Can Enterprises Trust AI to Do the Work, Not Just the Demo?

Rhys Fisher: Hello and welcome to CX Today. I'm Rhys Fisher, Associate Editor, and today I'm delighted to be joined by Gaurav Anand, Senior Vice President and Head of Customer Interaction Suite at Tata Communications. Gaurav, how are you doing today? Gaurav Anand: I'm doing great, Rhys. Thanks for having me. I'm looking forward to our conversation. Rhys Fisher: Same here. Thank you very much for joining us. I think it's going to be a really fun chat. We're discussing AI, which everyone is right now, but we're taking a more interesting approach: the direction of travel for AI in the customer service and experience space. We know AI is moving beyond pilots, copilots and standard automation. What enterprise buyers really want to know now is whether AI can truly be trusted to complete a meaningful outcome, whether that's qualifying a lead, securing an appointment or retaining a customer. So, Gaurav, in your experience, when do you know an AI pilot is ready for the real world, ready to really deliver, rather than just looking lovely and shiny in a demo? Gaurav Anand: It's a great question, Rhys. Before we jump into pilot readiness, it's important to step back and understand how we got here. It all started with traditional customer experience. Everybody had that top of mind: you wanted to make experiences as unfragmented as possible and have as much context on the customer as you could. Then we evolved to making things more self-service and more conversational. Then we brought in AI agents, which weren't just bots or the deterministic IVRs people used to deal with. Now those AI agents and that whole generative, conversational trend are moving into what we're calling the AI workforce. That's really where agents are starting, as you said, to deliver outcomes, and that's where pilots become important. A lot of companies and brands are running AI pilots in silos, trying different models. Where those pilots tend to fail is where they're missing context, orchestration and trusted, governed execution. That's really what it comes down to. When we're talking about pilots, it isn't about models. It's about looking at the different areas you want to cover. You want to look at customer readiness and operational readiness. You need to make sure governance is in place. And you need to be clear on what business outcomes you're trying to test and validate. A successful demo obviously isn't enough, so these four gates are extremely important, and each one must have a level of sophistication. We can set targets, we can set how we track towards them, and we can set how the different integrations come into place. But at the end of the day, what really determines AI pilot readiness in the real world is how easy it is to deploy, how easy it is to repeat and how easy it is to scale. With all four of those areas covered for each instance, you want to build some kind of infrastructure, or operating system, that can proliferate across the organization so you're not dealing with different silos. Rhys Fisher: Thanks for that. That's a really nice overview of how and why you need to differentiate between demo-ready pilots and solutions that are actually delivering outcomes. You alluded to this a little there, and it's something Tata Communications has spoken about in the past: the idea of touchless resolution. What does that look like in practice, and which customer or revenue journeys are best suited to it? Gaurav Anand: That's a good segue from the first question. You're running these pilots, and you want them to go into production. You want them to scale, to be repeatable, and to sit in different parts of the organization that communicate with each other. Touchless resolution isn't just about containment. It's about verified completion, or the verified delivery of an outcome. The enterprise objective must not be to run the pilot for the sake of running the pilot. With touchless resolution, the enterprise objective is achieved end to end. The relevant system of record is updated and the outcome is confirmed without requiring a human to complete any hidden work afterwards. There can be oversight, guardrails and an audit trail, and the system running these pilots should be able to execute on all of that. Security is also extremely important throughout. There's a fear of giving up too much control. But touchless resolution means moving from an AI that answers questions or suggests solutions to an AI that acts autonomously, within the guardrails you put in place. And it isn't a one-off thing, or some kind of nirvana. Certain tasks lend themselves more to autonomous touchless resolution. Some may require a joint effort. Some may require multiple agents working together to reach the end goal. So it takes various forms. Largely, touchless resolution suits high-volume, repeatable tasks with clear rules, where the data is easily accessible across different platforms and systems, and where verified execution and completion are achievable. In my mind, that sits in the middle of the vast array of tasks and outcomes enterprises are looking for. Rhys Fisher: You've referenced tasks better suited to AI versus tasks better suited to humans. If I'm a CX or contact center leader, I'm probably hearing this a lot. It's easy to say AI should handle some things and humans should still be involved, but how do you actually decide what AI should handle itself and when humans need to step in? Gaurav Anand: That's a great question, and as you said, it's top of everybody's mind today. When you're deploying AI to achieve end outcomes, as we said, it's not about the models, and it's not about being fully autonomous or giving away control. One thing we have to be absolutely clear on is that the unit of design should be the decision or the task, not the job. The question isn't whether a role can be automated, but which decisions