Kore.ai’s Raj Koneru Reveals the Multi-Agent CX Shift Leaders Can’t Ignore

As customer experience (CX) teams push beyond basic chatbot deployments, a new model is emerging: coordinated multi-agent AI systems that can execute end-to-end workflows across teams, tools, and policies. In this interview, Rob Wilkinson speaks with Raj Koneru, Founder and CEO at Kore.ai, about why this shift is happening now and what leaders should demand before they trust AI with real customer-facing execution.

Koneru argues that the last 12 months have changed the game because AI models have improved dramatically in both generation and reasoning. That progress moves automation beyond question answering and into task completion, where agents can take action across systems and processes. In regulated and complex environments, he says, a single agent is limited by context and scope. A multi-agent approach better mirrors how organizations actually operate, with specialized functions like billing, fraud checks, fulfillment, and escalation requiring orchestration.

Koneru also outlines what tends to fail in single-bot deployments: inconsistent answers, broken handoffs between bots and humans, actions taken without enough context, and heavy cleanup work for frontline teams. Risk and compliance leaders, he adds, often struggle because prompt chains offer limited traceability and control.

For CX leaders, he recommends insisting on deterministic policy enforcement, clear permission boundaries, human escalation controls, and runtime observability before going live. He emphasizes that governance should not be bolted on after deployment. It needs to be embedded into the platform and runtime so teams can reproduce, audit, and optimize outcomes.

Finally, Koneru shares the production metrics that matter: non-negotiables like failure recovery, auditability, and compliance, plus business outcomes like resolution rates, customer effort reduction, and time to resolution. He also highlights the “soft” impact of better experiences on brand loyalty and long-term value.

