A CIECA webinar held Thursday outlined different types of AI, how they are built, and how they could be used in the collision industry. 

Gaurav “Rav” Mendiratta, CEO of SocioSquares and chief product officer at Propel led the discussion. 

“As per a recent survey by GoDaddy, they estimate that a business owner or a professional can save more than $4,000 and 300 hours of work in a year by using GenAI tools,” Mendiratta said. 

He said another Statista study showed business owners increasing revenue by 6-10% by adapting GenAI. A third study from Accenture found GenAI will drive 95% of customer experience by the end of 2025. 

Mendriatta shared a video of two GPT-4o devices speaking to each other about the replacement of a customer’s phone. 

“I think you’ll agree that GPT-4o is very close to fielding customer calls,” Mendriatta said. 

ChatGPT runs on a Foundational Large Language Model (LLM) called Generative Pre-Trained Transformer, Mendriatta said.

Mendriatta said all human written text is estimated at about 10 trillion words. He said GPT is expected to be trained on 10 trillion words within a year. 

However, ChatGPT does have limitations, he said. This includes hallucinations. 

For example, a New York lawyer asked ChatGPT for a case that set a precedent, Mendriatta said. He said the lawyer quoted the information in a case to find out later that ChatGPT made up the information. 

“He was called out on that and then penalized,” Mendriatta said. “So always check when you’re asking ChatGPT for facts.” 

He said it also sometimes lacks common sense and has limited long-term memory. 

Other popular LLMs include:

• • Llama 2.0, open-sourced and introduced by Meta
  • Claude, created by Anthropic which claims to have extra guardrails against bias and hallucinations 
  • Mistral, a French company founded by ex-Meta and Google Deepmind employees
  • Gemini by Google

Mendriatta played an example of a vehicle owner asking ChatGPt what they should do after a car crash. 

Steps given by ChatGPT included: 

• • Checking for injuries and calling emergency services if any are found
  • Moving the vehicle to a safe location and turning on hazard lights
  • Calling police to report the accident, even for minor crashes
  • Exchanging information with other drivers including names, addresses, phone numbers, and insurance and vehicle information 
  • Documenting the scene with photos including the vehicles, any damage, and the location
  • Calling insurance providers to inform them about the crash

“This is today but I imagine in the near future AI virtual assistants would be installed in cars and will offer real-time advice and reassurance helping to reduce stress and ensure that critical information is accurately recorded,” Mendriatta said. 

A second recording showed how ChatGPT responds to a vehicle owner’s question about a cracked window. 

ChatGPT suggested small cracks or chips less than 6 inches could be repaired quickly. It said glass with larger cracks or cracks in the line of sight should be replaced. 

Mendriatta said ChatGPT will soon be likely to offer repair or replacement suggestions based on images of cracks. 

He said AI programs that insurance and car rental companies can use to estimate damage include: 

• • Altamira.ai
  • Tchek.ai
  • Tractable.ai

He said AI can also create fast claims management and support along with providing automated underwriting. 

Mendriatta said AI tools that he recommends for any business include: 

• • Fathom.Video Notetaker, a free notetaker. A paid premium version creates summaries with action points. It also records meetings with video. 
  • Perplexity AI finds recent news and research and integrates other AI models. 
  • Gamma AI creates slides with a single prompt 
  • ChatGPT which can be customized and trained to specific information 

An auto parts retailer recently customized ChatGPT to help field and direct calls, Mendriatta said. 

He said the company was missing 100,000 customer calls per week. About 80% of the calls asked for items in stock or pricing information. 

It would take traditional AI or answering services hours to assist a customer looking for a particular part for a certain make or model. 

Gen AI is able to understand the make and model of a vehicle and look for specific parts, he said. 

Mendriatta said AI companies often provide specifications that can be compared. 

For example, an estimating tool might claim they’ve fine-tuned their AI on 5 million photos of damaged cars. Another AI company may claim to have fine-tuned their AI on 15 million photos of damaged cars. He said the more photos the AI has learned, likely the better it is. 

Another way to compare products would be to provide each estimating AI with a very complex image and see which AI does a better job.

IMAGES

Gaurav “Rav” Mendiratta discusses AI during a May 30, 2024 CIECA webinar (Webinar screenshot)