Not Building Trust Before AI Agents Is a Mistake

Rob Wilkinson sits down with Amazon Connect Customer leaders, Tony Shen and Jeremy Puent, to unpack what slows real deployments down, where things go wrong, and why the biggest risk is not the mistake itself but failing to spot it in time.

They explain why ambiguous documentation creates inconsistent outcomes, why bad experiences drive disloyalty faster than good experiences build loyalty, and what “visibility” really means when an AI agent makes decisions you need to defend. The discussion gets practical on non negotiables like guardrails, transparency, customer consent for critical actions, and clear escalation paths to humans. You will also hear why containment and deflection can undermine CX, and what metrics and operating habits matter more as AI becomes a production channel.

Check out Amazon Connect Customer here.

Transcript

SPEAKER_02 0:08

Hello and welcome. I'm Rob Wilkinson, and today we're taking a closer look at a question that I hear repeatedly from contact centre leaders at the moment right now. It's not whether AI agents actually work. I think that question is starting to be answered, but it's more whether we can actually trust them, well, at least enough to let them operate at scale. That's another thing altogether. So if you're currently being pushed to move faster on your agentec AI efforts, but you might be uneasy maybe about what happens when autonomous systems kind of make the wrong decisions, especially when it's in front of a real customer, then stay with us because you're going to get some practical takeaways that you can use straight away in your business. Because today I'm lucky enough to be joined by Tony Chen and Jeremy Pewin, Principal Solution Architect, both senior leaders at Amazon Connect who basically spend all their time inside these deployments seeing where it's all built, where the trust is, where it breaks down, and why skipping that step is proving some people very costly. So welcome Tony and Jeremy. Thanks both for joining me very much. Looking forward to this.

SPEAKER_01 1:20

Thanks for having us, Rob. It's a pleasure to be here.

SPEAKER_02 1:23

Thanks, Rob. Okay, so let's kind of dive in. Let's add a bit of context though. So over this last year, we've seen quite a clear shift from customers experimenting with AI agents to actually very serious conversations about putting them into production. So we're kind of uh, you know, we're getting moving now. Um what you're seeing on the ground though, what what's actually stopping organizations making that jump?

SPEAKER_01 1:50

Yeah, um I I I'll jump in here, Rob. I I wouldn't say um I'm seeing customers stop or being stopped. It's it's more a slowing down, and it comes from I think something you said a little bit or hinted at in the in the introduction, and it's the lack of confidence, right? They're trying to build that confidence. Um they're they're they're trying to figure out what's right and what's ready and what's not. And I find that more often than not, the the slowdown comes around their documentation. Their documentation isn't clean and clear and ready for AI to use it, to consume it, to um provide the proper guidance, or it doesn't exist, right? It's all in somebody's head and they're trying to write it down. But but often um it really is that the documentation is very ambiguous. And I think that you know, those of us that are in the industry working with this stuff know that ambiguity is the enemy of good here. And so ambiguous documentation, like if two human beings are gonna read it and interpret it differently, the machine's definitely gonna give inconsistency in how it answers based on that documentation. And I think that's the bigger thing slowing everybody down. I I think we've all heard in the past garbage in, garbage out when it comes to documentation, and that's still true, or it's even amplified with AI, because you know, the the machine's gonna give a different answer based on the context of the question and how it interprets the answer in that moment from the documentation.

SPEAKER_02 3:28

Okay, that's interesting. Because kind of we don't always hear about uh people shouting that they've kind of maybe they're not trusting things or they're you know the delaying or stalling on things. Um that's something that happens maybe a little bit on a quieter side of things. Um when when leaders do hesitate though, um it can often be labelled uh as a problem or resistance or maybe even fear. Um in reality, do you do you think that's the case or are they protecting something?

