Video: Production-ready AI in customer communications: What it takes, from two sides of the market | Duration: 2744s | Summary: Production-ready AI in customer communications: What it takes, from two sides of the market | Chapters: Welcome and Introduction (38.18s), Agent Use Cases (172.66s), AI Personalization Use Cases (441.415s), Agent Audit Logs (671.17s), AI Deployment Checks (896.805s), Rising Model Expectations (1147.31s), Organizational Readiness (1305.825s), Empowering Bottom-Up Innovation (1472.43s), Measuring AI ROI (1667.61s), Agent Authorization Controls (1929.425s), Protecting Workflow IP (2138.94s), Monitoring and Rollback (2447.325s), Closing Remarks (2709.02s)
Transcript for "Production-ready AI in customer communications: What it takes, from two sides of the market":
Alright. Welcome, everyone. Good morning. My name is Jacques Glick. I lead market strategy at Cinch, including our market research team. And today, we're gonna be talking about AI agents in production and what it takes to get them live and actually keep them there. So before we get to our speakers, I wanna start with a little bit of context here. The way that businesses are talking to customers has really changed fast over the past few years. We've seen a rise of conversational capabilities moving to text, WhatsApp, RCS, email, and voice. Generative AI made those conversations easier to run, and AI voice and video now let agents serve customers anywhere all around the clock. But now we've come to AgenTic AI, agents that are acting on their own. Think of a flight rebooking app that's happening automatically or something that's checking your identity. Agents started in customer support, but now use cases are spreading into marketing, identity verification, customer updates. And all these agents, they really need a a way to reach people, and that's often where Cinch comes in as a provider of trusted, reliable communications infrastructure and channel access to messaging, voice, email. And so as this whole new world was forming, we wanted to understand what was happening with AI agents and what were the challenges of getting these live. So we did a survey where we interviewed 2,500 large enterprises to think about how they were implementing AI agents today, what were the challenges, what were the success stories, and that's what we're gonna try and share today. We had a little bit of a surprise that actually getting agents live was the easy part. It was actually about keeping them live in deployment that's much harder, and that's the the name of the story, the AI production paradox. So our two guests deal with this every single day from different sides of the market. We have doctor Stefan Wenzel, who's chief product officer of SAP Engagement Cloud, which Cinch actually has built in SMS, WhatsApp, Line. And as of this month, RCS, now integrated, and Daniel Sheard, who leads enterprise engineering at Lovable, where we're actually currently building Cinch Communications into Lovable Cloud, starting with email. So, Stefan, Daniel, welcome. Could you each just tell us a little bit about your background and how AI agents become part of your work? Stefan, maybe starting with you since you're. on the left hand side. Yeah. Thank you for having me. Super excited to be here. Yeah. I mean, agents, is is a hot topic for us, especially, for the marketers that we support. Obviously, a lot of brands leverage agents already to operationalize, a lot of repeatable tasks. So our job is not only to deliver the messages, but it basically help entire organizations get these things up and running. And scaling these campaigns, is a tedious work with a lot of specialization and, like, agents can help to bring these specializations down, help marketers to, set up communications engagements and journeys. But also on the consumer side essentially, we're using agents a lot to increase the level of personalization, context and use it for conversational engagements. Yeah. Cool. Daniel. Hey. Thanks for having me. I'm lead, two parts of of Lovable. One is around our enterprise product. I think what a lot of people don't know is we build Lovable on Lovable. So a a big part of my my day job is is figuring out how we build the product that lets us build the product, and and how do we make sure the controls and the the guardrails are in place to make sure that, you know, the app that we build, and run and the agent that we we build and run is able to deal with kind of, like, super high volumes of of direct customer interactions. And then the other part is I lead our forward deployed team, so really working with our our most complex, sophisticated customers on the the kinda hard AI workflows that they want to build and kinda roll out in in production. So, yeah, I think super excited about the the discussion today. Perfect. Alright. We're gonna have one last slide just to highlight some of the research that we did do. There's a full report that you can go to, cinch.com/ai-productionparadox. But, essentially, we surveyed 2,500 leaders across The Americas, EMEA, APAC. Each respondent was responsible for AI in their company's customer communications, and their companies are each sending more than a million messages a month. So these are people that really understand how this works. And what was interesting is that we found 62% of them already had agents live with customers today, and 98% were planning to spend more on deployments this year. What the challenge that a lot of these companies are having though is that 74% of them had to roll back an agent or shut it down. Now these were a lot of different reasons for why that happened. 