Lessons learned from agentic AI leaders reveal critical deployment strategies for enterprises

Companies accelerate artificial intelligence agents to production – and many of them will fail. But the reason has nothing to do with artificial intelligence models.

On the second day of VB Transform 2025, industrial leaders have exchanged lessons they achieved with difficulty publishing artificial intelligence agents on a large scale. A committee managed by Joan Chen, the general partner of the Capital Foundation, Champhehtra, included CTO in Rocket Companies, which uses agents through the home ownership trip from real estate subscription to customer chat; Shailesh Nalawadi, head of the producer at Sendbird, who builds customer service experiences for companies across many sectors; Thys Waanders, SVP from the transformation of artificial intelligence in CGNIGY, whose platform automates customer experiences of communication centers with large institutions.

Their joint discovery: companies that build evaluation and infrastructure for coordination first successful, while those that rush to production fail with strong models on a large scale.

>> Watch each coverage of our conversion 2025 here <

The reality of the return on investment: beyond reducing simple costs

The main part of the engineering AI agent for success is an understanding of the investment (ROI). The publication of the artificial intelligence agent focused early on the cost reduction. Although this remains a major component, institution leaders now report the most complex investment patterns that require different technical structures.

Victory of cost reduction

Malhotra shared an example of the most dramatic cost of missile companies. “We had an engineer [who] In about two days of work, I managed to build a simple agent to deal with a very specialized problem called “transfer tax accounts” in the real estate subscription part of the process. This two days of effort saved a million dollars a year at an expense.

For CGNIGY, Waanders note that the cost of each call is a major scale. He said that if artificial intelligence agents are used to automate parts of these calls, it is possible to reduce the average hand of the handling of each call.

Revenue generation methods

Savings one thing. Achieving more revenues is another. Malhotra stated that his team has witnessed the transformation improvements: since customers get answers to their questions faster and have a good experience, they are turning at higher rates.

Continuous revenue opportunities

Nalawi is completely highlighted by the new revenue capabilities through proactive awareness. His team enables pre -emptive customer service, even before customers realize that they have a problem.

It explains the example of this food connecting completely. He said: “They already know when the request is delayed, and instead of waiting for the customer to be annoyed and contacted, they realize that there is an opportunity to move forward.”

Why do artificial intelligence agents break production

While there are strong opportunities for the return on investing for institutions that publish AIC, there are also some challenges in spreading production.

Nalawadi has identified basic technical failure: companies build artificial intelligence agents without infrastructure assessment.

“Before you start building it, you must have an infrastructure in place,” Nalwadi said. “We are all used to be software engineers.

Traditional software testing methods do not work with artificial intelligence agents. He pointed out that it is not possible to predict all possible inputs or write comprehensive tests for natural language interactions. The Nalawadi team learns it through retail customer service and delivery of food and financial services. Standard Quality Assurance Methods lost the edge cases that appeared in production.

Artificial Intelligence Test: New Quality Assurance Form

Given the complexity of the artificial intelligence test, what should organizations do? Waanders solve the test problem through simulation.

“We have a feature that we soon launch about the potential conversations simulating,” explained Wandars. “Therefore, artificial intelligence agents test artificial intelligence agents.”

The test is not just a conversation quality test, it is a widespread behavioral analysis. Can it help understand how the agent responds to angry customers? How do you deal with multiple languages? What happens when colloquial customers use?

And Warsers said: “The biggest challenge is that you do not know what you do not know.” “How does it interact with anything that anyone can reach? You just find it by simulating conversations, by really pushing them under thousands of different scenarios.”

The approach tests the demographic differences, emotional states and edge situations that cannot be covered by human quality guarantee teams comprehensively.

The upcoming complex explosion

Perfect artificial intelligence agents deal with one tasks independently. Foundation leaders need to prepare for a different reality: hundreds of agents for each organization learn from each other.

The effects of the infrastructure are huge. When agents share data and cooperation, failure situations are dramatically doubled. Traditional monitoring systems cannot be tracked.

Architectural engineering companies must now complicate. The costs of updating the infrastructure of multi -agents are much more properly built from the beginning.

“If you are rapidly progressing in what is possible in theory, there may be hundreds of them in an institution, and they may learn from each other,” Chen said. “The number of things that can occur just explode. The complexity explodes.”