SPEAKER_01 4:00

Um, I mean, I think a little bit of it is a tiny bit of fear, right? Because they're trying to protect their jobs, right? Um nobody wants to be in the news for the wrong reasons, right? They say there's no such thing as bad press, but if you're in the press because you put a bot out that's you know tweeting racist statements or maybe you know selling a vehicle for a car manufacturer for like 18 bucks or 24 bucks or whatever it was, right? Like you don't want to be in the in the press for those kinds of things. But I I think more more importantly and more commonly, what I see is leaders are trying to protect their brand and more importantly, their customer. I'm seeing more and more of a shift in the conversations away from things like containment. I mean, it still comes up all the time, or deflection, and more into the customer experience and trying to make sure that you're not creating experiences that frustrate your customers or worse, create bad experiences. And I I think that's more where the hesitation is coming from for the good leaders, and I'm seeing more and more of that. Does that make sense?

SPEAKER_02 5:11

Absolutely, yeah. I think it's a really valid point. I mean it's it's kind of coming from a good place, I guess, in in in reality, the kind of customer focus. It's important that, right?

SPEAKER_01 5:21

Yeah, I mean, if if you ask in any room that's having a discussion around this, like show of hands, who's had a bad experience with a bot, right? Like every hand's going up, right? We've all had bad experiences, and nobody wants to be responsible for putting one of those bad experiences out.

SPEAKER_02 5:40

So I guess talking of bad experiences, what what how can that turn out? Kind of what does that look like from a customer's perspective? Because um we want to look at this. It's not theory anymore. This is this is kind of the reality of the situation. And as you've mentioned, things can go quite wrong, and there's been some quite public issues over the kind of recent times. So, how does that look and and and in the moments that can damage trust with customers, what does that kind of play out like?

SPEAKER_01 6:08

Yeah, I I think um for years people have said that great customer experiences drive customer loyalty, right? Like I've heard this, I've read it a million times over the you know three plus decades that I've been in this industry. And and I find that the inverse is actually the greater truth. It's not that great experiences drive loyalty, it's that bad experiences drive disloyalty. You can treat me great 99% of the time, but if you treat me bad enough once, I am never gonna be your customer again. And so creating the customer experiences that allow you to be notified and having the tools that notify you when something's going wrong, when things are off the rails, becomes far more important because you need to be able to step in and save that customer experience in that moment. Because once they hang up, it might be too late. They may never reach back out, they might not call back and allow you to create a better experience or give them a VIP treatment. So the reality is that if your AI agent is doing something really wrong, it's giving incorrect medical advice that's gonna lead to harm. Or um, you know, it's telling you, hey, take those two wires out of the back of your machine and touch them together, you know, and you're gonna get shocked, or it's gonna short out the box that you just paid hundreds of dollars for, or whatever it is. If if the AI gives advice that becomes harmful or frustrates you by just talking you in loops without solving your problem and giving you an escape patch and a way to get to a human being who can either fix the problem or find the right person who can, you're creating those scenarios where you risk losing a customer. Um and so if if the AI agent is creating those frustrating or harmful experiences, you're gonna lose customers, and that's what people don't want.

SPEAKER_02 8:17

It's a super valid point, and a key there's a there's a key difference in what you've said in terms of looking at those failures uh through through the correct lenses. I think there's um from the customer's perspective, that's one thing. I think we've also got to consider internally though, because um there's someone's got to pick up the pieces when those things go wrong, right? Um, and that's an impact on our frontline, our human teams, and the supervisors of those teams. Um so they're left managing the consequences of uh you know bad decisions from the AI that they didn't see coming. Is that is that fair to say?

SPEAKER_01 8:55

Yeah, for sure. And it's not new, right? The same thing happened with human agents. Like human agents say the wrong thing, they're having a bad day, they frustrate a customer. Some personalities just clash, you know, and sometimes somebody's not gonna like somebody and they're gonna have a bad experience, and another human agent has to pick up the slack on that. The same thing's true with AI agents, and and that's true across the board. And that's something that I go back to in these conversations with customers is as you're testing or thinking about these experiences, like think about what happens with a human agent in that same scenario, right? If Tony and I answer the call and read the same document and we interpret it differently, um, we're gonna give different guidance. The AI agent is gonna do the same thing. If Tony or I frustrate a customer in some way, the other one of us is gonna have to pick up the pieces on that next call if we're lucky enough to get a second call from that customer.

