AI: First overrated, now underestimated
Readingtime: 9 min
Exclusive The initial enthusiasm among insurers for artificial intelligence (AI) has faded. But Arndt Gossmann, founder and CEO of AI start-up DGTAL, is certain: important technologies are initially overestimated, but then underestimated in terms of their long-term impact. This also applies to AI, says Gossmann in an interview with Versicherungsmonitor. He explains the DGTAL business model and describes how the introduction of AI methods works – and how it fails.
The start-up DGTAL has some prominent customers. The claims settlement specialist Darag uses AI applications built by DGTAL [1] to support the valuation of claims and accident portfolios that it intends to take over. The reinsurer Scor uses the artificial intelligence provided by the start-up [2] to prepare and process claims for its experts.
This is reported by founder and CEO Arndt Gossmann. “A broker in Greece uses the software for the entire claims processing workflow,” he adds. The broker in question is NN Hellas. “A US insurer, whose name I cannot disclose, is using our system for triage.” This is scheduled to go live at the end of the year.
“Like most insurers, the insurer conducts an annual claims audit,” explains Gossmann. “Employees ask themselves which cases are actually urgent and what special circumstances exist.” If a claim touches on New York State labor law, it goes straight to specialized lawyers. DGTAL is supposed to help prepare these decisions.
Gossmann, 55, knows his way around the industry. He advised insurers at the auditing and consulting firm KPMG, then headed Darag as CEO, and later took over and wound up the transport insurer Sovag. He then founded Gossmann & Cie. as specialists in the run-off of current motor vehicle portfolios. “The pandemic put an end to the business model,” reports Gossmann. It was a stroke of luck that he got to know a Swiss-Danish IT company. That's where his partners at DGTAL came from.
“We build generative AI for insurers,” says Gossmann, explaining the business purpose. “Our main starting point is the fact that insurers work predominantly with unstructured data.” Only 2 to 5 percent of the data is in the systems and thus available in a structured and analyzable form. “Only then is it AI-compatible.” Around 95 percent of the data is bound in PDFs or other documents. “Then it is available, but you always need people to be able to do anything with it.”
DGTAL has developed methods to make this unstructured data machine-readable. “We then build use cases on this basis,” reports Gossmann.
One of these is the satellite perspective. “This allows us to evaluate large volumes of files or case files, for example to identify the cost drivers in the portfolio or the damage drivers.”
AI reads claims files
Gossmann: “In the second perspective, let's call it the helicopter perspective, AI helps with the investigation of the relevant files.” DGTAL has just introduced the system at Scor. “Claims files are read in, which in special cases are sent from the primary insurer to the reinsurer.” These are usually 500 to 600 pages long. “They are summarized by software, which identifies 20 key questions that the claims adjuster has, and a draft for the internal assessment paper is also added.”
This saves the claims adjuster several weeks of work. “The assessment is not carried out by our AI,” adds Gossmann. “These are not decision-making systems, but decision-making aids.”
DGTAL has 15 employees. Annual sales are still below one million, but Gossmann is optimistic that they will quadruple by 2024. The start-up is making a loss, but has enough money for further development. External investors include VGH Versicherung, Hanover, which holds 7.5 percent.
Four customers, seven additional pilot projects
DGTAL applications are in production or will go live soon at four customers, and there are seven pilot projects. “We also accept funding for pilot projects,” reports Gossmann.
How can a small company like DGTAL make a difference? The IT departments of large insurers have several thousand employees. What can DGTAL do that the experts at Allianz or Axa cannot?
“It's mainly about time and proximity to the user,” he replies. “Anchoring AI at insurers does not work through a central introduction or central purchasing.” A scalable technological platform is needed for this. “The implementation of use cases must take place close to and with future users.” It makes no sense to buy AI if the claims expert or underwriter isn't sitting next to you and telling you what they need.
It took DGTAL more than two years to build its platform. “But now it only takes two weeks to put the use cases into production on the platform – provided that the people who work with it are actually sitting next to it.”
DGTAL has AI agents in use
DGTAL has now further developed its applications in the direction of AIA, or Artificial Intelligence Agents [3]. The technological approach is not insurance-specific. “Imagine a
factory floor where we have set up one AI machine per customer,” explains Gossmann. “It was built for the customer, albeit with essential elements that we already had.”
