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

Companies hurry to produce AI agents – many will fail. However, the reason has nothing to do with AI models.

Two of the two VB Transform 2025Industry leaders shared a tougher classes from accommodating AI agents on a scale. Joanne Chen, a panel managed by a general partner Stock capital, shawn malhotra, co Rocket companiesAgents using agents on the homeowner travel journey from the mortgage entry into the customer’s conversation; Shailesh Nalawadi, Head of Product BirdKoran agent customer service practices for companies with many vertical; and THYS WAANDERS, AI Transformation Svp ModestThe platform automates customer experience for large enterprise contact centers.

Their shared discoveries: Companies that create evaluation and orchestra infrastructure are successful, and those who fled with strong models are failed on the scale.

>> – siSee all transforms 2025 coverage of our entire transform<

ROI Reality: Out of the cutting of simple costs

The main part of engineering You have an agent To understand the return of investment (ROI) for success. Early AI Agent Placement is aimed at reducing costs. Although this is the main component, enterprise leaders now express more complex ROI samples that require different technical architectures.

Reducing costs wins

Malhotra shared the most dramatic expense sample of missile companies. “We had an engineer [who] The mortgage underwriting of the approximately two-day work process was able to solve the problem as a multi-niche called ‘transfer tax calculations’. And he spent a million dollars in this two-day effort, “he said.

For Cognny, Waanders noted that every rich value is the main metric. He said that if AI agents It is used to automate the parts of calls, it is possible to reduce the average processing time for each call.

Revenue generation methods

Saving is something; It is another thing to bring more income. Malhotra saw the conversion development: Customers are becoming higher degrees because they have a faster response to their questions and have good experience.

Proactive income opportunities

Nalawadi stressed completely new income opportunities through proactive broadcasting. His team allows active customer service before understanding whether there are problems with customers.

This perfectly describes this example of a food delivery. “They have already realized that an order will be late and expecting to be late and waiting for the client to be sad and calling.”

Why AI agents are broken in production

Although there are strong ROI opportunities for enterprises placing Agentic AI, there are some problems in production areas.

Nalawadi, the main technical deficiency determined: companies build AI agents without evaluation infrastructure.

“Even before it started to build it, there must be the infrastructure of the earth,” Nalawadi said. “We all were software engineers. No one places for production without a single test. I think this is a very simple way about the assessment this is a single test for your AI agent system.”

Traditional software test approaches do not work for AI agents. He noted that it is impossible for any possible access to the forecasting or vital testing of natural language. Nalawadi’s team learned it through customer service placements for retail, food delivery and financial services. Standard quality guarantee approaches released external circumstances.

AI TEST AI: New quality assurance paradigm

Given the complexity of the AI ​​test, what do organizations do? Waanders solved the trial problem through simulation.

“There is a feature that we have left soon, that this is to imitate potential conversations,” he said. “So testing AI agents is essentially AI agents.”

The test is not only a test of conversation quality, but the scale is an analysis of behavior. Can the agent help understand how angry customers respond? How is it managed in more than one language? What happens when customers use slang?

“The biggest problem, you don’t know what you don’t know.” “How do everyone react to anything else that he can come? You find the conversations by simulating the conversations only by pushing them under the different scenarios.”

The approach is demographic changes, emotional states and emotional states and emotional states and outsider.

Incoming complexity explosion

Current AI agents manage single assignments independently. The heads of enterprise must prepare for a different reality: hundreds of agents to the learning organization from each other.

The results of infrastructure are massively. When agents share information and collaborate, failure mode is exponentially multiplied. Traditional monitoring systems are unable to monitor this interaction.

Companies must now be architects for complexity. The retracted infrastructure for many agent systems is much higher than the start.

“If you put forward the theoretically possible, they may be in hundreds of organizations and maybe learn from each other.” “The number of things that can explode again. The complexity explodes.”