This article was originally published by QBE Ventures. We have reposted it here with their permission.
AI is a key focus for QBE as we continue our ambition to be the most consistent and innovative risk partner. There are many applications for AI across the entire insurance value chain, with insurance itself being a highly specialised industry with unique datasets and use cases.
As such, there are several key industry challenges that create opportunities for start-ups, technology companies and incumbents to collaborate.
Training data quality is the single biggest determinant of model performance. Insurance data is typically highly inaccessible: reports suggest that 80% of insurance data is unstructured, unlabelled, and not ready for AI model training. Finding ways to utilise unstructured data for AI/Machine Learning (ML) use cases requires platforms that not only make the data accessible, but do so in a way that can be built on by non-technical stakeholders.
In addition, ‘off the shelf’ Generative AI models are constrained in their ability to meet niche industry use cases. Models such as GPTx, Bard, Gemini and Claude will often require a high degree of prompt customisation to perform with the level of consistency, accuracy and reliability deemed sufficient for a regulated market. The limits of Retrieval Augmented Generation techniques (‘RAG’), which retrieves relevant supporting information, but does not necessarily improve model understanding of the task at hand, will mean that as we enter 2024, we expect more large enterprises will start to explore fine-tuning foundational models. This will push insurance companies to open-source base models, which allow an organisation to bring them in-house and fine-tune at an accessible price point (with the important advantage of retaining full control of their data, intellectual property, and newly created proprietary tokens).
As key challenges are solved, our view is that the most successful insurance market participants will succeed, not with one AI model leveraging proprietary data, but with hundreds (if not thousands) of models. This means that an insurer’s ability to select, adopt and customise ML and Generative AI at scale will be a key competitive battleground.
Enterprises are already starting to maintain these libraries of models, striking a specific balance between accuracy, cost, performance, and stability per use case. At the project level, there is significant opportunity to make it easier when having to decide which model, and how to inject business context into the data sources in a transparent, efficient way.
The QBE Ventures team has been scouting for best-in-class founders leading the world in providing the rails for simple, scalable, and affordable ways to customise ML models and fine tune foundational Generative AI models.
Snorkel AI started as a research project in the Stanford AI Lab in 2015, where Alex Ratner, Chris Re, Paroma Varma, Braden Hancock, and Henry Ehrenberg worked together to help use AI to tackle human trafficking. They found that the lack of labelled training data was a crucial bottleneck.
After five years of developing the product and deploying it within organisations including Google, Apple, Intel and the US Department of Defence, the open-source collaboration evolved into a platform for data-centric AI called ‘Snorkel Flow’, which enables programmatic model iteration, collaboration and labelling at scale.
Snorkel Flow makes the classifying and managing of unstructured data easier and faster which reduces the challenging and expensive need for collaboration between business experts (who have the business context) and data scientists (who have the base models and ability to work with them), as well as the technical requirements for data-centric AIOps (Artificial Intelligence for IT Operations).
QBE Ventures’ introduction to Snorkel AI came from our QBE data science and claims analytics peers. QBE’s North American teams use Snorkel Flow across a variety of predictive analytics use cases. The immediate value has come from reducing the friction involved in converting vast amounts of previously locked-up corporate data to improve the outcomes of ML solutions being applied to claims and underwriting business processes. We invested to help accelerate the evolution of Snorkel AI as it pushes further into Generative AI.
The investment in Snorkel AI marks a significant step in QBE Ventures’ commitment to embracing and advancing cutting-edge technology for the insurance industry. It’s a long-term strategic partnership anchored in establishing industry specific patterns for the responsible and explainable use of data-centric models.
The most impactful innovation often happens when cutting edge academic concepts are applied in practice, at scale, for real-world problems. Achieving this requires strategic and significant collaboration between early-stage companies and industry incumbents.
“Ensuring carriers have the capability for customising models in safe and scalable ways is of paramount importance,” said James Orchard, QBE Ventures CEO.
“We’re excited to be working with Alex and the founding team to help pioneer the capabilities needed to enable the insurance sector to adopt ML and Generative AI in ethical, fair and data-informed ways”.
Alex Ratner, co-founder and CEO at Snorkel AI believes we are at the early stages of understanding the potential of Generative AI: “Our partnership with QBE brings valuable industry experience that will help us productize insurance specific use cases in Snorkel Flow, making it easier for carriers to get faster and better value from their AI projects”.
Over 2024, we’ll bring together founders, QBE executives, leading applied researchers and industry experts to experiment and learn together. We’re excited to advance the existing QBE-Snorkel relationship to develop Generative AI operating models, accelerators and patterns for the insurance sector.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。