Tom Arbuthnot: Hi, and welcome back to the show. This week, we are talking real world voice agents. Chris at Maersk gives us a great insight into how they're using voice agents, both for inbound calling and also, in some cases, outbound calling. They've assessed various vendors, so he talks us through how to assess the options, how to work with the business, and where voice agents fit between the contact center, IT, and the voice team. Some really great insights and advice for Chris. Many thanks for him jumping on the show. And also many thanks to Ribbon, who are the sponsor of this show. Really appreciate all their support of the community. On with the show. Hey, everybody. Welcome back to the show. Really excited to have this conversation. We're gonna be talking voice AI. There's a lot of talk out there at the moment, but Chris is actually doing some stuff in the real world, both inbound and outbound, which is super interesting. So, Chris, do you wanna just give yourself a bit of an intro and background, please?

Chris Bailey: Yeah, sure thing. Hey, Tom. Yeah, so I'm Chris Bailey. So I work in our Modern Workplace Engineering Team at Maersk, we're a shipping company. And yeah, like you said, we've been exploring, testing, doing POCs, some of which have made it into production with voice AI with different vendors, for the best part of a year now.

Tom Arbuthnot: Yeah. So you really — I remember we had this conversation. I was at one of the user groups, and we were on Teams, so I was like, "Oh, wow, you're doing outbound." Like that's exciting. So before we get into that, give us a little bit of background to what you've done in the past and your role and your areas of responsibility.

Chris Bailey: Yeah, sure. So I come from a voice background, so I've worked both with Cisco and kind of Microsoft Voice and AudioCodes for a long, long old time now, longer than I can remember. So yeah, this is definitely on my first kind of soiree into voice AI. And it's been very different, I would say, to the traditional voice setup of like your auto attendants or your contact center call flows. It's definitely a different animal, but it's been a good learning curve for sure.

Tom Arbuthnot: And how does a voice break down at most? Like are you responsible for like classic Teams and Cisco voice and contact center? Is contact center a different discipline and service area?

Chris Bailey: It's all... Yeah, no, it's all in our area. So we have Cisco Contact Center, and then we also use Microsoft Teams as well for general IPT. So they both fall under me.

Tom Arbuthnot: Awesome. And then the context of that question's quite interesting 'cause that's an area where voice AI seems to be hottest in the contact center.

Chris Bailey: Yeah. Yeah, for sure. And the one thing that's very different about this kind of work is that normally when we go to the business and say, "Hey, we've got this new technology," or, "We need to migrate you to this platform" or whatever, normally it's us in IT going to them. But actually in this instance it was the business, you know, not just customer services but also finance and kind of IT services. They were coming to us and be like, "We want to start using voice AI." So that was quite a nice change that we didn't have to push it on them. They were happy to work with us on it.

Tom Arbuthnot: Yeah. That's fascinating, isn't it? It's always interesting, and I think that's AI across the board when the business is engaged with IT and they're the ones saying, "We hear about this. We wanna do this. We wanna do that." Like, as you say, most of our careers are like, "Come on, it's time to go to the cloud. It's time to move." Yeah. But it's a very different dynamic when the business is saying, "I've heard X, Y, or Z can change the way we work."

Chris Bailey: Yeah, for sure. For sure. And I think it's really key that you do need to find those people in your business that can work with you on this because I can tell you now, it would not work if it was just IT pushing this. Like, it wouldn't work. You need the people who actually answer the calls to be working alongside you on this and telling you what their SOPs are, how they actually handle the calls and what the customers are calling about, and most importantly, what a good customer journey looks like 'cause that's what it's all about at the end of the day.

Tom Arbuthnot: Awesome. That is really exciting. Like, different organizations are handling AI in different ways. So some have got center of excellences that kind of focus on AI. Some have got... It's just gone into each of the business areas. I think voice is pretty unique in terms of its requirements. So I wanna talk through kind of, 'cause you've done some vendor assessment as well, which is really interesting. But when it first came about, who was it, and have you got any feel for how they ended up knowing about it and what they were trying to get as an outcome?