31% for exposing customer data, 22% because the AI agents were providing incorrect or off brand answers, and another 16% couldn't even trace what went wrong, which I think is an interesting, kinda dilemma in itself. How do you actually monitor this? But it's not for lack of effort. You know, trust and security and compliance, this is the top investment area for all the people that we spoke to. And today, 84% of their engineering time is spent at least half their time on building guardrails for these AI agents. So just a massive amount of investment and needs and engineering time that's spent to actually get this right. But that's why now I wanna turn to the conversation, and we'll stop sharing the slides. And we'll start with just a broad question, Stefan. You know, thinking about the brands that you work with or just the way that you work within SAP Engagement Cloud today, where are customers putting AI agents to work today? What are some of those use cases? And where are they maybe still holding back on putting agents live? Yeah. So so as as I mentioned before, so one of our key users are essentially marketing users shaping and set up setting up campaigns. So a lot of agents are basically being used to create campaigns, to scale campaigns from one market to another, to translate, to create variations of it, but also to to do testing, yeah, of campaigns before they get... Go out, testing in a sense, like other links the right ones. And there's a lot of tedious work in the operational process of getting a large scale campaign or journey ready. And a lot of agents are helping today to, to make this process easier. Now... And then in in front of consumers essentially, like consumer facing agents, And I think it's important to differentiate between those two. Yep. It's really about, like, pushing the boundaries of personalization, bringing in, additional data elements, in the moment, a one to one conversation goes out to really have context, in that next touch point. But even down to really, like like doing personalized imagery on... In in the moment that happens. So we see a lot of use cases that push the boundaries of of personalization with with what they do. And customers approach it very incrementally, very very outcome oriented. So, like, we don't necessarily see that, like, everything is turned upside down, but, like, people taking it piece by piece into, existing processes and move forward very incrementally, which I think is the right way. Yeah. Yeah. What about you, Daniel? How are you seeing this play out? So, obviously, Livable is being used, kind of daily, and kind of AI is very core to to the kind of product propositions. We have, like, people building applications and and operating our businesses on Livable. I think there's probably two things that I'm seeing. One is in, like, in terms of our own internal operations, we've not necessarily had the luxury of of kind of being able to establish operational processes over ten years, and we've had to kind of rapidly get to the point where we can scale to meet the the customer volume and the demand, that kind of we're we're seeing for the the product. And that means that we've had to and and kind of wanted to and it's kinda like core to to who we are is, like, let's build the operations of the business AI natively within Liverpool. And that... That's everything from, like, customer support. How do we triage kinda customer support tickets? How do we kinda drive our, go to market workflows? How is our finance team operating? Kinda how is our procurement run? We've we've built that kind of into the the DNA and the culture of the the organization. And I think the kind of outcomes that have have really been... You know, we've we've been able to, like, five x industry metrics on on, some of the kind of key, go to market metrics as a result. I think if I think about how do you then apply that to... As for our customers and kinda what some of our top customers are doing in this space, I think the the kind of biggest opportunity in the unlock depends a little bit on where you are. But I think the ones who are having the kind of outsized returns are the ones who are experimenting and kind of placing bets. And I think there's depending on kinda where you are in your journey, it depends a little bit where where... To to do, kind of the investment. I think if you're... You've not started yet, there's a lot of kind of traction you can make internally with kind of your internal operations, your internal functions. And then kind of as you build the muscle and and build the the the platform and the toolkit to be able to kinda handle volume and scale and and kind of, reduce the risk and some of the things we'll talk about today, I think then you can start to kind of pivot that into the the kind of external facing use cases, you know, whether it's customer support or marketing as as Stefan was talking about. Yeah. And, Seven, do you sense that also, like, this is really about risk taking a little bit, and maybe it's being comfortable with the fact that you are gonna have rollbacks of some of these things, you know, and and just being able to iterate and test. Are you seeing that as well that it's mostly, like, people that are willing to take that risk are getting huge outsized, advantages on return on that investment, and maybe those that are more hesitant are just not truly seeing or haven't opened