SPEAKER_02 9:56

It's such a valid point, and and it's such a it's a it's a significant risk, actually, um, of getting it wrong. So um it feels like there's a a bit of a visibility thing there. So people flying almost like flying blind, not knowing what's coming next. Let's talk about the cost of moving too quickly, maybe, um, before you've built the right foundations, because we've spoken to leaders who say that the biggest risk, it's not it's not that AI gets something wrong. Like we know it is a risk, but it's not the biggest one, it's that they won't find out about it. So what why is that so dangerous, that lack of visibility?

SPEAKER_01 10:37

Yeah, I mean, in in the contact center for years, we've said if you can't measure it, it doesn't matter, right? But if it's impacting the customer negatively and you can't measure it, it's it's a blind spot that is gonna be very painful. Um, visibility isn't just about knowing what went wrong, it's really about understanding how we got there and why it happened. And that's where that visibility comes from. If you think about what you do with human agents, right, you're gonna do quality management, right? You're gonna score them, you're gonna have questionnaires, and you're gonna say, you know, did Jeremy greet the customer appropriately? Did Jeremy use the right tone? Did Jeremy refer to our products correctly, and all of those things. You need the same things to happen with your AI. You need to be able to test it to make sure before you roll it out that it's gonna respond the way that you think, that it's not gonna start saying harmful things, that it's not going to lie or give information that's incorrect, uh, that's going to be harmful to your business or your customer. And then you need to be able to score those and use that visibility and that data to inform future interactions, right? I mean, that's what you're doing with the humans. Hey, how did Jeremy respond to this? Man, he's really good at handling these types of questions. Let's use that as guidance to inform how everyone else should do this. You need that same type of life cycle with your AI agents. Tony, what are your thoughts on that one?

SPEAKER_00 12:14

Yeah, yeah. Thanks, Jeremy. Yeah, definitely a great visibility. It's not just knowing what actually went wrong. You need to know how did you how how did the AI agent even get there, right? Why did it make those decisions? So the way we think about at Amazon Connect is how do we help leaders to defend their AI agents' decision or their decision-making process. And really, I think it comes down to getting that visibility through an extensive amount of testing and simulation so that we let AIs to make all the possible mistakes in the testing instead of rolling it out there with real customers. So, in the example that Jeremy gave earlier, the you know, like selling a car, right? So if there's an AI agent that's selling a car and customer asks for a better pricing, and you know, provide a budget or like$18 or I don't know,$50, and when the AI agent was implemented, its goal is to close the sale as soon as possible, right? And at the same time, the AI agent wants to satisfy the customer while meeting their budget. So what could happen is the AI agent could sell a$30,000 car for$50. So so without that visibility Yeah. So exactly. So without that visibility, leaders wouldn't be able to see what actually went wrong in that decision-making process. But if they have it, they would see, okay, you know, there's might be a garbreal problem about like pricing authority. Maybe offering the pricing must be escalated to a human agent instead of AI making that decision. Or maybe there's another type of rules they need to add, adding that price boundaries. So AI agent wouldn't accept like a price low as like$50. So, yeah, overall, I think when you combine the visibility and learnings uh together with what we're what we are supporting and developing as the observability tooling, uh leaders can really course correct their AI agent before they are even in production.

SPEAKER_02 14:26

I love that. And I think it's um I think it's something we really need to just kind of pause on because we have to build this stuff in before things are deployed. So in the the experience that you you guys have both got doing this like lots of times, there must be some real kind of non-negotiables, is probably how how I'd call that. What what what are those things that really need to be in place before an AI agent can ever speak to a customer?