Chris Bailey: Yeah, I think they... I think everyone's aware of AI 'cause it really, you know, blew up—

Tom Arbuthnot: kind of thing. Yeah. You can't not be, can you, really?

Chris Bailey: Yeah, it kind of took us a little bit by surprise, not, you know, as to how quickly... And I think everyone in the industry has found that as well, just how quickly people are adopting it. And yeah, I think just through... And also when one team got wind of it, it kinda spread like wildfire in the organization. Like, people heard demos and they were like, "Oh, we want this in our team." And it just got people talking and got people coming to us. So yeah, I don't know where that first spark came from. I guess it was just someone hearing a demo online or maybe interacting with a different company. But they were soon after us and wanting us to do demos and proof of concepts with them. And like I say, some of them we couldn't realize any benefits from, and those proof of concepts didn't move any further, but certainly quite a few of them have made it into production now.

Tom Arbuthnot: And during that cycle of the different teams and different products, you've assessed quite a lot. Like, what did you find during that process? What solutions have you landed on? What's that look like?

Chris Bailey: So yeah. I think the key thing is, first of all, before you kind of go and look at the technology, really understand what is the use case or the problem area that you wanna focus on for your proof of concept or for your testing. I think that really is the key thing, is understanding what is the problem we're trying to solve.

Tom Arbuthnot: Mm-hmm.

Chris Bailey: And what are the benefits you're trying to realize from that. Because if you just start trying to do it for the sake of it, it becomes quite problematic 'cause obviously it's not free, it does have a cost. It's also—

Tom Arbuthnot: Yeah, but both a project cost and an incremental cost, which is really interesting, is like the incremental costs are not inconsiderable depending on what solution and what model you're using, especially as you scale concurrency.

Chris Bailey: For sure. For sure. There's loads of hidden costs as well. So there's obviously the AI cost, but then if it's obviously gonna be touching the PSTN at some point, maybe you're gonna see an increase of PSTN costs. So you need to factor all that in. And also resourcing effort. For sure this was more effort than I was anticipating when we got involved in it. At one point, kind of earlier on in the year, in the back end of last year, it was taking up so much of my time, which is a nice thing, good thing, but I hadn't properly planned for that, so I was still trying to do my everyday job alongside this—

Tom Arbuthnot: Yeah, all your BOF for the global voice deployment and this—

Chris Bailey: Exactly. Whole new thing. So yeah. Do bear that in mind that it is that upfront implementation and that planning does take a lot of time, more so than a standard kind of traditional IVR call flow setup. And then also it's not just set and forget, you know, like a standard IVR menu. You know, you set it up, you test it, all good, move on to the next one. For this you have to obviously monitor how it's behaving, fine-tune it, and that process really just goes on and on. So yeah, prepare for needing to spend a lot of time on this as well as having to get the cash out as well.

Tom Arbuthnot: And there's some quite different solutions and different architectures as well, isn't there? 'Cause obviously you've got Microsoft and Cisco as incumbents and—

Chris Bailey: Yep.

Tom Arbuthnot: Cisco have got some interesting solutions. Microsoft have just come out with theirs in Frontier, but then you've got different third parties that have line of business solutions or have specialist solutions, but then it becomes, well, how do I route into those and out of those?

Chris Bailey: Yeah, absolutely. So I think when choosing a vendor, obviously it makes sense to look at your incumbent, whoever you're using now 'cause they're gonna be strong in terms of the telephony integration, which is a really massive requirement. But also definitely I'd recommend to look at ones outside of that space as well. Some of the more niche players... What we've found is that they're typically very, very strong on that AI voice piece of it, so the quality and the naturalness of the voice is excellent. But where they may fall down a little bit is the integration with your existing phone system. So things like being able to escalate that call back to a human agent, is it gonna mess up the contact center reporting, the recording, and all of that stuff that the business still care about. So it's really getting the balance between providing something that's gonna be a good customer experience in terms of the conversation, but also not gonna impact your existing telephony setup too much as well. So you've gotta get the balance right.