up the value that AgenTic Systems can offer them? Yeah. I mean, like, rolling back might always be part of, of a larger process, but I... As I said, I think the the smarter way is to take it an incremental approach. We are having a lot more attention to data that is being produced. People looking at what happens basically. Every action is basically being locked and through through these iterations, changes are very natural and, like, avoiding that, like, we have to break complete customer experiences and, like, like, roll back an entire process as we saw, like, for many years back with the first generation of chatbots on websites where basically people have been sent to, to an immature chatbot that also did not provide a solid alternative or anything like that. I think, like, it's it's like people are more... Or customers are more aware of that, introduce that incrementally backed up with data analysts that see what happens. And that that helps them also to iterate iterate quickly. Unlike the audit logs, for those that aren't super familiar, you know, can you give a little detail and just break it down for folks? Like, how would you recommend they create some of these logs to track what an agent is doing? Because I think that's often part of the challenge here is that people struggle to actually know how to track when an agent doesn't perform the way you want it to. So maybe just a small example, if you have one to share, that would help bring this to light. I mean, there's there's not... This is a huge topic in general. So so for internal agents, for example, agents that are being used by business users, we have a so called responsibility management. That means every agent has a assigned responsible human. There's no agent that does anything, a business transaction without a a human that is responsible. That's that's first of all, like, important organizational model. And then, of course, if you have that assignment, there's, let's say, a marketer and the marketer has, their their, agentic companion, then the marketer needs to fulfill that responsibility. How do you fulfill that responsibility in a way that you get notified about transactions that the agent has done? You're getting notified. You have access to these logs in a human readable way so that it can also be processed. So responsibility management needs to be translated into practical responsibility management. And if you translate that also to, the consumer side where you might not have always like that opportunity to always have or trigger human in the loop process because, like, you might have millions of consumers that in parallel interact with a consumer facing agent. There's additional. measures that need to be taken. Right? So, like, you can't have for every agent interaction, the human read through all the logs. So you have... Need to really have a sophisticated mechanism around, like, notifying handovers, and and and guardrails. Basically, building out that harness that allows this this, agent to operate within its with it within its, its guardrails, but always have them, yeah, these these logs available so that data analysts can follow-up what has happened, both in the individual case, but also, in the in the aggregated statistical case, so to speak. And, Daniel, for for your enterprise customers, what are some of the checks they're putting in place before an app that they've built on Lovable or some kind of agentic workflow, you know, ends up going live in production? What are some of the checks that you're recommending they go through? Yeah. So I think probably the first thing to think about is when we talk about AI and we talk about agents, I think not all kind of workloads necessarily look the same. I think there's, like, it's kind of two broad categories. There is agent kind of doing the orchestration, deciding what to do. This is like a chat experience where it's very open ended. You don't know kinda what the the attack vectors are gonna be. And then there is, like, workflows where AI will do some of the steps, but you can kinda bound it in kind of the end to end control in the same way that you would think about kind of any other software process or automation. And I think. the way to think about that is slightly different depending on the paradigm that you're kinda solving for. I think it's a lot easier to start with, in my view, workflows, where there is a a kind of an input and an output, and and you you weren't able to automate it before because there's a there's a step that historically has had to been done by a human. AI allows you to solve that problem, but it doesn't mean that AI is the only thing can solve the end to end problem. So I think the first thing to figure out is, like, can you have a deterministic check whether the output is correct or or incorrect? And I think if you're able to use AI for the decision making process, right, able to kinda use the AI to to accelerate the the interim state, but you can bound it in in determinism and and kinda traditional software, I I think then you can treat it like traditional software, and you can kinda write the test. You can write the evaluations to make sure that you're kinda meeting that that end goal, and it's it's kinda not operating outside of the the the guardrails. I think in the the former case where you're kinda doing something like a chatbot, you know, know, you're gonna put it on your website, you don't know what a customer is gonna ask. I think then you have to be a lot more careful with, you know, writing