um, that that needs QA as well. You want to be monitoring those, and then that's gonna need updating and training and and and things like that along the way as well. So it's a mistake to think that it's you're gonna build it and then it's gonna handle your customers' issues for the next six, twelve, eighteen months. It's gonna need some constant kind of care and feeding and modification as well.

SPEAKER_00 10:04

Yeah. Well, that makes a lot of sense because as you say, I think sometimes um you know companies assume they can just install the AI agent. Like you say, it needs uh the initial kind of training, but also maybe some ongoing adjustment as well.

SPEAKER_01 10:19

Yes, for sure.

SPEAKER_00 10:21

So when it comes to um customer trust, you know, um, which is you know such an important word in this space, um, why does this particular moment of handoff carry so much weight for customer trust?

SPEAKER_01 10:33

Uh well, I think it's back to what we talked about. Um I'm in distress, I'm frustrated, my sentiment is low. Otherwise, uh, you know, most of these conversations will uh very successfully be handled by the AI technologies moving forward. Um but when it's not, that's when you know, brands and and our clients need to really take a step back and say, okay, how do we get this from a negative experience and turn it very quickly into a positive experience? And the ability to do that might be the difference of retaining or losing a customer. So um we all know that sometimes your best customers come out of negative situations where you you handle it quite well and you recover and you treat them well and you move on. Um, but if you go from you know bad to bad, hopefully not bad to worse, but uh what we think starting over that journey, how can I help you? Who are you? Let's get you re-authenticated, uh, isn't isn't the way to go. So making sure that those handoffs take place. There's different layers as well. It's not a binary zero or one type of thing. Um, the the first is knowing the identity of who's coming through, and there's uh different ways we've got technology that we're investing in uh to do that uh in with increasing accuracy. Um, but knowing who it is, uh having some information about the customer on the other side in real time as you get it, and then um the other side of it, that's the it's kind of 50% the um uh the identity and and and what's going on there, and then 50% what has your journey been? What's your experience been over the last two or three or five or seven minutes? How can I understand what that's been like? And then the the follow-up is always searching for this next best action. Here's what I need to do next to make sure that I can take care of you and to get a negative experience uh turned back on uh back on the right track. And when they're talking to the live advisor, we want to be um in real time, still getting the sentiment analysis. And you could take different actions if it's not improving. Um, right? You need you might send a message to say, hey, I um uh uh Nicole, I I I see that doesn't seem like I'm able to really uh take care of you today. What am I missing, or how can I help you? So you that you can prompt the advisor to say, maybe I could get you a supervisor that's got more experience in the space that could help you uh more quickly, or something like that. So you you want to follow that through throughout the entire journey.

SPEAKER_00 13:08

And do you think this becomes a competitive differentiator over the next few years, you know, with everybody introducing various forms of AI agents, is getting that part right gonna bring them a competitive advantage?

SPEAKER_01 13:21

Uh I I think for some, it will for sure be a competitive advantage for the ones that do it right and and do it uh sort of best in class. For some, um it will be uh a disadvantage and they'll lose clients if they roll out the automation technology, they do it too quickly. We we talk to clients a lot. The technology is such that building these automated bots, we could do it very quickly. We could do it in in in minutes to hours these days. Uh, but then it takes um usually weeks of testing QA on these systems, making sure the answers are in the right brand voice and with the right terminology, how you would speak to your clients, and just making sure that it's bulletproof. So it's um trivially simple to build most of these things these days. The technology is advanced very quickly. Uh, but then we like to caution our clients that we should spend plenty of time in our program plans and our cycle here for testing and making sure that you've got confidence in it. And then the the last part of it is okay, we've got our date for launch. What's our process or mechanism going to be to make sure that we're monitoring this on an ongoing basis? Either the client is or or uh we've got a large, of course, at Concentrics, a services arm that can do that for them if they they'd like us to do that. And then just like we QA human advisors, uh, we want to be QAing these bot conversations. And AI has allowed us to move from um most clients, if you look back, you know, five years or 10 years, I don't know the average percentage of conversations that get QA'd, uh, less than 10% for sure, if not low single digits. We now have the technology using AI to QA 100% of the human interactions, and we certainly have the ability to AI, uh QA um 100% of the AI conversations. And interestingly, um uh we I shouldn't say all, but certainly a lot of clients they don't think about the QA on the AI conversations. And we say you should think about it identically to the AI and the human conversations. And um do you want to do it on 100%? And now all of the everything underlying AI uses these tokens, right? That people are becoming familiar with. So AIing 100% of your bot calls uh will cost more than 50% of your bot calls. So there will be some cost trade-offs there. But um the costs are coming down such that most of our clients are are thinking about uh 100% of it, and then you still would then have uh human um QA managers that would be looking at the exceptions and and dashboards from the AI bots, just like they do today for human uh QA exceptions when they need to go in and provide coaching, and then we have to go back and coach our bot to do better. So which we think is gonna be if you get it right, it'd be something that maybe once a month or once a quarter you might do a few tweaks on. Um, but we don't think that these will be static systems, and we think that's a mistake to think about them that way.

SPEAKER_00 16:32

Okay, yeah, that's interesting. Um, because kind of thinking about that that practical step that companies need to take, um, what capabilities do brands need to prioritize if they want to get it right here?

SPEAKER_01 16:46

Uh we've talked about a few of them. I think a big one is making sure that it sounds like you, sounds like your company and your brand. Uh so we've got ways to um we we can upload vocabulary lists and things like that. So, for example, um uh different terminology meaning different things in different industries. The word copilot in the airlines industry means things differently than in the AI world today, where AI is a copilot for a lot of things, would be an example of getting that right. And um, how how serious or informal of a tone versus how informal and fun. And there's no right answer for that. It's every company and every brand's got a different style around that. You just want to make sure that these systems are speaking to your customers in a similar style as you would speak to them and using the words that you would speak to them. And that's just uh the tweaking and training of these AI systems. So getting that right is important, and then um I I would argue that probably zero percent of these AI systems when they launch, you you could ever be guaranteed that 100% of questions that come through from your customers will get handled adequately. But how do you make sure it you know what the other thing is humans actually can't answer 100% effectively either? So you don't want to set the bar so high for the automation that it's significantly higher than the humans. You just want to move forward with um when you're confident enough that you're getting uh 80%, 90%, 95%, what's what's the right number for you? And there's no magic number there uh as we talk to different customers, they tend to have a um they tend to think about this has to be 100% perfect. We suggest to them that if you've got a hundred or even a thousand or ten thousand uh human advisors supporting your your company or your brand today, that they're not going to be a hundred percent effective. And you shouldn't uh shouldn't let perfection be the enemy of good here as well. Like what's good enough for your voice bots? And then just like the humans, you monitor it, you tweak it, you coach it, um, you onboard new skills and things like that as you go.

SPEAKER_00 18:56

Sure. Yeah, that's that's really important advice. Uh so thank you, Craig, for taking the time to share your insights.

SPEAKER_01 19:03

Yeah, it's my pleasure.

SPEAKER_00 19:05

And thank you to our viewers for watching. For more interviews and articles, head over to cxtoday.com and join the to join the conversation, engage with our community on LinkedIn. Thanks for watching.