Since AI agents enter the real world deployment, organizations are under pressure to explain where they belong, how to develop them effectively, and how to run them on a scale. In Venture Bats Change 2025Tech leaders gathered to talk about how they are turning their business with agents: General partner Joan Chen in the Foundation Capital; VP of Project Management with Shailesh Nalwadi, Sandbird; SVP of AI change in thys wanders, coggygy; And Sean Malhotra, CTO, rocket companies.
Some high agent AI use cases
“The initial attraction of any of these deployment for AI agents is around the savings of human capital – mathematics is straightforward,” Nalawadi said. “However, it reduces your ability to change with AI agents.”
In the rocket, the AI agent has proved to be a powerful tool to increase the conversion of the website.
“We have noticed that with our agent -based experience, with the experience of conversation on the website, clients are likely to change three times more when it comes to this channel,” Malhotra said.
But it is just scratching the level. For example, a rocket engineer made an agent in just two days to automatically make a highly special task: Calcing transfer tax during mortgage undertake.
“This two -day efforts saved us $ 1 million a year,” Malhotra said. “In 2024, we saved more than one million team members, mostly from our AI solutions. This is not just saving expenses. It is also allowing members of our team to focus on people who are often the largest financial transactions of their lives.”
The agents are basically super charging the team’s individual members. This millions of hours of savings are not completely copied many times. These are different parts of the task that are not enjoying employees, or were not included in the value of the client. And this millions of hours savings give the rocket the ability to handle more business.
Malhotra added, “Some members of our team managed to handle 50 % more clients than before last year.” “This means that we can have more throptics, run more business, and once again, we can see the conversion rate high because they are spending time to understand the client’s needs, with a lot of work -working AI now.”
Deal with the agent’s complexity
“Part of our engineering teams’ journey is moving beyond the mentality of software engineering,” said Nilwadi. Not only software engineers, but also product managers and UX designers. “
The thing that has been helped is that LLMS has made a long journey, Wonders said. If they made something 18 months or two years ago, they really had to choose the right model, or the agent would not perform as expected. Now, they say, now we are at a stage where most mainstream models behave very well. They are more forecast. But today, combining these challenging models, ensuring the reaction, arranging the right models in the right order, and becoming the right data.
“We have users who move tens of millions of conversations every year,” said Wonders. “If you automatically talk, 30 million a year, how does the LLM do this in the world? These are all things that we have to discover simple things by getting the model with cloud providers.
Malhotra said that a layer of agents that organized the LLM is arranging a network. In the dialogue experience there is a network of agents under the hood, and the orchstrower is deciding which agent from the applicant to the application.
“If you play this ahead and think of hundreds or thousands of agents who are eligible for different things, you have to face really interesting technical problems,” he said. “This is becoming a major problem, because the delay and time of time. In the coming years, this agent will be a very interesting problem for resolving routing.”
Tapping into a vendor relationship
To this point, the first step for most companies is to build Agentic AI launcher at home, as there were no special tools yet. But you cannot create a difference and value through the construction of ordinary LLM infrastructure or AI infrastructure, and you need special skills to go beyond the initial construction, and to maintain the infrastructure, and to maintain the infrastructure, along with debugs, repetition and improvement.
“Often the most successful conversations with potential consumers are someone who has already built something at home,” said Nalawadi. “They quickly realize that it is okay to reach 1.0, but as the world develops and as the infrastructure develops, and when they need to change the technology technology of something new, they do not have the ability to set all these things.”
Agent AI complexity preparation
Theoretically, the Agent AI will only develop in complexity – the number of agents in an organization will increase, and they will start learning from each other, and the number of use issues will explode. How can organizations prepare for a challenge?
“This means that your system will put more pressure and balance in your system,” said Malhotra. “کسی ایسی چیز کے ل that جس کے پاس ریگولیٹری عمل ہے ، آپ کے پاس لوپ میں ایک انسان موجود ہے تاکہ یہ یقینی بنایا جاسکے کہ کوئی اس پر دستخط کر رہا ہے۔ اہم داخلی عمل یا ڈیٹا تک رسائی کے ل you ، کیا آپ کے پاس مشاہدہ ہے؟ کیا آپ کے پاس صحیح آگاہ اور نگرانی ہے تاکہ اگر کوئی غلط بات ہو تو ، آپ کو معلوم ہے کہ یہ غلط ہو رہا ہے ، اگر آپ کو یہ سمجھنا پڑتا ہے کہ آپ کو لوپ میں انسان کی ضرورت ہوتی ہے ، اور پھر ان پر اعتماد ہوتا You need a human being in the loop, you have to do it.
So how can you believe that when an AI agent is developed, he will behave reliable?
Nilwadi said, “If you haven’t thought about it initially, this part is really difficult.” “The short answer is, before you start building it, you should have an Evil Infrastructure. Make sure you have a tough environment in which you know that like an AI agent, and you have this test set. Keep pointing to it.
Wonders added that the problem is that it is non -biased. The unit test is inevitable, but the biggest challenge is that you don’t know what you do not know – what an agent can possibly do wrong, what can be reacted to any situation.
“You just imitate the scale conversation, pushing it forward under thousands of different scenarios, and then analyzing how it is intact and how it reacts,” said Wonders.