evaluations around, like, what are the, things that the... A customer is gonna try and do, or, you know, a bad actor is gonna try and do to to be able to to kind of, kind of process something in a slightly different way. And I think there's obviously stuff you can do from, like, a brand, the messaging point of view. Obviously, you don't want the agent to kind of respond in a way that's not, in the same way a human at the company would respond. But I think a lot of the actions that the agent can take can then kinda follow the workflow in the kinda more deterministic route, and you can put the checks in place on the tool call as opposed to necessarily try and do that in the agent instruction. So I think it always comes down to, like, how do you try and decompose the problem into into the different component parts? And then how do you check each step in that decomposition of the prob... Problem as opposed to trying to, like, get the agent to do everything and and kind of really kind of make the the instruction super broad. And then I think the the other thought is the other thing is you... Not all software is the same. There are some applications which are used by, like, a team of five people and kind of internal. access out of the way. I think you should put less control and invest less resource in in kind of managing that. There are other applications like our Livable main agent, you know, which is building 1,200,000 apps a day. Oh, sorry. A week. If if, you know, we we need to make sure that's very high quality and, you know, isn't isn't gonna kind of, do something that... That's, gonna make us look silly. So I think the level of which you kind of ladder that up, I think, is very much an organizational question. How do you actually kind of create the, kind of IT operating model that allows people to create software in a way that is is so proportionately governed to the to the use, to the reach, to the the mission criticality of the the software. And I I don't think AI changes that. It changes, like, how do you test that? And it changes, like, the the determinism and the probablism of of how you can can check it. But I think fundamentally, Yeah. you just try and bake it back into kind of, like, what is the the new paradigm with which we want to given that that end to end software? Yeah. And it almost feels like, you know, advancements in models, like frontier models, are act... Obviously important. But when you start to think about agentic workflows, like, it's also the foundational pieces. It's infrastructure. It's data that you're putting in. It's it's exactly what you said, the control of deterministic flows versus more general flows. But I am curious because the the models are getting so much better as well. Is that making it easier in your experience when, you know, you you get a big upgrade to a model? Is it improving experiences or or lessening the load of some of the teams having to build some of the checks in place? Like, are we seeing that there's an improvement there or no? Well, I I think I think the expectation goes up with every model shift, and therefore the capabilities of what the the agent is able to do and should be able to do also goes up. So I I don't think the it becomes easier because I think the goal post move with every model enhancement, this is gonna... Something that's gonna keep happening. The acceleration, you know, at least as far as I can predict, kinda continues. But I think also with that, the ambition level is is rising as well, which I think is super exciting for us as an industry. I think the core muscle, though, of of writing the evaluations and of kinda, like, you know, doing the incremental kinda rollout and how do we think about kind of, staggering, you know, stuff in those in more kinda high risk situations, I think is is equally important. I think I think what is good is, the the model vendors and and kind of in post training, we're putting the guardrails closer into the model. I don't think that was necessarily there. Two or three years ago, I think you could more easily break the core model and kinda get the extract the instruction. I don't think that's that's as easy these days. But but, fundamentally, I think if I think through, like, what people are building, like, every week, I'm impressed by a new thing that someone's built, both on our platform and and kind of within the industry as well. And I think from a controls point of view, we need to keep leveling up as well. So, like, we... We're giving, you know, builders access to see all the agent logs and to see the traces and to to be able to observe, like, why did the agent make those decisions. You know, we're able... We're having to put more checks in place around, like, how do we detect personal information and kinda make that accessible to the agent and to the builder to to be able to manage and monitor that. So I think I think every week it it changes. I think the operating model is a thing that needs to continue to evolve in parallel. Otherwise, you know, the the the art model gets left behind and and kinda gets stuck in the old world. Yeah. I think from from my end, as I think about, like, the models getting better, it primarily broadens the applicability of, the use cases that we can serve with it. So, but on the introduction, there's a certain pattern that remains relatively stable. So you... The agent and the LLM is only as good as the data you put in, as we said. Yep. And a lot of focus is basically around the harness, like a special term