SPEAKER_00 14:58

Yeah, yeah. I I can offer some of my perspective. So for me, I believe there are a few things that just have to be in place, but definitely not limited to the ones I suggest. Uh, first, I would say the guardrail. You gotta have a guardrail, like Jeremy said, you gotta prevent those harmful speech in proper languages or even sensitive data from being exposed to the customer. Um, and now with those actions that AI agent can do, you have to prevent AI agent to pull like you know other customers' sensitive information or even the company's confidential information before that's you know before all that's happening. And second, I think about transparency. Um, what that means is basically you need to let AI agent to explain what it will do before it actually does. So this gives the customer feels like they're in control. So the things like payment processing, you definitely need the AI agent to get the consent before it proceeds into those like those types of critical actions. And third, I believe is the escalation path. So honestly, having the fallback, it is critical. Because some customers they're frustrated or they don't even know what question they need to ask. So you definitely need to teach AI agent how when or how to handle those types of situations and then hand off to a uh human, uh human agent. And the last one I think is obvious, you need to understand the AI agent's behavior across all of the scenarios possible, not just only the happy path. Uh so basically collecting all those real data evidence that you know your AI agent can actually handle those edge cases without actually breaking the customer experience.

SPEAKER_01 16:48

This this is something that Tony and I have been talking a bit about, Rob, lately with a number of different customers. Um I I think back to the earliest days of speech recognition, and it sounds like, like me, you've got several years' experience in this space, right? Um, in the early days, I remember the the contact center managers, supervisors, or the IT designers would sit down and think that they could come up with a list of the way the customer was gonna ask for things, and then they would roll it out and be shocked and amazed at all of the crazy ways that the customers actually said it, right? Like they never had the list right. And we would have to go in and then edit and add all this stuff, and like back when we were doing grammar files, I wanted to gouge out my eyes with a spoon, right? From staring at all those documents. But but um here with Connect, you can turn on transcriptions, we can have the actual conversations that the customers are saying to the bots, to your agents, so you can take that data and use that to inform your test cases and how you're thinking about this, because you've got what they're actually saying, not what you think they're gonna say, right? Because like you think they're gonna call your products by the full product name, and they're saying the widget that does the thingamabobber, you know. How are you gonna have an agent that knows what's going on if you're not using that and then testing for those and providing the right guidance? So um that's that's something we're we're constantly talking about, Tony and I, uh, in that loop to to inform, you know, have the current interactions inform the future experiences.

SPEAKER_02 18:35

You make a really valid point there, Jeremy. I I think when you think about contact centers and organizations, they always I think they always open up with we're different than everybody else. Um but then they come with this list of stuff that anybody could have found and and and any company could suggest is what their customers are before. It's only when you listen to the actual conversations that you find out you know you're right, you are different, and your customers are different, and that's why you can't just take this like standardized approach to everything, it's got to be tailored, and to be able to base that on your real data, that's just the powerful thing, right? So that's that's great. I get that, and I think it's a really valid point. I think what I'd like to just move into that a little bit is trying to understand how um un how helping our leaders in the contact centre to understand how all this stuff works, um, how how does that help them? How does that look to them? And and what changes when they understand how these AI agents are making their decisions rather than just seeing the outcome of the decisions, it's kind of when they understand it better. It must be a really positive thing that.

SPEAKER_00 19:51

Yeah, yeah, I can offer some of my perspective. So I think it changes everything. So when you know leaders can actually see how the AI agent arrived at a decision, they know the source of the decision and the process how an AI agent made the decision to the step that caused the outcome. So when leaders know that, they can actually define decisions based on their clear rules or guardrails. So they can actually tailor the decisions to really fit their customers, and also they could find gaps in their knowledge base or documentation where they see there's there's a gap of knowledge or ambiguity, uh, and they can improve based on that knowledge. So this is I believe this is a foundational piece that has to be in place for leaders to really improve and iterate on their AI agent to improve the future outcome. And even I think the benefit is like even something goes wrong in production, they will know how to fix it really quickly and also prevent from happening the next time. So that every time something goes wrong, you're not scrambling to look at, okay, you know, what happened, right? Who who who did something wrong here, who implemented this? But you actually understand what the root cause is. And Jeremy, you want to add anything?

