Companies today face a dilemma in any fast-developing tech arena: should they explore the technology with their in-house IT team to build a unique solution, or should they buy a pre-packaged system?
The question is urgent for insurance organizations looking to deliver a step-change in efficiency, talent utilization and preparedness for radically different operating models as systems driven by artificial intelligence (AI) become more common.
“Not adopting generative AI is likely to get you left behind,” Alex Taylor, Global Head of Emerging Technology at QBE Ventures, the innovation arm of veteran global insurer QBE, told viewers of a recent Roots Automation webinar, Harnessing AI’s Potential in Underwriting & Claims.
“The real question,” Taylor continued, “is how to adopt it.”
Building proprietary technologies has some incredible benefits. The main one is a competitive advantage. With a system in place that no rivals can offer, you can deliver innovative products and step-change efficiencies that put your business streets ahead.
In the early days of computing, large organizations needed significant investment in in-house development teams to reap big benefits. Today, the calculation is much more complex, and an ecosystem of tech providers is continually innovating. (After all, who writes their own accounting package these days?)
But the emergence of AI technologies – particularly the LLMs – makes the “build-it” approach tempting all over again. Companies such as OpenAI and Anthropic make their LLMs available relatively cheaply. They’ve already created engines capable of understanding and regurgitating information.
With the really tricky engineering done, why not build the tool you need yourself? Point your off-the-shelf LLM at your data, train your underwriting and claims teams in prompt engineering… and now you’re an AI-driven insurance organization… simple, right?
The problem: even knowing which LLM to choose is a major technological challenge – especially in insurance, where data security, accuracy and regulatory oversight are priorities. “The democratizing nature of these models makes it almost too easy to get started,” said Alex Taylor. “The risk of being wrong could be even more challenging than not starting at all.”
Another issue is that individual insurance businesses are limited to their own datasets, and even then, if their historical data isn’t well structured, the LLM might misinterpret or hallucinate.
“The tools are openly available, but if you don’t have the historical data from a number of sources, it just won’t be effective,” warned Dr. Anand Rao, Distinguished Service Professor of Applied Data Science & Al at Carnegie Mellon University.
And that’s before we even get to the money. “If you’re building an insurance LLM from the ground up that understands underwriting and claims management, [and] unless you have a really big benefactor writing checks for you, DON’T,” stressed Dr Rao. Even hiring specialists in AI and data science is prohibitively expensive. Demand for IT specialists with skills in generative AI more than doubled both in 2022 and 2023, the only IT specialty to post consecutive increases in demand.
Gartner estimates that even just developing a virtual assistant would cost between $8m and $20m for an insurance business, which leaves little mystery as to why Gartner also predicts 30% abandonment rates for all AI projects by 2025. “Building in this space is a bit like catching a falling knife,” said Alex Taylor. “It’s changing so quickly that it’s difficult to create systems that don’t age in place.”
Click to view Harnessing AI Document Processing Across Underwriting & Claims now on-demand.
Luckily in a dynamic tech market of competing innovators, there are also “buy” options for insurance organizations that want both to learn about AI’s impact on their future and deliver immediate operational benefits.
“Tech companies have the advantage of moving at a much faster rate than others, especially for a technology as complicated as GenAI,” said Ratish Dalvi, VP of Al & Machine Learning at Roots Automation. And Dalvi highlighted the core mistake made by many companies:
“People say is, here’s my data, make AI do something with it,” he explained. “I prefer the opposite: here’s my problem, how can we solve it?” This approach emphasizes that AI is just another tool – albeit one that’s continually growing in power and capability.
The most innovative organizations treat it as such. They explore the future by solving today’s problems and delivering results they can learn from. Deploying pre-trained Digital Coworkers on defined problems in underwriting and claims is showing triple-digit ROI… now.
“You still need strong MLOps [machine learning operations] practices, continuous training, checks on reproducibility, monitoring,” stressed Ratish Dalvi. That’s good news: deploying a specialist LLM like InsurGPT™ that’s pre-trained and can be targeted for quick results around, in particular, document management does not curtail AI learning within the organization. Far from it: the human-in-the-loop approach makes it even easier to spot fresh opportunities to deploy AI as it increases in utility and learns from data specific to the business.
“This is going to be too transformative, if you miss out it’s going to dramatically change the success of your business,” concluded Alex Taylor. With those stakes – and the costs of building a proprietary model so high – it’s no surprise so many insurance organizations see buying industry-specific, trained and tested products as their route to the future.