that we use, like like we talked about guardrails, but guardrails is just one element of the bigger harness. So so is the organization around it basically ready to, to operate that agent within, and which tools do you provide to to the agents to make that work? So a lot of effort is basically spent around providing that environment, so to speak, and the organizational surroundings to take, like, most advantage, out of the agent. So so that element, is is not fully understood by many organizations just yet. Yeah. Because, this also involves the human organization around it, the human harness, so to speak, and that very often lacks behind or takes simply more time, yeah, to to to do it. Like, let me give you one example. So a lot of, like, we have seen a lot of specialization in marketing. Right? So we have seen specialized teams for channels. We have specialized teams for content, imagery, copy, translation. Agents really allow us to rethink these specializations. Maybe turn that specialization back and have more empowered integrated teams again. But to make that happen simply simply takes time, and, like organizations need to be prepared. And if they don't do so, they might not see the outcome they they wish for. Right? Yep. It's funny because when we did this research, only 3% of leaders said that resistance to change was their biggest barrier to getting agents live. I suspect that's a lot lower. Or or, you know, there are important other factors like, you know, the cost and, just overall, like, infrastructure and pieces like that. But I guess, what what do you think, you know, coming from an AI native side, Daniel, like, what has to change in organizations that didn't start AI native in order. to be able to move at speed, and to and to just kinda take this on from, like, an organizational process, an ownership, a skills level on the human organization side. Yeah. I I think, I was reflecting a little bit on, like, if you think back to, the, like, Toyota Kaizen model of, you know, how did kind of innovation happen on the factory floor, like, kind of back in the day, and and the kind of innovation that they made in lean manufacturing, everything that came from it. You know, the person closest to the problem gets to to fix the problem. And they they they can solve it, you know, and and kinda empowered and they're incentivized to do it, and that... That's how you get the return. I think then, you know, when IT became the constraining factor, it was a, you know, an expensive scarce resource. You know, you you needed to have done it for ten years to understand how it worked. And I think what that meant was that fundamentally innovation broadly stopped. You know? And I think we've had this period where people haven't been able to innovate because a great idea needs a, you know, 10 page business case and a 100 lines of Excel in order to be able to build the thing that that moves the business forward. So you need to have high certainty that it's gonna work before you you ever kind of put your hand up and say, alright. I think we should do this, and I'm gonna fund it. And I think what's changed now is, well, that constraint has gone away. Like, the the cost of building software is converging on zero. And I think what's not changed yet is the operating model in IT organizations in large organizations to to really harness that power. And I think this is where AI natives and a lot of kind of the more nimble organizations are having, kind of these outsized returns because they're not thinking about it in the traditional sense where you configure an IT team to look after 300 applications. And and, you know, if you want 301, then, oh, how are we gonna manage that? I think actually now, like, software is becoming a lot more dynamic, and you have your systems at record, and you you build this... The kind of new software on top. But you need need to think about that new software in a in a slightly different way to you... The way you thought about it before where you you don't invest a huge amount of resource in it as long as you're comfortable that the guardrails and the the the platform is taking care of the policies that the organization have for what what good looks like. So I think what change is is as leaders in the kind of industry and as kind of leaders in... As I advise my my clients, like, I think you need to be setting the expectation with the people in the organization who are closest to the problem. Now you can can change it. Now you can fix it. But I think we need to do that from, like, kind of the bottom of the organization. Like, empower the people to be able to have the tools to do the work, but also set the ambition and set the incentives from the top down to say, actually, this is what good looks like. Let's celebrate these success stories. Let's place, you know, these bets. And at the point where somebody, is is, you know, having an early impact on a pilot program that they've driven from the business in a in a nontechnical way, put the resource there. Don't put it onto these these big programs of work, you know, that you don't know if they're gonna necessarily prove the impact. So I think I think we wanna see more results first now. I think the kind of world where you can ask, you know, for a huge amount of budget and then not show any results is is kinda gone for me. But, yeah, I think that's the the the main change. Yeah. We're gonna head to q and a soon. And if you're on LinkedIn live, we'll pull those questions in. But let's switch to kind of