SPEAKER_01 21:11

Yeah, I would say, you know, the way I think about it is like I mentioned earlier, it's the same thing we do with human agents, right? When it's a human agent in that situation that did something wrong, the leader wants to understand how and why they made that decision, what documentation they were looking at, what what guardrails or internal tools they were using, and then saying, how do we fix it? The same thing happens with the AI agents and the tooling that's being used here. Um, it just can't be a black box because they can't just go ask the computer, Hey, why'd you make that decision? They have to have the ability to look at, you know, the the logic, the reasoning, the documentation, the bit that it chose based on the context. Um Um and it it's really about the context, right? Like we we all know on this call uh the the power of prompt engineering and how it can change the way the same model answers a question, but most customers don't understand that yet. And even, you know, most you know normal people that are gonna be calling in and interacting with these situations, even though they might play with chat GPT, they're they're not really good at prompt engineering yet. They don't understand it, they don't understand the power of it, and uh, you know, it affects the way the answer happens. Those leaders have to be able to understand that and defend that, or to Tony's point, build in to prevent it in the future if it was an outcome that they don't want or can't allow, right?

SPEAKER_02 22:51

Absolutely. I think um it's really I this what the way I like to phrase this is it's a it's remembering that it's it's generative. The key is in the word, it it's not the same every time, it will be different, and there will be wants in it. But um you're right, it's to us it's simple that that is something to call out, but it's it's it's kind of one of those no-brainer things that people need to be aware of. Um okay, so if we're kind of a lot of organizations are still measuring success based on some really basic stuff, containment, maybe deflection. Um is that still what we should be doing from your perspective, or is that kind of undermining anything?

SPEAKER_01 23:36

Uh it's definitely undermining the customer experience. Historically, it's done it a lot. Um if you force a customer to stay in a channel or down a channel that they don't like because you're trying to say, I'm gonna contain it in the cheapest resource, the IVR, that's why we all raised our hands earlier and said we had bad experiences, right? Like, um, we've been forced down this. We've been trapped in horribly designed IVRs that you know had endless loops instead of treating us like a customer that you care about. Um, containment and deflection is generally about avoiding a contact, getting to a more expensive channel, like with a human agent, because they feel like it's something that could or should be answered. Um but it that's not always true, and the customer doesn't always feel like they're getting the answer that makes sense to them. So while they still matter, I think looking at it from a customer experience perspective is far more important. Um, I can't remember the stats exactly, but I remember a few years ago we we released a stat that said something like that the average number of bad interactions that a customer would would be willing to tolerate was somewhere around 3.8, if I'm remembering right, from the earlier days when we launched about nine years ago. And uh more recently it's like a 1.x, like 1.2 or 1.4, right? It's nearing if you get it wrong once, they're gone. And so if you're trying to force containment without weighing the customer experience, you're risking creating those bad experiences and losing those customers to the to the point we were talking about earlier. So focusing on um allowing the customer to get the right experience, lowering the customer effort, the customer frustration to drive customer satisfaction or net promoter score, however you're measuring that, um becomes far more important, right? You want the easy things to be answered quickly, and the customer wants that too, right? Like if I'm calling to find out the status of my order, it is not worth me listening to five minutes of smooth jazz the girl from Epanema is walking to get to you, Rob, to have you say, uh, what can I help you with? And I say, Where's my stuff? And you say, What's your order number? And I say one, two, three, four, and you say it shipped yesterday, here's your tracking number. Cool. Honestly, you added no value to that conversation, right? And it wasn't worth me listening to five minutes of hold music. So automate anything like that, sure. But if I don't have my order number or I can't find it, or I don't know, and I just want to talk to somebody, give me a quick, simple path to get there as well. Don't try to force containment and say, we failed. Sure. Look at that data and see how you could do it better, but if you're pressing containment or deflection as the key metric, you're losing out on the customer experience side of it. Sound good? Makes sense?

