How to deploy AI in insurance: Your top 10 questions answered
Charles Dugas Charles Dugas
October 8 10 min

How to deploy AI in insurance: Your top 10 questions answered

It’s no secret that the insurance industry is primed for disruption. All you need to do is type terms like “AI for insurance” or “latest insurtech” into your search engine of choice, and a veritable cornucopia of articles will pop up preaching the advantages of adopting new, intelligent technologies. With so much information available, it’s no wonder that many insurers are struggling to figure out how to embark on their AI journey in a way that ensures projects yield the desired results.

Over the past two years, we’ve met with insurers throughout North America, Europe and Asia. Here are answers to some of the most common questions we’ve been asked.

1. We know we need to be thinking about AI, but we’re not sure where to start. What are some of the things we can do to kick-start the process?

Like most large-scale projects involving multiple stakeholders, developing a strategic roadmap for rolling out AI is absolutely crucial. We recommend you start by identifying use cases and ensuring they are relevant by engaging the right business stakeholders and users in the process. Each use case should be carefully framed, with clearly specified objectives and desired outcomes. Once identified, the use cases can be assessed and prioritized based on a set of key criteria, including desirability, viability, feasibility and risk:

  • Desirability reflects the importance of the specific problem considered for end users (i.e. how badly do we want it solved?)
  • Viability evaluates the expected economic benefits or ROI (return on investment)
  • Feasibility assesses whether a solution to the problem is achievable, including data, technology and people considerations
  • Risk gauges the potential adverse consequences of a model failure and the ability to mitigate against them

It is also important to choose a manageable scope, as developing an AI roadmap for the entire company, one line of business, or a specific workflow will require wildly different efforts. It is advisable to start with a more focused scope and expand across the organization in successive waves, each leveraging the learnings from prior iterations, rather than attempt to do it all at once.

2. What kind of data is needed to deploy an AI product?

Data is critical to deploying any AI technology. We group data into two buckets: structured and unstructured data.

Data required to deploy an AI product

Structured data is highly-organized and formatted in such a way that it's easily searchable in relational databases. Structured data includes information on policies, claims and payments that is stored in a company’s systems of record.

Unstructured data has no pre-defined format or organization, making it much more difficult to collect, process and analyze. It can be in the format of text, images, sound or video and is typically much more of a challenge to search. Examples of unstructured data could include emails from brokers, satellite imagery (google maps), financial statements, call center recordings, pictures of damaged cars and social media posts.

While most companies see a mix of both structured and unstructured data within their organization, we’ve observed that the majority of insurers' data is unstructured. While structured data is preferable when deploying an AI product, modern AI algorithms can leverage both types. In fact, recent advances allow AI to effectively process many types of digital information, including scanned documents, images, videos and even audio recordings.

3. What can insurers that are lacking data do if they want to deploy AI technology?

The first thing that needs to be determined is whether there is truly a lack of data or if what is really missing is labelled data.

Unlabelled data has not been tagged with one or more labels, and it’s these labels that enable AI models to accurately read the data and understand what it represents. Photos, audio recordings, videos and, x-rays can all be forms of unlabelled data. If it is an issue of data not being labeled, there are tools that can accelerate the labeling process and help insurers build AI-ready datasets.

If the issue does indeed stem from a lack of sufficient data, then an insurer needs to rely more heavily on pretrained, off-the-shelf models. These products, however, tend to have lower predictive accuracy, since pretrained models do not account for the insurer’s unique specificities.

4. If we deploy an AI product within our business, will we be forced to migrate from our current digital platforms?

Migrations are messy, expensive, time-consuming and generally a painful process for most involved. This is one of the reasons why it’s important for AI products to integrate seamlessly into your current workflows.

The seamless integration of AI products.

In order for AI algorithms to access data and provide a recommendation or automate a task, data needs to be available in digital format. This is certainly the case when a digital platform has been deployed. Even in the absence of digital platforms, some insurers have managed to apply AI for certain use cases, but these cases are limited to areas where the data is available in a proper format. So in short, no migration necessary!

5. The insurance industry is highly regulated, and regulations vary by state. How does AI take this into account?

AI models are not plug-and-play products. They learn continuously through mechanisms designed to capture human judgment, knowledge and experience, further improving the performance of the AI model over time.