Tom Arbuthnot: Yeah. And I know we don't wanna get into like specific vendors on this podcast and just like who you use for what, but you've assessed kind of all the different variants of those options. When they needed different call routing, like is it PSTN forward? Is it SIP refer? Did you do that on your SBCs? Did you do that with your carrier? What did that look like?

Chris Bailey: Yeah. So we've seen a real mixture. So some of the ones I would say that are very new, they seem to be edging towards like almost let's do a PSTN transfer. Like, we'll give you a PSTN number from—

Tom Arbuthnot: And that's what I've heard as well from our side. I'm like, that makes me nervous. Yeah. I'm like, sure to prove it, but like, what, are you just gonna have like 50 forwards going on for every contact center?

Chris Bailey: Yeah. Exactly. And it's really interesting 'cause all of the kind of engineers that I speak to who have, like yourself, got a voice background, they hear that and they kind of are like, "Oh," you know. And they're always like in their heads thinking of like 100 reasons why you shouldn't do that. But then the more kind of, I guess like software engineers or the guys that are doing the development work, they're just like, "Yeah, let's do that." So it's hard to have that conversation and be like, "Okay, we can do this, but let's not scale with that design 'cause you are gonna hit problems." And we did try that, and we have had problems with that approach. So yeah, we have tried the PSTN transfer. We've also, you know, SIP trunking, so—

Tom Arbuthnot: And I'm guessing the business kind of missed that as well 'cause that's very voice and very technical. Like they see an amazing demo that has a really good voice model, that has industry knowledge, and they're like, "That's the one. It's amazing." And obviously they can ring it from a cell phone and test it and like, "Yeah, it works on phones." It's like, "Well, yeah, wait a minute. Let me explain why in our hundreds of people contact center that's not gonna scale very well."

Chris Bailey: Exactly. And also geography. So if you've got contact centers EU based, for example, European based, and then you're looking at AI vendor and all their infrastructure is hosted in the US, all of a sudden you're transferring calls across the Atlantic. How's that, you know, it's gonna mess up your latency and also things like GDPR regulations. So if your recordings and your transcripts are supposed to be stored in the EU, all of a sudden they're being stored somewhere else. You've got to factor all that in as well. So it's very important to look at what's under the hood, because yeah, like you say, the business, you know, they won't be thinking about all of that. So it's our job to explain the risks and why to steer away from that.

Tom Arbuthnot: Yeah.

Chris Bailey: But yeah, we have started to see some vendors using a more like an API media streaming integration, which for us is the gold standard. That's where we ideally wanna get to so the call can stay anchored in your contact center, in your phone system. But then the call is then being — the media is being streamed to the third party agent. That is the best integration model for sure, but not all of them offer that.

Tom Arbuthnot: And in that model, are you working with the Telco and their infrastructure streaming the media, or are you bringing it to your local SBC and you're streaming the media?

Chris Bailey: So right now we're bringing it into our contact center, so our Cisco Contact Center. Right now it's then, we're doing a SIP transfer model, so we're actually just like doing a SIP transfer. It's not even a referred transfer at the moment, so it's quite primitive. But yeah, we are working and pushing the vendors that we're working with to start adopting a more API integration approach.

Tom Arbuthnot: And have any of them come up? You mentioned a challenge of kind of end-to-end reporting. Have any of them proposed a good solution for, like, when you're using vendor X for the clever AI stuff and vendor Y for the contact center or the phone system?

Chris Bailey: They tend to just open up like the APIs, the reporting APIs. So they acknowledge, yeah, that we get that it's a problem, and then they open up their API, so you could build out your own internal dashboard.

Tom Arbuthnot: Right. So you bridge the gap.