SPEAKER_02 26:49

Yeah, absolutely. I'm just kind of I'm reliving some terrible experiences as you as you I think anyone watching this can probably do the same thing. And so we're aware of this, you know. So but I think we're just kind of rushing and trying to do stuff almost for the sake of doing it, like roll out the new shiny technology without actually getting those kind of little bit of rules around it and making sure that we consider the bigger impact. Um, I guess what we should do is if we should we should talk about if if those are the wrong things, uh and you kind of have touch touched on this around kind of maybe looking at the kind of CX outcomes uh as as as as ways to measure uh instead of uh things like containment. Um what what what are the what would be your top uh recommendations for metrics that should be looked at uh based on what you've seen companies uh succeeding in using?

SPEAKER_01 27:44

Yeah, I I think before the metric, the guidance I would give the leaders is you've got to align the use cases to the right technology, right? AI, generative AI, large language models, it's not like the Lord of the Rings. It's not one ring to rule them all, right? It's it's right to a right place, it's really good at certain things and it's not so good at others. Uh you said like it's probably gonna get it right. It is a probabilistic model. When we look at how it works, it is predicting the next word in a sentence based on it it's a token to LLMs and and to AI. Um but it's not deterministic. And for many of the customers' use cases, they need deterministic outcomes. They have things that are ruled or governed by compliance. You know, billing has to be done the right way. If you mess up my billing and overcharge me, you're not just gonna have an angry, upset customer. I'm gonna go tell everyone I know not to do business with you, right? Um, so you've got to be able to use AI in the right places, but then hand off to deterministic tooling where it's needed to guarantee those outcomes that have to be handled in a specific way. And so when customers go into it thinking that AI is magic that's just gonna solve all the problems, their their outcomes and their measurements are gonna be wrong. Then it becomes those those containment, uh, deflection versus customer experience, net promoter score, and customer effort measurements that really matter. And you should always be using two factors, right? Don't use a single thing. Just like you do with, again, human agents. If you tell human agents they're gonna be judged on average handle time, boy, all their calls are gonna end in under two minutes, even if that means they hung up with the customer or a transfer failed, right? But their metrics are gonna look great, but you have to combine it with customer satisfaction or something else to say if you're doing it in two minutes and every customer's angry, you're still failing, right? But if you're doing it in three minutes and every customer's happy, we're gonna be okay with that. So you've got to find the right combination, you've but you've got to be thinking about it from that customer experience point of view because we're getting closer and closer to that one and done. If you get it wrong once, you're losing that customer, and you can't do that, right? The cost of acquiring a new customer is far greater than the cost of keeping an existing one.

SPEAKER_02 30:20

Right. Absolutely. So you've uh do you know what I've loved throughout this whole conversation is the amount of times we've come back to treat you like you would a human. Think about what you've done with the humans, right? So we've learned these lessons before. We know we kind of know what to do. We just need to remember that we the AI, we need to kind of treat them in the same way, so and and and kind of don't skip any of those steps that we would step if it was a human. And that's my biggest kind of takeaway here because we we're not learning anything brand new, we're just shifting it onto an onto another entity almost. So thanks for that because I think that's that's a really powerful takeaway. I think um what I'd like to do kind of just before we wrap up is that there are going to be audience members watching this who are currently feeling quite caught up, really. There's a big pressure to scale um at right now across organisations. There is this kind of fear or whatever the root cause of it, of exposure of doing things and not getting it right. Um there must be something, a top tip, or a f a sensible first step that you guys can recommend that they could kind of make in their business like this quarter to get things rolling.