within that role can be automated safely and compliantly. Where the two start to work together is where AI provides insight-based assistance, a digital assistant, to the human. Take a step back and think about how an enterprise is set up. Companies spend millions of dollars setting up their data lakes. They want to capture every touchpoint with customers, suppliers, partners and employees. The problem is that it all sits in different systems, sometimes with different vendors. I'll give you an example. Just yesterday, I was booking international travel for six of us in my family through a travel agency. I won't name them, to maintain their privacy. I talked to a human agent, who set up my itinerary. It was fairly complex, with multiple cities. At the end of the session, I said I didn't want to purchase the tickets yet, because I wanted to explore other options and talk to my family. I asked whether I could access what they'd just offered me online for reference later. They said no. I could have done exactly the same thing through their online system, entering all the information without a human agent at all, and I'd probably have had similar options. But because I'd already spoken to an agent, I couldn't go online and adjust things myself. This is a large multinational serving global requirements, and I'm a premier customer, yet they just can't connect those two systems. That's really interesting, because I'm sure both of those platforms, the human-assisted one and the online one, are using AI. They have knowledge bases, and their AI is suggesting the next best action. But it's still very siloed. What I'm getting at is that there's a data layer, with different systems sitting across the globe, and this is a very simple customer interaction. In the back end there are ticketing systems, inventory, pricing and billing, all systems that don't talk to each other. Then on top you have applications pushing out to customers. What we need is a layer in the middle, between the data layer and the application layer, that connects the two. It's like an operating system, and I think that's the answer to your question about how AI and humans work together. The AI operating system can connect these different data points, bring them together and create context at an individual level, whatever that unit might be: a customer, a supplier, a machine or a network element. It can then feed that to an AI agent acting autonomously, to multiple agents interacting with each other agent to agent, or to humans in the loop. I apologize for going a little long here, but to put a cap on it: humans and AI work together in so many different ways, and this operating layer enables that interaction to happen in far more sophisticated ways, driven by its insights. It compounds outcomes and complements what already exists, completing the stack rather than competing with it. That's how human and AI interaction comes together to create an outcome far superior to any individual model or pilot. It brings the whole ecosystem together. Rhys Fisher: Absolutely no need to apologize. It's always great to hear real-life examples, because they make it much easier to picture. That was a great example of a common issue: the lack of connection, the silos, that prevent this technology from really maximizing things for the customer. It leads into another common pain point. We've all heard customers complain, and we've probably all experienced it. You go through an AI agent, provide all the details it asks for, and when you're finally handed over to a human agent, they ask for the same details again. That's a real frustration. We hear a lot about seamless handoffs, but in reality they often aren't as seamless as they're made out to be. How do companies make that handoff genuinely seamless? Gaurav Anand: Look, it's an age-old pet peeve. Everybody hates it. I've had to do it, you've had to do it, and it's incredibly annoying. A lot of it comes back to what I said before: these are different systems, largely from different vendors. I'd even go as far as to say it's a lack of AI readiness, and perhaps the lack of the right integration partner or solution, that's at the heart of the problem. Say a call comes in and you have voice AI. Getting a voice AI agent to take that call already requires three levels of integration. A PSTN call hits a switch or an SBC somewhere, and that SBC then needs to set up WSS streaming so the AI agent can start interacting. Throughout the interaction, the AI agent needs assistance from the data and the context. I'll keep coming back to this context layer and the context graph, which we'll talk about in a minute. If it's a premium customer calling, you should first be able to manage whose calls go to AI agents and whose go straight to human agents. The system should already know the customer's preference. Some people will say they don't want to talk to an AI agent. Others won't want to talk to a human. Others might prefer to switch channel to text or WhatsApp. Whatever the channel, that AI assistance needs to be live and available to the AI agent. And if the call needs to transfer to a human agent, all the integration from that first leg has to transition the full context of what was discussed with the AI agent across to the human agent. The human agent then continues with access to the knowledge base, SOPs and any other intel the customer might need, which at times may sit across 10 or 15 different applications. To have access to all of that and transfer the context over, you need an integrated solution, number one. It has to manage channels, endpoints, transport and security guardrails, plus all the intelligence that needs to be processed and passed on, the next best action suggestions, and any live inventory or other searches. There could also be a scenario where, after the human agent resolves an escalation, the call transfers back to the AI agent. Take an