return on investment and some of, like, the immense benefits that you both are seeing in the world with AI agents. I mean, one interesting tidbit from our research that I found really interesting was that the top goal for the executives that we surveyed and and the leaders on IT and marketing that we interviewed as part of our survey, the top goal was not actually to reduce costs in the organization when leveraging AI agents. It was to increase customer satisfaction, at 36%, and then reducing costs or driving revenue was 24%, and 16% concurrently. So to me, the value is harder to prove on increasing customer satisfaction compared to revenue and cost. But how are you seeing that investment pay off in the real world with your customers, and the agents they're deploying? Stefan, if you wanna start. Yeah. Of course. I mean, like, as as a customer engagement platform, a lot of what we do is basically driving campaigns to a point of sale, but we're also attributing goals towards loyalty customer retention, and more and more also operational, engagement processes along an entire customer life cycle. So attribution of some of these moments, whether that's point of sale or whether that's like retention elements, is part of our platform. So the application of AI is directly being measured as as part of what what we regularly do. And I think that's a that's a great advantage because we, yeah, immediately can see the outcome, basically on on on those business business targets. And this is what then also customers directly basically feedbacks they and where we see the usage. What works is being used, what not works is not being used, as simple as that is. And... But, of course, we we definitely also see this element that potentially often is mentioned on the cost or efficiency. But reality is it's, I think, more of what I touched before a little bit on on taking back empowerment, to some of these these users where super specialization has driven them to some inflexibility. And now we can think things quicker and faster and ultimately with that agility then also deliver better customer experiences and meet meet these these, these these these new new levels of expectations that are also out there. Right? Daniel, your thoughts? Yeah. I I think fundamentally, we shouldn't invent new metrics to understand whether AI is returning. I think we should take the metrics that are already in place at. the kind of functional level and and see if we can move them. Because I think the... You know, to kinda come back to the point on, like, leadership setting ambition, you know, I think there's, cost is, is the lagging indicator. Right? It's very hard to do that, you know, in a very fast way and kind of an in quarter, especially. But I think there's leading indicators that every organization is tracking, you know, in every team to say, like, are we are we really kind of performing? So, you know, is that meetings booked in sales as a leading indicator, or is it... You know, what's the contract spend in in procurement, or, you know, what's the time to resolve in support? I think these are broadly productivity metrics that have a lagging cost impact or, you know, a lagging revenue impact, you know, in in in the sales case. But I think fundamentally that when we kind of looking at where to deploy AI, and I think a lot of boards and and kind of leaders are are wrestling with this, I think you should be deploying AI to to move the needle on the things that are already important to you, not inventing new things that that might be important in order to justify the the AI. So, you know, if customer satisfaction is already a thing that's top of mind, great. Like, you've already proven that's important because the way you've been running your business for for the last five years. So so go move the needle on that with with AI. Yeah. I'm seeing no questions in our q and a right now, but we do have some questions from LinkedIn. How do you stop an agent from committing to things it isn't authorized to give? They're giving examples like refunds or discounts, delivery dates. Is that a... At a prompt level policy. or hard checks before the message sends? I mean, this kinda goes to the conversation we had before on... It it really just depends on the workflow, depends on, like, the magnitude, you know, of the scale of how many interactions would have, whether you do this deterministically or not. But, Daniel, maybe since you kinda touched on that in the session, do you wanna take that? Yeah. I mean, if if the alternate act to the agent is the call center operator or the the kind of other person who who might have been doing the work before, or I guess the question is how would you stop them doing it? And I think it's not. usually training. It's usually tooling. So I think, you know, if if you only want a, you know, a customer service representative to give a refund if certain conditions are met, again, you should deterministically check that those conditions are met as as as... You know, and and do that outside of the agent and and make the... You know, have the core decision making outside of the agent. I think if you don't want the agent to have access to to to make certain decisions or certain tools or, you know, to make certain calls, and just don't give it access to the tools. And it comes back to Stefan's point around the harness. The. harness, Yeah. you know, really should should only give the agents access to things you want it to do, and it should only have access to those things when certain