SPEAKER_01 31:33

I I'll I'll start and then I'll hand it to Tony. Uh, for me, strategically, I think I think of the first step is take a step, right? Lots of organizations suffer from analysis paralysis, and you need forward movement. So pick somewhere small, pick uh, you know, uh a safer uh example or use case, and you've got to think about those use cases like we were just talking about, you know, something that that is okay probabilistic or that is a combination of we're gonna use the probabilistic component to do intent derivation, but then we have a deterministic flow to handle, you know, if you're the airlines, cancel, confirm, or modify, or rent cars, whatever else, um, or uh, you know, billing or something like that, uh scheduling, you know, pick uh a use case and peel it off and start to test and iterate on it. Go through your testing, give it some thought. Um, but you've gotta get going or else you are gonna get left behind, right? It it's I don't know how many of the listeners are gonna be fans of the Clayton Christensen books, The Innovator's Dilemma, the innovator's solution, and seeing what's next, right? Like I've read those books multiple times, and that's something I think about is if you're not moving in the forward direction, you're gonna get leapfrogged, and then you're gonna get left behind. And so you've got to get some motion going. Um, don't don't overthink it. Think about the things that we've we've said and do the testing, do it in a right, safe way, but start small, iterate, and then grow from there. Tony, what would your take be?

SPEAKER_00 33:19

Yeah, yeah, I think at first I want to start with, you know, take risk, it's fine, right? The biggest risk is not taking any risk. So taking calculator risk, so that I feel like I encourage you to experiment with your minimal viable or minimal lovable product in a control environment. What that means is really having that small set of real customers to reach out to your AI agent on the scenarios that Jeremy mentioned prior, so where the risk of failure is low. Right? Things like common FAQs, maybe things like order status tracking or even appointment rescheduling for non-urgent routine type of like deployment. Right? So, and then with those control experiments, have your measurement in place and also define what success looks like so that you know when you're ready to scale to other use cases. And what I want to say is at Amazon Connect, our roadmap is purely driven by our customer feedback, and we want to know what we can solve for you. And here I lead the testing and simulation um area at Amazon Connect, and we are continuously building better tooling so you can test and evaluate AI agent at scale before they ever reach production. So, to close this out, Jeremy and I are eager to have a conversation with you if you're facing any of these challenges challenges that we talk about uh today, uh, and we're looking forward to hearing from you. Thanks.

SPEAKER_02 34:55

That's great. I think um I think they can put I think that's something that people put put into practice straight away. Uh and I think um I think you're right. I think um I think there's loads of uh opportunities, loads of examples for things that are safe to deploy and and can't cause that much damage. So just run at it, let's have a go.

SPEAKER_01 35:15

Um unfortunately I'm there's there's internal use cases too, right, Rav? Like start with your internal help desk or uh HR question line. Like there's easy ways to get started internally before you expose it to customers.

SPEAKER_02 35:28

I hadn't even thought of that. Yeah, absolutely. It's another another example of how you can limit that risk without stopping and not starting. So I think, yeah, I think uh yeah, if people put their mind to it, they'll find all these things available to them. Um, unfortunately, uh, gents, that's that's all we've got time for today, at least. Um Tony, Jeremy, thanks very much again for joining me and answering all my questions. Um, just before we do close, just going back to what you mentioned there, Tony. If if anyone out there does want to get in in touch, uh wants to explore this in more detail, what's the kind of best way for them to find out more about Amazon Connect or to get in touch with either of you guys?

SPEAKER_01 36:08

Uh so I think we're both on LinkedIn. You can find us there, happy to make new connections. Uh, but definitely visit uh Amazon.com slash connect, right? Or Google it. I don't remember the exact URL there. Uh but you can go there, you can fill out a form for contact anywhere in the world. We have specialists all around the globe. It's not just the four guys we launched the service with nine years ago. Um, so we can get the right specialists aligned to help you with your needs, and we're we're very happy to do that. And thanks so much for having us today, Rob. This has been a great conversation.

SPEAKER_02 36:43

Absolutely, I've really enjoyed it. I think what I'll do is I'll also make sure that we get the right URL on the screen here for the audience so that they can refer to that. We'll definitely do that. Um, and obviously to the audience, don't forget that you can also find a wealth of related resources, all the stories, videos like this one, all at cxtoday.com. So check those out. Um but for now that wraps things up. Um I'm Rob Wilkinson from CX Today. Thanks very much for joining us.