Why building a good AI agent is harder than you think

A decade ago, when you watched a lonely Theodore fall in love with a faceless AI assistant, Samantha, in the movie Her, you wouldn’t have had many questions on how she became a 'good AI agent'. However, in 2025, as AI agents become pervasive and impossible to avoid in day-to-day life, questions about their intentions and functions are bound to be raised.

So, what is a good AI agent and what goes into making one? This was a key topic that Anand Jain, Co-founder, CleverTap and Shivakumar Ganesan, Co-founder and CEO, Exotel discussed in a panel discussion at TechSparks 2025, YourStory’s flagship startup and innovation summit in Bengaluru. 

For Jain, a good AI agent experience is a seamless one. "It just feels natural and responsive. That’s what a good experience looks like," he said.

But a lot goes into creating that experience, particularly good data. “Imagine you’re chatting with a brand’s support agent about a refund that hasn’t come through. The conversation feels seamless, but what’s happening in the background is far more complex. The brand’s systems are pulling together data about your past transactions, refund policies, payment status, and even interactions with the bank—stitching all of it together to give you a smooth, human-like experience,” Jain added. 

A bad AI agent, on the other hand, is one that ignores brand, budget, or regulatory guidelines and goes rogue. “That’s when users immediately feel like they’re talking to a wall. We’ve all been there, stuck in endless loops trying to reschedule a flight or fix a payment issue and desperately hitting “talk to a human” because the bot just doesn’t get it,” he explained.

Jain’s company, CleverTap, is a customer engagement and retention platform used by global brands to analyse user behaviour and build personalised communication. 

Meanwhile, Ganesan talked about how over-prompting and treating AI like software often spoils the experience.

“Engineers often ‘over-prompt' and test AI systems the same way they would test software. In the process, they end up restricting what the model can do, turning a smart system into something that behaves like the old rule-based chatbots we were trying to leave behind. You ask it a simple question like, 'What’s the weather like?', and it replies, 'Sorry, I don’t know what you mean.' We end up killing the intelligence we’re trying to build,” he shrugged. 

Also Read

AI in healthtech: Why a fine balance between innovation and data privacy is crucial

“At the core, though, it’s about how we think. If we keep approaching AI with a rigid, software-engineering mindset, we’re not going to get very far,” he added.

Exotel is a cloud communication platform that provides voice and messaging solutions to enterprises.

However, even with the explosion of AI agents, Ganesan noted that India is yet to fully figure out agentic AI, and that’s where the real opportunity lies for startups. “There are already at least four US companies building voice agents with annual revenues exceeding $100 million each,” Ganesan said.

Talking about the future of AI agents, Jain envisioned a zoo of agents interacting and collaborating seamlessly.

“Think of it this way: OpenAI or similar models will continue to advance and become incredibly capable at a general level. But when it comes to specifics, they won’t automatically have that knowledge. That’s where enterprise-level agents will come in. These agents will be embedded within your business systems, gathering contextual information and sharing it with one another,” he said, adding, “For instance, if one agent is running a promotion, it will need to coordinate with another agent managing inventory to ensure there’s enough stock to fulfil orders. This kind of dynamic, context-aware collaboration is where the real power of agentic AI will emerge.”