conditions are met. And, you know, there's there's ways to be able to do that such as, like, you know, asking what the next best action is given a a scenario, and then, you know, kind of... Then the agent only has access to the next best actions. You know? I I think I think we we... I think we give too much power often to agents when we're not experienced building them. And and and the first thing, again, it comes back to problem decomposition. Stefan, you agree? No. 100%. I I think, like, if you don't want a shopping agent to give out, discount code, basically, you need to control the tool that produces the discount codes and, like, have eligibility checks. So it, again, comes again down to the point that, yeah, you you build the surroundings in a way that, like, the agent within, basically can operate at its best. So sometimes, like a conversation with a colleague, think of the think of the new LLMs like a really powerful bull. Yeah. Like, it has all the muscles twenty four hours seven. So you don't need to worry about the muscle of the bull, but you need to worry, like, where he runs. And so you have to build the walls left and right, so that it doesn't run into the crowd. So so there's more attention to that. Don't worry so much. Like, we are all trained, I think, when we think of humans putting a human on a task to train the human on the task. But now what if the assumption is the human that we will put there or the agent is always trained? Like, we need to ensure that the the the surroundings that it, like, does what it does within within that controlled environment. So, like, thinking a little bit invert, so to speak. Yep. Another question from LinkedIn. With the increasing ability to reverse engineer systems and processes, how can organizations protect their original workflows? What strategies can be implemented to secure and preserve workflow integrity within mainframe based environments. I don't wanna take a stab at that. That's a tough one. I think that was some more context, but then gets a difficult one to answer. I know Stephanie, if you have a view. Yeah. I mean, like, I mean, the the the the question is, well, like, how can you protect, your company's IP in that sense, if that's the question. Right? So, I mean, lot of companies' IPs coming down to data, reverse engineering data is, more difficult. So companies need to have strategies, like, also, to, yeah, protect this data, intellectual property. Of course, if you produce software, I've seen, like, there there are slight clones of known softwares out there that is published on GitHub every time... Can download. I think these things, happen now and, regulations somehow have to deal with it. I can't say much about it. But from a com... Company perspective, a lot of a lot of what's to be productive in data and relationships and... Yeah. Yeah. I I definitely agree. I think it's the partnerships you create that ends up being the value, like, where you sit in the ecosystem and the value that you're driving with those other trusted partners that have their own, kind of protective systems and things like that. I think that's that's definitely what we see, on the Cinch side. How can we protect REA? Any compliance needs to be implemented in the AI community. I'm not sure what REA refers to. Do you? No. I would need you? to? to go to that instance. I mean, from a cinch perspective, we we think about this a lot. Because when you're talking about communication channels and reaching end customers, you are thinking about how do you compliantly do that across different geographies and markets. So a lot of our time spent on compliance is thinking about how do we build that in. So for example, when we are creating a native workflow to be able to send a rich conversational message, you know, within an ID platform, that the agent is taking some of that compliance understanding and know how for the right way to send that message to a European recipient versus an American recipient. A lot of that is is built in from that perspective. It's less broadly applicable to AI compliance, I think, which is a little bit different, but, one one example of how we approach that problem at Cinch. How. do you... I'm curious. guess how do you guys think about compliance in general for the workflows you're putting in? yeah. I was I was maybe just on the REA point before we move that. I I think... I'm guessing REA, is the the open source repo that lets you reverse engineer the the workflow. I'm just kind of reading a couple of these. It's like very much like, Yeah. how do you protect your existing products from, like, basically AI being a new paradigm and and getting disrupted? I I think that it's almost the wrong question to ask because I I feel like, you know, whether your your software is being reverse engineered or whether someone else is building it based on know how they have from the industry, you know, because software is becoming free to create. Like, I I think fundamentally you need to think, like, what's the kind of next value proposition or kinda what's the next, like, what's the second act where we we're gonna have on our existing product in order to stay relevant in in what is fundamentally a changing market where the customers of your product are probably know not just humans now that they're also, you know, personal agents as well. So I so I think fundamentally, you know, I I think if you're only thinking about, like, IP protection on, like, software and and kinda workflows and and not to to a Jacques and