Experts Explain How Agentic AI Is Changing The Way The Collision Industry Functions

The automotive collision industry like so many other industires is already tapping into the power and efficiencies of artificial intelligence, but as agentic AI matures, industry experts expect the adoption of this new form of AI to continue ramping up and change the way the collision industry functions

The Collision Industry Electronic Commerce Association hosted a webinar on Jun 19 with two speakers with AI backgrounds who are keeping their eyes on how the emerging technology will impact the collision industry. Experts at the CIECA webinar, which is available on demand, aimed to balance excitement with a realistic overview of agentic AI’s capabilities—and emphasize the continued necessity of human involvement in customer service.

“From a customer experience part of it, we are going to see traditional, generative and agentic AI coming together to solve a myriad of challenges within our industry,” said Abhjeet Gulati, senior director of AI and machine learning at Mitchell International.

Unlike traditional and generative AI, agentic AI goes beyond supervised machine learning mechanisms and content creation from large language models. Gulati called agentic AI a “quantum leap” forward.

“Basically, these systems are autonomously planning and reasoning and taking coordinated action across multiple tools and systems,” he said, describing the technology’s ability to take specific actions that were previously required to be undertaken by humans.

This development not only increases automation—it fundamentally reimagines how cluttered workflows are processed. Agentic AI is all about taking independent actions to accomplish complex steps, according to him.

Information retrieval gaps, time-consuming data collection and inconsistent and erroneous decisions in assignment all amplify claims life cycle delays. Traditional AI can automatically detect damage. Generative AI can explain damage estimates and use natural language interfaces during the claims submission process, Gulati said.

“Agentic AI takes it even further,” Gulati added. “Now you have different, multi-agents. Conversational agents. Informational retrieval agents. They’re all gathering information at each touch point and … and then they’re adapting to the customer responses.”

In other words, agentic AI is task-based and capable of autonomously executing multi-step processes. In regulated spaces, Gulati said the intent is always to keep a human being in the loop.

How Can Agentic AI be Used in The Collision Industry?

“It can coordinate complex, actionable activities between the policyholder, repairers, adjusters, insurers and other stakeholders,” Gulati said. “It’s kind of an orchestration layer. And then, it can also maintain audit trails for all the actions and decisions that were taken to satisfy some of the peer requirements, including the regulation side of it.”

Moving forward, he believes that computer vision—the backbone of traditional AI—will continue to define the classification of damages in the collision claims process, but he also sees multi-agentic agents entering the process and executing specific tasks. Gulati believes that from FNOL to settlement, information agents can serve as retrieval bots that are capable of gathering necessary documentation, validating coverage and providing support for the entire claims journey.

Meanwhile, scheduling agents can act as repair coordinators during the full claims process, Gulati said. Repair shops, parts suppliers and customers can all work with them to order parts and schedule work that needs to be completed as part of a claim. AI agents for fraud, waste and abuse can “take a keen viewpoint of where there are opportunities to look at a particular claim and flag a suspicious activity,” he said.

However within all of this, a human being would be required to review the work the agents perform before they take action, he added.

“The hype is relevant,” he said, but “it remains to be seen how the orchestration will come about. One thing we have to understand is that the human-in-the-loop concept is going to be a key differentiation for us, especially within the regulated industries like insurtech, fintech or healthtech.”

In follow-up correspondence, Gulati explained that that many organizations are either exploring or implementing agentic AI in the collision claims process. One area of focus is the use of these autonomous decision-making models to gather information at FNOL that drives actionable tasks. For example, an agentic task can retrieve all relevant recall information for a particular VIN. If the vehicle is repairable, another agentic agent can take over to manage scheduling, followed by parts procurement and delivery.

“With just one input, multiple, sequential tasks are performed,” he explained. “These tasks require minimal human oversight and allow organizations to deliver timely customer updates while still maintaining empathy throughout the claims process.”

A Case Study

Gaurav Mendiratta shared a real-world story of a large auto parts provider with a problem.

The provider was missing roughly 100,000 calls per week across their 4,000 U.S. stores. Overall, 80% of their callers asked the same two questions: “Is this part in stock? How much is it?” The solution: a generative AI-powered voice assistant.

“The AI finds the part and places the order all during the call,” said Mendiratta, who is the CEO of SocioSquares, an AI software development and online marketing firm, and the chief product officer at Propel, a SaaS company. “Is it perfect? Nowhere close. I would say currently [it’s] between 90 to 92% accuracy. Getting an AI agent to be 100% reliable will be a work in progress for many, many months—if not years.”

In addition to the human in the loop, as agentic-based collision solutions are built, decision transparency, explainability, consistency and feedback integration will be key aspects of building consumer trust in the technology, Gulati said.

AI agents will have access to data systems across multiple databases, and Gulati said checks to safeguard sensitive information will be necessary, as will regular bias audits, decision-pattern monitoring and edge case identification.