It has been just over a decade since the Harvard Business Review named data science roles as “the sexiest job of the 21st century.” The years since have been a roller coaster – we’ve seen massive waves of hiring activity for data science positions, widespread layoffs, divestment in research during times of economic downturn, and, today, threats of these jobs being replaced altogether with recent advancements in LLMs as popularized by ChatGPT.
There will always be a lot of buzz around this topic, but here at Insurance Quantified, we are laser-focused on driving business impact. We’ve settled on a practical approach to investing in and getting value out of data science, making it a consistent and integral part of our efforts to transform insurance through data, analytics, collaboration and workflow efficiency. Here are a few keys to success we’ve learned from all the ups and downs:
So many projects are developed in the proverbial “data science dungeon” – they never see the light of day. In many organizations, data scientists operate as part of an isolated research team with few connections and limited cross-sharing of information with other business units.
While a data science model might seem generally useful, it can be difficult to identify its exact use cases and determine precisely where it fits into a company’s complex network of products and business operations. Even when a model is specifically designated to support particular user- or product-facing teams, its utility can be limited severely if the data scientists are not integrated into those teams or do not show incremental progress to key stakeholders. Finally, some models with viable use cases are unable to be deployed and drive business value simply because the resources to do so are not available.
To mitigate these issues, we strive to strike the right balance between collaboration and specialization. For many projects, we integrate our data scientists into relevant product teams, while still allowing them to develop more research-oriented projects outside of those settings. And the process doesn’t stop with them. Once projects leave the data scientists’ hands, we heavily prioritize securing early buy-in from stakeholders and planning for potential deployment or productionization efforts, maximizing their chances of success.
Given the increasing attention and hype around data science and AI in both academic and commercial organizations, there is an enormous amount of research and numerous open-source models and tools/frameworks available. We’ve found that when starting a new project, spending one week on literature review and exploration can save weeks or even months of effort down the road with comparable or better performance.
For most organizations, this freedom to explore is by no means a given. The data science team needs buy-in from management so it can research new avenues to creating value, and not be singularly focused on execution-oriented work. This approach requires a careful balancing act with lining up projects against real business use cases. Data scientists must constantly expand their knowledge to remain on the cutting edge, but ultimately, everything should be oriented toward delivering value to the product and users.
As alluded to before, you cannot build data science solutions that generate business value without engineering. And without serious investment in engineering capabilities and engineering standards within the data science team, it can be difficult to bring new models to production.
This is largely because it is easier for things to become lost in communication in a partner engineering team structure. Whether it’s rewriting the code developed by the data scientists or coding bugs in research projects that misrepresent the true accuracy of the system, there tend to be domain knowledge gaps on both sides that make handoffs and collaboration challenging.
At Insurance Quantified, we’ve mitigated these challenges by encouraging data scientists to develop their engineering skills and creating a partnership model with our engineering team, as opposed to a handoff model. This has resulted in projects where data scientists develop and even ship production-grade code, with assistance and review from our engineers through joint design sessions and code reviews.
Furthermore, because our data science team understands core engineering considerations such as latency, throughput and infrastructure requirements, it can keep engineering impact in mind when working on research projects. We also prize flexibility in tasks and job responsibilities, encouraging data scientists to push forward with certain aspects of production in the event that engineering resources are unavailable. And the benefits run in both directions, with engineers encouraged to contribute to data science efforts.
We’ve learned many lessons from our work to integrate data science into our business operations. Perhaps the biggest is that the necessary changes extend much further than the data science team alone. In addition to engineering, it is essential that product managers have the right framework to effectively drive and utilize data science in their work.
This is easier said than done. Given that data science is still a relatively nascent field in industry, most companies are unfamiliar with best practices for adoption. Additionally, data science projects typically have a higher risk-reward profile than engineering projects, where it is easier to discretize the investment and achieve success. This can make data science projects particularly difficult to plan and prioritize, especially weighed against engineering initiatives. Finally, industry-wide reliance on execution-oriented frameworks like Agile means it is often too easy to allow research-based data science tasks to fall by the wayside, among other challenges.
In order to overcome these challenges, we heavily involve our data scientists in all product planning initiatives. We’ve seen firsthand the importance of clear communication between our data science and product teams, particularly when it comes to identifying which steps of a data science project are more research-based (more likely to result in changes to the plan) versus more execution-based (less likely to impact the plan). When such research-based tasks do exist in a product execution plan, it is critical to account for variability of outcomes and build flexibility into the plan, rather than react on the fly when things change.
We do not claim to have all the answers when it comes to data science – nor should any organization out there. The state of data science is changing rapidly, as is the world at large, and it is important to be adaptable to these changes. Nevertheless, we believe many of these principles and best practices will serve us well for years to come as we continue to iterate on driving practical data science at Insurance Quantified. And we look forward to continuing to apply these ideas to help P&C carriers and MGAs grow their business by developing their data-driven edge.