In the new world, the hall should be imagined as not having one machine, but ten to twelve people working, each with their own workbench. Each AI agent is a software element. “The AI agents can communicate with each other and with the human user and even make decisions within a procedural framework.”
The European insurance industry receives around 5 billion documents annually in the property and casualty sectors alone, 3.5 billion of which are related to claims. “All of this has to be read, and that's where our AI agents help.” An agent evaluates an incoming email. In doing so, it determines that an Excel file is attached. “This AI agent is not equipped for this, so it searches the workshop for the AI agent that does nothing but evaluate Excel files.”
AI agents learn independently and try out what works
The AI agents take care of the coordination themselves. “It's no longer rule-based,” says Gossmann. “The AI agents can learn independently and try out what works.” The result: higher quality. “And it leads to traceability,” adds Gossmann. “One of the criticisms of AI is that it is a black box. That is no longer the case.” This is because the AI agents can log and explain every step they have taken.
This enables insurers to meet an important requirement of the supervisory authority. BaFin does not want systems where it is not possible to trace why a particular decision was made.
DGTAL started out with large language models such as ChatGPT. “But they hallucinate too much,” reports Gossmann. "With them, you can achieve reliability of 70 percent to just under 80 percent. " That's not enough for an insurer.
Positive experiences with large language models
“We have had positive experiences with small large language models,” he says. “There are many LLMs from universities and other institutions.” In the hands of the right AI agent, reliability can exceed 90 percent.
In the insurance industry, the initial enthusiasm that flared up with the release of ChatGPT has now faded. Many expectations have been disappointed. “This doesn't just apply to insurers,” says Gossmann. “When a significant innovation appears in this world, it is initially completely overestimated in terms of what it can do. And then its long-term impact is significantly underestimated.” This is the case for some insurers. Some do not understand how they can successfully approach the problem of digitalization and AI. “There is one company that has set up a central purchasing department for AI,” he says. “That's not going to work.”
The way forward is to bring many use cases into production. “It doesn't help to organize lots of pilot projects that I then discontinue.”
Using AI to combat the shortage of skilled workers
Will AI lead to job losses? “Companies tell us they need AI because otherwise they won't have enough people,” reports Gossmann. For most of them, the aim is to gain capacity for customer service. “Some talk about how they use chatbots in customer interactions,” he says. “If I've had an accident, I don't want to talk to a chatbot, I want to talk to a human being who will listen to me.” It is then important that the person affected receives a decision on the claim within two days. “Even if I don't get the benefit, I'd rather have quick information than have to wait six to eight weeks to hear from the insurer for the first time.”
Are insurers prepared for the fact that both industrial and private customers have completely different digital expectations – also because they themselves use AI? “I don't think so,” replies Gossmann. “Customer demands for speed of service delivery are increasing incredibly.” The industry can only meet these demands with AI.
Will AI become a competitive factor? Will companies that are poor in this area lose out in the market? Gossmann: “That depends on how quickly the industry develops. If everyone remains slow, it won't lead to enthusiasm among customers, but technically backward insurers will be able to keep up.” However, if individual companies develop more quickly, this will lead to noticeable distortions in the market.
AI-savvy insurers at the forefront of development
“In my opinion, an AI-savvy insurer will be at the forefront of this development, not an insurtech company,” says Gossmann. “This is because insurers have the data treasures.” Once the first insurers have tapped into these treasures, he believes it will have far-reaching implications.
Fraudsters are also upgrading with AI. However, Gossmann believes that insurers have a good chance of successfully defending themselves. “With good technology, it's easy,” he says. “We have just completed a pilot project with a British car insurer that involved identifying certain accident constellations.” The insurer fed the AI 300 (long-closed) cases, 50 of which the company believed contained elements of fraud. “We found all but one of them, plus two more that the insurer was unaware of,” says Gossmann. An interesting tool: if accidents are described in similar terms, this is an important indicator of fraud.
To date, DGTAL has invested more than €10 million, and the total will continue to rise. Gossmann expects further investors. “We don't yet know when we will be profitable, but to get there, we have to be very fast and bring products into production,” says Gossmann. “We can't live on pilot projects alone; that doesn't work.”
German Version: KI:Erstüberschätzt,jetztunterschätzt
Original published version: AI: First overestimated, now underestimated