Transcript

0:08

Hello and welcome. I'm Rob Wilkinson, and today we're taking a closer look at the next phase of AI and customer experience. This is where teams move from individual chatbots to coordinated groups of AI that can actually get stuff done. So if you're challenged with scaling automation without losing control, or if you're exploring how you might be able to make AI safe and measurable in the contact center and beyond, then stay with us and you'll get a practical takeaway that you can use this quarter in your organization. Because today I'm joined by Raj Kanaru, CEO and founder at Core AI. He's an expert in this space. He's got a really unique insight as well into the decisions and the risks and the opportunities that CX leaders are facing today. So it's great to have him here. Welcome, Raj. Thanks so much for joining me. Thank you, Raj. Good to be here. So last week, Core AI launched its agent platform building on the previous kind of broader push to make agentic AI more measurable. For our audience, just to kind of help uh set the scene for them a little bit, can you just kind of talk to what's changed over the last sort of 12 months? What makes um a multi-agent approach inevitable now for CX rather than what was previously more of a nice to have? Well, fundamentally, what's changed in the last 12 months is uh the models have become much, much, much better, both in terms of generation and in terms of reasoning. So, what that means is software, as we know it, is getting developed by the models more than by humans. That has a significant impact on CX. Now, CX, as we've known it for many, many years, in terms of IVR automation or chatbot automation, voice bot automation, was single-agent. It was a deterministic flow, uh sequential in nature, with some conditional branching. But then, with the availability of the models right now, there are two main things that have changed. Number one, these agents are not just about answering questions, they're about actually executing actions and completing a workflow. Two, the agents are able to work with multiple systems, multiple policies across multiple teams. And that's the value of these agents. A single agent being able to do that is limited in terms of the context that it would use, basically, and involve a single customer journey as opposed to a concierge type service that goes across, for example, in banking, it would go across multiple things like billing, fraud checks, fulfillment, compliance, escalations, humans in the loop, basically, which all requires orchestration across different agents. Because when you think about an organization, each of these are different functions. So the agents represent those different functions with an orchestrator at the top. So multi-agent applications is the future for CX, basically, and it's inevitable. Every organization is going to put out these multi-agent applications. So with the help of orchestration done with the models, with context which is managed at an agent level and then populated up to the orchestrator level, and the orchestrator being able to route to the right agent, that brings it all together for a for a wholesome CX journey, if you will. I love that phrase that you used there, concierge approach. I love that. I think that is a really I mean it's the holy grail of CX, isn't it? Being able to provide that concierge service. One which I think we've struggled to achieve traditionally, but it sounds like with these new models uh and and you know the advent of multi-agent, that that that might be something we're going to be able to improve upon. So that's exciting. Um when we think about organizations moving from just a single bot into those more autonomous executions, what tends to go wrong kind of in the real world for the customers and the frontline agents, maybe even those risk teams, when this isn't kind of handled correctly? And what can our audience learn from your experience? Well, a couple of things. You know, when deployed, sometimes provide inconsistent answers. They break when the handoff has to happen from one bot to another bot or from a bot to a human. The agents take actions without enough context because there's no central context, the context is at a single bot level. And then the employees have to get involved, humans have to get involved to clean up bad escalations. And risk and compliance teams struggle with it because much of the single bots are done with prompt chains, with very little traceability or control. And as autonomy increases, the orchestration is the key. Observability and traceability, you know, of exactly what happened is the is the key. In addition, for regulated industries or even unregulated industries, compliance to business rules is extremely important. That cannot be delivered through full autonomy. So you need to have bounded autonomy along with deterministic flows mixed in with an orchestrator at top, which manages the context at a central level. So the world of single bots, the problems in building them, the problems in maintaining them and observing them go away with a multi-agent architecture. But multi-agent architecture has to be thought through carefully. You know, some of it has to be designed with deterministic flows, some of it with autonomy, where you need personalization, you need dynamism in the experience. So this new architecture that we have now released with Artemis, our new platform, gives you this dual-brain architecture. It is about building AI with AI. You don't build AI manually. We use AI to build the multi-agent application and each of the agents and automatically determine what needs to be deterministic and what needs to be reasoning-based within each of those agents. So that's the future of Agentic AI. It is AI with AI, which is AI to build AI, AI to test it, AI to deploy it, AI to manage it, govern it, and AI to most importantly optimize the application. So it's it really is teams of agents working together, but you can't just put a group of single agents together. You've got to you've got to restart the from the from scratch and plan it properly, design it properly and execute it. So it's it the yeah, there's a bit more to do. Um but it sounds like once you've done it, it's kind of it's set for success and it's going to then continuously kind of uh evolve and evaluate. So that's I mean, it gets it does it, it is exciting. I guess um looking back to the launch, um, you led with the uh agent loop uh agent blueprint language and and a set of orchestration patterns, so handoff, escalation, fan out and federation. Um what problem is that solving that historically kind of prompting and wiring can't actually do now? See, prompting alone is not an enterprise architecture. It may work for demos, but large-scale CX operations like we deploy needs explicit orchestration. It needs reusable execution patterns, governance, and the whole lifecycle management. Asian Blueprint Language is an intermediary language. It's declarative, it's YAML-based, so it can be versioned, it can be audited, basically, to coordinate across multiple agents, systems, and humans with very, very clear traceable execution paths and very controlled behavior. Basically, prompts are interpreted by an LLM, and then the LLM decides what the prompt means. That will not work for most organizations. That's why a multi-agent architecture with the appropriate orchestration patterns that we provide in Artemis allow these organizations to scale these agentic systems in a very structured but very controlled manner, basically. Instead of building a fragile collection of prompts and API integrations which need to be tuned and fine-tuned and fine-tuned to get it exactly right, basically. You use the AI, you give it the standard operating procedures, it will build the right agent topology, it will build the agents by generating the ABL for each of the agents, it will select the appropriate orchestrator, it will do the federation to an external agent where required through A2A. It would do MCP calling or creation of the tools based on the APIs that are available. So it brings it all together. Artemis is the only platform of its type, which then gives you this controlled execution through a language called ABL with a super agent called Arch AI, which does the design and the build, basically, with a dual-brain architecture. Basically, and that is what the need of the enterprise is today and going forward, as opposed