Stefan's point on on, like, data or what is the service we deliver that kind of wraps the software? I think it's a difficult place to be if if you're kind of the only frame of reference is, like, how do we kinda protect what we've got? Because the the world's moved or is moving. And I think. you you need to continue to to kind of push the the the the thinking on, like, well, actually, how are we gonna evolve the product to to serve this this new world? Or or even beyond the product. Right? It's like you have so much more abilities of things you can do from a go to market perspective with AI as well. So how are you going to leverage that to reach new customer groups or existing customers that you struggled with? Like, the cost of doing business is going down even in a go to market space. It's not just product development. So I think it's also like, how do you rebalance your thinking and your approach to that full life cycle that actually can make a huge difference? Another question. How do you scale agentic workflows that require human supervision? How do you resolve that tension? I can maybe take this. I mean, our, customer support, operation within Livable is is mostly run kinda with with agents doing the the kind of... The bulk of the work. So they're fundamentally kinda triaging issues. They're kinda looking into, like, you know, problems that that people report. And and about 80%, of our, customer support tickets, like, require no human intervention. But there obviously are 20% that that do. I think to... I think it was Stefan's point earlier, like, the way to get there is one, like, start. I think if you if you never start, you'll you'll kinda never, kind of have it have it as a problem to worry about. But I think what you... What you want to do is is have the agent kinda support the human and and, like, kinda have this kind of pilot or a pilot group or kind of AB test running at all times where you're trying to say, you know, can the agent 100% solve the problem with the human kind of maybe taking the final approved step? And then I think at the point where you have got enough volume through, I think you start to build the confidence, but also the muscle of, like, what are the things that are breaking and how do we kind of think, you know, a step ahead on on writing the evals to to kind of protect and prevent that. I think one thing that we find is super helpful is the kind of the worst person to test an agent or an agentic workflow is the person who built it. Because then you come in with this, this frame of reference that, you know, I built this thing is good. Right. And this is why, you know, historically we had software testers who kind of weren't in the development team, like in the. the the... Have rolled waterfall days. Now I think, fundamentally, then what what do we do? Well, we we we internally run, like, breaking things sessions where, like, people from different teams just come along and try and break the thing that someone's made partly because it's fun to do that, but partly because it creates a better company and creates a better process. So I think I would leverage kind of the internal, kind of human capital to try and break the things that you build before they ever get in front of a customer. And then leverage your customers to break them, but you don't need the customer to see the broken output. You can still kinda have the human layer there, while you kinda train the the model and train the motion. Yeah. That makes total sense. I got one last question. Just wanna squeeze in as many as possible. But what specific monitoring evaluation rollback practices have actually worked for keeping agents reliably live, especially the difference between open ended agents and bounded ones. Stefan, what are your thoughts on on that as... From a CPO perspective? I mean, again, I I think I I I think I touched on it. Like, for for us, the... From simply also a software engineering perspective, the role of a data scientist or data engineer has grown significant importance. So when we roll out agents basically into production, we incrementally roll them out. And then from the first day, that dev ops paradigm gets, like, enhanced by really monitoring the live usage of of the... Those agents and we directly, like, intervene when we see things go go go south, so to speak. And this is an enhancement basically of how we construct our product teams even and how we roll out essentially. So it's not necessarily about, okay, we ship another version of the software, but it's actually operating that agent and doing, doing adjustments as we move. And in that way in that way, the rollback indeed, like, the big rollback is just like the last exit because, like, we ideally would learn and enhance as as as we go. Right? So so that's the general approach around, like, delivering our agents into our software. Now our software also can power basically customers to roll that out, but that that that basically is this this other step. Great. Well, thank you all for joining today. This was a really great conversation. Big thank you to you, doctor Stefan Wentzel from SAP Engagement Cloud, and Daniel Sheard from Lovable. Really great conversation. And, please, if you wanna learn more about the research that we did at Cinch, please go to cinch.com/ai-productionparadox. Looks like we have a link going live now. But thank you, Daniel. Thank you, Stefan. Very, very good, conversation, guys. and appreciate all the questions that we got as well. Cool. Bye, Bye bye. everyone.