airline example. I call in to make a change, and I'm not happy with the options the AI agent is giving me, or it can't offer a certain discount or seat choice. I talk to the human agent, who takes care of it and sends me back to the voice agent. The voice agent then concludes the transaction: here's your text message, click on this to make your payment, and so on. So there are a number of things at play, and a number of systems that need to integrate because there are so many touchpoints. There's the AI operating system I talked about, which brings all the data and context together and creates what we call a context graph across all the different systems. It provides insights and live information back to the agents, whether voice AI or human, and it should be truly seamless. All of those things have to come together. I'll take another 10 seconds on one more important point. If you talk to contact center folks, human agents can spend up to 18 to 21 minutes or more after a call summarizing it and updating all the back-end systems. They go into the CRM, perhaps log a credit that was given, and so on. Now, with all this information and multiple agent-to-agent conversations happening, the human agent's time is freed up to be more productive: checking that everything was done correctly, making sure the audit trail is there and closing things out efficiently, rather than getting into the minutiae of taking notes. Last but not least, there's post-call analytics: what you learn from the calls that have happened, which feeds into agent training and so on. That feeds back into how the handoff should happen, including what information should be transferred and the context, intent and sentiment behind it. This operating system learns from every new data point, every new conversation, every new resolution and every outcome achieved, and it gets better and better. So handoffs, whether between AI agents or from AI to human, become more and more efficient. That's why this operating layer is so important. It brings all the systems together and offers them up as insightful data points that enable seamless handoffs. Actually, I wouldn't even call it a handoff in this case. It's seamless working together, or seamless collaboration, among different agents, human and AI. Rhys Fisher: Thanks, Gaurav. I always find it really enjoyable and interesting listening to people like you who are clearly knowledgeable and in the weeds with this technology. It really emphasizes that in many cases we're only scratching the surface of what it can do. If an organization wants to move beyond some of the more basic AI features and tools and experiment with what it can do, where should it start? And how can it scale what works without trying to change everything at once and risk breaking its existing AI infrastructure? Gaurav Anand: That's another really good question, and it's top of mind for a lot of enterprises right now. They all want to scale and go beyond experimentation. A successful AI program is going to be built around repeatable operating models, not individual use cases. I'd say you need a partner who can provide an integrated solution, like we just talked about. A partner who can run the AI operating system that sits between the data layer, the applications you're running and the AI agents you want to deploy, orchestrating between them within the guardrails of security, compliance and privacy. Privacy comes up a lot, as does role-based access. So you really need to find the right partner and the right operating model. It's also about cross-functional teams spanning all the different departments, drawing on the same set of data and the same set of insights: the context graph we've been talking about. And it isn't backbreaking. For instance, we're working with a company that evaluates 425 gigabytes of data and transforms it into consumable insights that can be spread across the organization. So you can run multiple AI pilots that don't sit in silos and that interface with each other. To answer the question directly: start with one or two high-value, measurable business outcomes. People usually talk about the left-hand side and the right-hand side. The left-hand side is business acquisition, and the right-hand side is business retention, such as service and loyalty. Pick what you want to work on, define the outcomes you want to achieve, and set a baseline to start from. Then build the security and guardrails around governance, auditability and privacy. And have a partner who can facilitate that in a co-created way, rather than forcing a particular approach on you. There's a lot of complexity and nuance, and a lot of individuality across different verticals and enterprises, that needs to be taken into consideration when building this context graph. So start with a repeatable operating model. Figure out the outcomes you're trying to achieve, pick the elements with the high-value returns you're looking for, and then find a way to scale that and replicate it in a way that's governable and secure. Rhys Fisher: That last point about the complexity of the technology, and making sure the vendor relationship is a really good fit, is a very important one. AI has been around for quite a while, but this latest iteration is relatively new and, as you said, there's a lot of complexity to it. That's why it's so helpful speaking to people with that expertise, who can pick through the complexities and really unpack them. I think the audience is going to get a great deal from this conversation. Thank you very much for your time, Gaurav. Gaurav Anand: Absolutely. Thank you for having me, and I look forward to talking to you again soon. Rhys Fisher: Absolutely. I'd also like to thank our audience for tuning in. If you enjoyed this chat, please remember to like and subscribe to the channel, and head over to cxtoday.com for more stories like these. Until next time, thanks for watching.