Chris Bailey: Yeah. You bridge the gap yourself. Yeah. I mean, if you go with... So Cisco do have a concept of, like this bring your own virtual agent. So they acknowledge that not everyone is gonna wanna use those native Cisco virtual agents. So they've brought out this concept of bring your own virtual agent, so it's a supported way of integrating via API with third parties. And the good thing with that is you still get all the reporting visibility in Cisco. So that is a really good approach. But if you do have to go to a more primitive integration model, then really just building your own dashboard is something you're gonna wanna do.

Tom Arbuthnot: That's true. Yeah. So again, project scope, it's worth kind of factoring that in, like we're gonna have to build or buy some kind of solution to make all that piece come together.

Chris Bailey: Exactly. Yeah. And something that makes your life easy when you've gone live with these, you don't wanna be listening back to hundreds of call recordings every day and going through transcripts. So you need to build in some kind of mechanism when you start to scale so you can actually check what is the performance of the virtual agent? Is it dropping calls, for example, which we had early on? Is it failing to transfer calls back to the human? Is it saying the wrong thing? Is the ASR gone haywire? All that kind of stuff. You've gotta check it. And what I say when we try and explain this to the business now, if we get new requirements, we try and frame it that you're not bringing in like a new solution, which, as you are, but think of it as you're bringing in a new hire, right? Think of it that you're bringing in a new person to your team. You've gotta train them. You've gotta give them access to the various systems and the data for them to do their job. You've gotta be clear what their job is and what their job isn't. And also you need to not just set them up and let them loose. You've got to monitor how they're performing as well, just like you would do or hope they would do with a new hire.

Tom Arbuthnot: Yeah. I think that's a really great paradigm, and that shows the level of investment on the business side or the contact center side or the team side of like, yeah, great, you're all excited about this, but you've got to have the buy-in, as you say, to like explain to IT and the vendor what's good, what's bad, what data is, and maintain and train. And not just one time either. Like things are gonna change over time in terms of how the business works and therefore what the voice AI, what's the right answer, what's the wrong answer.

Chris Bailey: Yeah, absolutely. And I think another key thing is be really clear with the different responsibilities. So if you bring someone in from the business to help on this, be clear with what their responsibilities are. Because long term, do you wanna be the person monitoring the performance and constantly training it up? It's better if you can bring someone in from the business or from that contact center team to kind of take that, who helped set it up in the first place, but get them to do an element of the monitoring and giving feedback to us.

Tom Arbuthnot: Yeah, I think that's a really good note for anybody listening in who's on voice teams or AI center of excellence teams. Like there's a risk here that the business is like, "Oh, great. Now you built this. You maintain this, right? Make it better. If it goes wrong, you fix it." Like you've suddenly acquired a new service and a service where you're not the knowledge SME to even make those changes. They should be the knowledge SME.

Chris Bailey: Absolutely. Absolutely. And yeah, 'cause if a contact center team, if they're realizing some benefits from AI, but actually behind the scenes you've got a whole team of IT engineers making it work, then actually it's kind of a bit of a false positive, isn't it? So you do need to—

Tom Arbuthnot: Yeah. They get that for their contact center P&L. They're like, "Oh, we've doubled the calls we're managing with the same amount of agents." And you're like, "Yeah, but there's two people in my voice team maintaining your agents that you aren't accountable for that weren't doing that before."

Chris Bailey: Exactly. Yeah, exactly. That definitely can happen.

Tom Arbuthnot: Nice. And talk to me about the outbound, 'cause that's really exciting.

Chris Bailey: Yeah, outbound. So we had a requirement from our finance collections team. So they had this manual process where they would make outbound calls to customers on a daily basis to chase up outstanding invoices and collect kind of in a promise to pay and the status of that payment. So they came to us late last year and wanted us to do a demo or proof of concept using AI. So we've built out — it uses an outbound campaign dialer. So the finance collections team every day generate an Excel file. They upload it to the campaign dialer, and it makes outbound calls to the customer numbers, and that triggers an AI virtual agent which actually speaks to the person who picks up the phone and then it will have a natural conversation with them. "Are you customer XYZ? Can you give us an update on invoice XYZ?" Has that conversation with them, and then the information or the data that the customer provides then gets written back to a back-end platform that the finance collections team can then update and track what's going on. So we started off with a proof of concept to a demo just in one location earlier in the year, and that's now live in like 50 countries, I think. So that's had a really big uptake, and that's all because our finance collections team were really the ones behind us driving it. They really wanted to go live with it, and it's been a really good success.