Thanks to this continuous learning, industry regulations, including state-specific regulations, can be applied, for example by imposing constraints on the AI model structure or by limiting the data that is made available to the model.

6. Our company is currently deploying another new system. Is it possible to deploy an AI model simultaneously?

In short, yes, in some cases it can be done. When deploying an AI model for the first time within your firm, the model will rely on historical data to be trained. That data would come from the older system.

Insurers deploying multiple systems or platforms at the same time as an AI model must make sure that the data coming from the new system is represented the same way as the data coming from the older system. Typically, newer systems provide more data with more granularity than older systems (which, when it comes to AI, is preferable). It is therefore possible to deploy an AI model while deploying or migrating to a new system.

7. Our actuaries are already using advanced analytics and machine learning (ML). Do we really need to look outside for other AI solutions?

While some insurers’ in-house analytics teams have seen success developing AI and ML models for specific uses, the reality is that it is very difficult to bring analytics to production. According to Gartner, by 2022, 85% of AI projects will fail in production.

Given the level of complexity of advanced analytics and ML, compounded by the high probability of delays or failure of AI projects, in the long run working with an AI product provider can be a time - and cost - saving solution. AI products can be deployed on-site quickly and they are highly customizable, thanks to human-in-the-loop features like tolerance thresholds settings for automation and corrective feedback provided by the end users.

8. What business areas are most primed for AI?

While there are opportunities to deploy AI models across every stage of the insurance lifecycle, with the technology currently available to us today and based on our experience working with various clients within the insurance industry, we believe that the areas where insurers will see the most return when applying AI are in the underwriting and claims processes. However, there are also opportunities to apply AI to other areas, such as marketing, customer targeting or product recommendation, to name a few.

At the end of the day, the best use cases to select are those that will best serve your organization’s priorities.

9. What happens when we are ready to deploy an AI product?

Once your team has selected the use case, prepared your data and fine-tuned an AI model the next step is to integrate the AI product into your technology environment, including all the systems it will need to interact with within your organization. This is done through API (application programming interface) calls. What’s an API? It’s essentially a set of procedures and tools that help developers build software applications. The API will specify how the different software components will interact. An API call is your APIs being put to work. Any time you perform an action at your computer (e.g., send an email, download an app, enter a password), you are making an API call. In the context of an AI product interacting with, say, your systems of records, an example of an API call is requesting a price quote for a new business submission.

It can take anywhere from eight weeks to a few months to complete this deployment process. Unlike traditional robotic process automation (RPA) software, which is rules based and static, the AI model, once it is set up, will continuously learn and improve its decision-making over time based on thresholds you determine and control, as well as on corrective feedback provided by end users (e.g., clerks, underwriters, adjusters).

10. How should insurers plan for change management when adopting AI technology?

Beyond developing an AI strategy, securing technical talent, investing in data and technology infrastructure and putting in place the right governance framework and controls, insurers should not underestimate the importance of change management in ensuring the AI solutions they adopt are successful (whether developed in-house or by third parties). As with any major initiative, insurers can plan for effective change management by identifying and empowering internal champions within the organization, in addition to developing and delivering a thoughtful communication strategy, both internally and externally, and providing adequate training and coaching, before and after deployment.

Here is a practical list of activities and deliverables that insurers can start with to effectively plan for change management:

  • Look to your use cases to ensure that impacted business processes and standard operating procedures are up-to-date. It is important to know what you are changing from, to understand what support, training and organization change will be required to be effective in the future state.
  • Update job descriptions and capture the tasks associated with impacted jobs. This will help identify training requirements and determine whether you need to reorganize job tasks or work assignments.
  • Review the content and completeness of your internal training and onboarding processes for the job roles affected. Once AI is deployed, certain staff may be taking on new responsibilities and may require training or support to be set up for success.
  • Communicate, communicate, communicate. It can’t be said enough! Companies should not only share the vision and roadmap for how AI will be developed in the future but also how employees will be affected and the support systems that will be available to them to ensure a smooth transition for all.

Learn more about how to start planning for AI adoption within your organization by visiting our Insurance page or contact one of our AI experts today.