to when you look back, it was prompt heavy, and before that it was deterministic only. I get it. Okay. So that covers quite a lot though. So and as you've been going through how comprehensive it all is, that's really driven home to me why it's taken till today for this to be able to be available, because this this isn't this is too complex for the models up until not actually that recently, really. Uh, and although that technology is changing super fast and it's still continues to kind of basically see no no no stop in it. Um, I can completely understand now when you said at the beginning how that was something that needed to happen before we could go here. So thanks for unpacking it a little bit and giving us the the meat on the bone, so to speak. Now, you you said there around uh governance uh observability being almost built in, not enforced even before you're going live, that traceability of agent decisions is in there as well. That's great, but what controls should an operations person, a CX leader, insist on having themselves before they let the agents even execute any real tasks because especially at the start, I guess, of the kind of journey. I mean, look, when you're a CX leader, you need deterministic policy enforcement. You need full traceability of every agent decision and action. Human escalation controls, you know, when do I escalate it to a human? And clear permission boundaries, you know, for system access. You know, what system do you access, you know, what data did you do you access. But importantly, you need runtime observability so they can understand why agent made decisions, how workflows get executed, where failures occurred. And why is that important? For optimization. No matter what Arch AI designs for you and builds and deploys, you may not meet your business metrics. And Arch AI then automatically automatically looks at your traces, your production data, and gets intelligence from it and optimizes it and regenerates the ABL. Governance cannot be a bolt-on after deployment. It has to be embedded into the runtime itself, basically. So it needs to be part of the design of the multi-agent application. See, enterprises' trust comes from visibility, reproducibility, auditability, control, not just simply from model accuracy and latency. There's too much focus on model accuracy and latency, and they think that's the silver bullet. No, the silver bullet is in your multi-agent application, in building control in it, in building observability in it, you know, watching for errors. That happens in the platform, not in the model, basically. So it is very, very critical for CX leaders to understand this and build it right and build it into the platform, into the application, so they have the control, and they're not as dependent on models to implement that control and governance. Great. So it's a that's a really important clarification that so I'm glad you called that out, and the difference uh between the platform and the model. And when you look at this, the platform is using the dual-brain approach. We've got agentic reasoning alongside the deterministic flows, and you've got that shared memory and that single runtime that you talk to there. How do you decide what needs to be deterministic versus what can actually be delegated to reasoning? Well, first of all, RCI decides based on what we've trained it to do, but at a very broad level, you know, reasoning is extremely valuable when dealing with ambiguity, personalization, or dynamic planning, you know, on a set of tasks or a set of steps. But deterministic systems are essential where accountability, policy enforcement is involved. You do not want probabilistic behavior for compliance rules, for approvals, for refunds, for identity verification, basically, for example. So in Artemis, the reasoning layer handles interpretation, adaptive decision making, while deterministic orchestration governs execution, workflows, and business rule enforcement. But the combination allows enterprises to gain the flexibility of AI without sacrificing operational reliability, because that's what a CX leader needs. They want the benefits of AI, but the operational reliability and control. That's the dual-brain architecture in Artemis. It is very important because there's been a lot of hype by AI native startups that, like, hey, we we will do everything with reasoning and with prompts and you know, we'll leave it all to the LLMs. In fact, think about guardrails, right? You don't want the LLM to enforce the guardrails. You want to enforce that before you send something to the LLM and when you receive something back from the LLM, basically, because guardrails, by their very nature, basically, are your protective mechanism. And you want to enforce that in the application, in the platform, basically. So that's the reason for the dual-brain architecture. And Artemis is one of its type, and the only one of its type right now that implements it very, very cleverly, you know, for a large-scale secure CX deployment. Thank you. And I yeah, there's so much for uh our audience to unpack. Uh, for this has been awesome. Uh, we are running out of time, but I do before we before we kind of finish off, I want to um almost step step back a little bit just from core AI specifically just for a second, and kind of put our buyer's hat on or our um you know our CX leaders' hat on and consider what are the things that you'd recommend people validate uh when they're looking at multi-agent orchestration whilst it's in production, um, what metrics they should consider monitoring that are going to actually prove value rather than just you know activity and that it's doing some stuff. Um, you you you speak to organizations all the time, uh, and you must see this play out you know every week. So we can really benefit from your insights into that, I think. Yeah, I mean, listen, I think there are basically three elements that a CX leader should focus on. Number one, most important is failure recovery, observability, auditability, you know, uh compliance, adherence to the rules, the security rules, and the business rules. That's you know, non-negotiable. That has to, has to, has to happen. Otherwise, nothing will go live, nothing will stay live. Then you have the business metrics, you know, resolution rates, reduction of the customer effort, first time, you know, first call resolution, containment quality, not just quantity, the time to resolution, the revenue impact, and all of those business metrics which moves the which brings in the ROI. And then there's the soft metric, the customer experience, which actually, if you if the customer has a great experience, it propagates the brand. It propagates the brand in a soft way, which gives the soft feeling in the customer that, hey, this brand is taking care of me. So I'm going to stick with the brand, essentially, which may result in revenue impact as time goes on. So these are the three things, but those are your each one, each one is important, but the first one is the most important. The second one is why is the why you're doing it, and the third one is your design. Think of think it, think of it from the customer point of view, and then you will design it in such a way that they will love it as opposed to it's not so great, basically. So that's how I advise CX leaders to think about the metrics. That's um that's that's brilliant and super clear and very refreshing for uh someone in tech to bring us back to the customer, which which doesn't always happen. So kudos to you, Raj, for that. Um thank you. I could I could I could talk to you all day about this, but unfortunately that is this is all we've got time for, uh Raj. Thanks very much uh for joining me, answering all my questions. Uh really appreciate that. Before we do close, uh for anyone watching this who wants to explore this subject uh in any more detail, what's the best way for them to find out more or to get in touch with the team? Email me, raj at crow.com, go on our website, you know, you can engage with experts, look at demos, schedule a demo. We want all of you to understand and use Artemis because it will achieve your objectives. That's awesome. Awesome, thank you. And don't forget, um, you can also find uh a wealth of related resources, uh, stories uh and videos at cxtoday.com, just like this one. Uh, but that wraps things up for today. I've been Rob Wilkinson at CX Today. Thanks very much for joining us.