Tom Arbuthnot: Yeah, that's a great one where it's super, super measurable, right? Like, we're bringing in collection time, which... And that's a slog of a human job as well. Like, it's mostly gonna be non-answers. Like, you need to try multiple times. So that's really interesting. And multilingual as well.

Chris Bailey: Yes. Multilingual, yeah. Like, different languages in there as well. And what we found was, after we went live in one or two countries, they were like, "Okay, we can see for sure that there is a benefit," and they can measure the benefit, but it was small. Yeah. So it wasn't really — and that's why we've had to scale it up to kind of 50 countries before we actually realized, okay, now they're realizing some benefits from it. But you almost have to really scale it up before you can kind of claim any kind of benefit or saving.

Tom Arbuthnot: Yeah, that's a really important point, and thanks for sharing that because I'm seeing a lot at the moment of proof of concepts and tests, and it's like, well, we've still got all the old costs and we've got the new AI costs and we haven't hit a tipping point. So actually, like, we're on the right trajectory, but if the business doesn't understand the direction of travel and the level of commitment, it's like, well, all we've done is added the AI, not taken away any people effort or people heads and we're net up, whereas suddenly, oh, now we can be 24/7 or now we can be multi-country, that tips it the other way.

Chris Bailey: Yeah, absolutely. And I think, don't be scared to pull the plug if a specific use case isn't working out. Like, we did another outbound use case, not finance, but a different one, and actually, yeah, we kind of pulled the plug on it because we weren't seeing any benefit at a small scale, so therefore you're definitely not gonna see any benefit if you scale it up. So yeah, definitely be choosy with the use cases that you're going for.

Tom Arbuthnot: Nice. And how has it been ongoing assessment of the quality and the experience? You mentioned transcription, you mentioned recording. Like these are non-deterministic systems, so they're not going to do exactly the same thing every time, and that's part of the power— Yep. But also part of the concern potentially.

Chris Bailey: Indeed. And yeah, I think there's two areas. One is that kind of change management becomes quite difficult because if you like wanna update the instructions or you've got a new SOP, you don't know how it's gonna handle that. Yeah. So you have to be careful. If you do make changes, like be really strict with your kind of version control and make sure everyone's aware, right, okay, we're gonna push this change into production. Let's all make some test calls to make sure it's not doing anything completely crazy.

Tom Arbuthnot: And it's not just your changes either. The models behave differently over time. Yep. So even if your business case is static, if they say, "Well, we're moving from model X to model Y," the way it works might well change based on those prompts or that information.

Chris Bailey: Yeah, absolutely. And like you said, those models are being updated quite frequently now, aren't they? So yeah, that is one that's kind of we've noticed as well. And different vendors seem to have a different level of maturity on the observability side. So some are really, really simple. You can access the transcripts and the recordings via an API, and you can almost build out another agent, like a separate agent that monitors your first agent. Hmm. Because if you do start to scale up, you do need another AI model to kind of measure how it's doing. But saying that, some other vendors are still quite poor at that and we're not actually able to access the transcript. So do factor that into your planning as well. Can you get, how do they deal with performance monitoring? Is it something that's already built into their platform, or do you need to go and build something out yourself?

Tom Arbuthnot: Yeah. I feel like this is an area where you can potentially, or the business can potentially get seduced by a really impressive demo that might not even be a phone call, might just be on a website. And it's like, oh, yeah, but to get from there to scale or production, reporting, compliance, API, change control, cost model, like those are things that you've weighed on a lot in this conversation. It feels like you've been very good and/or learnt a lot about assessing the different vendor options beyond just the demo.

Chris Bailey: Yeah, absolutely. And I think some of the testing can be a bit subjective as well, so different people can have different thoughts about how the AI is responding. Now some people think, "Ah, that voice is excellent." You know, you can select different languages, so different accents, can't you, for different ones. And some people are like, "Yeah, that accent's really good," and other people are like, "No, I don't like that one." So it can be a bit subjective and also you can customize things like the speed of the voice to how it sounds. And we've had some people are like, "No, we need to slow it down." And then the next day, some people are like, "No, we need to speed it up." So yeah, you have to tread carefully there.

Tom Arbuthnot: Yeah. And that's back to where it's not IT solution, it's the business' solution and we're a facilitator. Like—

Chris Bailey: Yeah.

Tom Arbuthnot: Whoever is in charge of this project, they can make the call. Like, don't ask me to speed it up one day and down the next. Ask Sarah — she owns the strategy. She owns that.

Chris Bailey: Yeah. For sure. For sure. And that's where you've got to be really clear with the responsibilities as to who's in control of what.

Tom Arbuthnot: Yeah. Awesome. And you've mentioned APIs a couple of times, like your voice and contact center. Now obviously, our world in voice and contact center has gone very, very IT, very cloud. Was this a next level in terms of IT and models and MCPs and reporting, or was it a fairly natural evolution for you in terms of what you're used to with Teams and cloud, and now it's just a little bit more?

Chris Bailey: It's definitely... It's interesting 'cause I would say voice AI wouldn't automatically... I don't think if someone has a traditional voice background or sits, like I do, have a dedicated voice team, I would say it's not 100% guaranteed that that would be a good fit for your team. Because, yes, there is an element that you need to know about telephony for sure, but then there's also an aspect of you need to know about prompt engineering and API integration, which may not, you know, depends obviously on the org, but for us it was a skill set that we had to improve on. You know, obviously your PowerShell scripting and traditional telephony we're strong on, but some of this was definitely new to us. So it's been a great learning area, but we have had to work in good collaboration with other teams outside of our area to get this done because we don't know every single platform that gets used in the company and how we can connect to them. Yeah. So you cannot do it alone, I guess is what I'm trying to say. As well as your telecoms team and as well as having someone in the business, you need to work with software engineering team or whatever they may be called in your organization. So it does require a fair amount of collaboration, and be clear as to who is gonna own AI voice in your org because it's everywhere almost. We're seeing it in ServiceNow, we're seeing it in Salesforce. So all of those CRMs have got their own AI voice. Microsoft have obviously got Copilot. Whoever your contact center provider is, they will have it as well. So if you don't try and control it early, it will start popping up everywhere and everyone's doing their own thing. So that's definitely a risk that you've gotta get a handle on it quite early.

Tom Arbuthnot: Yeah. It feels like, from a personal career development point of view, like a really hot space to be getting into in terms of skills and capabilities. And then you're right, on an org strategy level, like if you put your head in the sand to this, you're gonna turn around and find out the Salesforce team have gone ahead and enabled a bunch of stuff and off they've gone with their AI voice and it hasn't been considered in the wider voice strategy or some of those things that we picked up on in the conversation about weird forwarding or geo regs on telco. They're not gonna necessarily be aware of that stuff. It's a fraught area if you don't understand the background in telephony.

Chris Bailey: Indeed. 100%.

Tom Arbuthnot: Awesome. Well, Chris, thanks for sharing so much. It's a really exciting area. So I've got quite a few organizations I'm talking to who are getting started or doing a little bit, but you've shared so much there on not just one vendor, but how to think about the problem space and then a vendor assessment. If it's all right with you, we'll check in again at some point in the future to hear, it sounds like you're on a trajectory to do more here, so I'm sure there'll be more to talk about.

Chris Bailey: Yeah, absolutely. I think the first part of the year has really been about testing and we've been evaluating and moving on. It's now about scaling up. So yeah, happy to come back in the future.

Tom Arbuthnot: Awesome. Well, thanks so much, mate. Really nice to catch up,

Chris Bailey: And we'll talk again soon. All right. Thanks, Tom.