AI in Early 2023

How will continued integration affect actuarial work? Jeff Yang

The field of generative artificial intelligence (AI) is advancing at a rapid pace, with new developments in text, image, audio and video generation emerging regularly. We are beginning to see the widespread implementation of AI across various sectors, and the speed at which progress is being made is astounding. If the current level of enthusiasm and innovation persists, AI may very well be integrated across many areas of actuarial work in the near future.


In November 2022, OpenAI released ChatGPT, a popular chatbot interface built on top of its generative pre-trained transformer (GPT) models. ChatGPT is an autoregressive model that uses machine learning algorithms to probabilistically predict the next sequence of characters, known as a token. In short, it is a sophisticated autocomplete engine.

ChatGPT is the fastest-growing app in the history of the internet, having reached 100 million users in just two months. The instant popularity sparked the start of the AI race, prompting Microsoft and Google to publicly announce their respective AI-based search engines. OpenAI’s public application programming interface (API) also allows any developer to integrate AI into any application, giving rise to numerous AI-based apps and startups.

As of June 2023, the free version of ChatGPT uses GPT-3.5, which was trained on 175 billion parameters.1 A paid subscription to ChatGPT Plus gives the user access to GPT-4,2 which is shown to perform considerably better than GPT-3.5 by several measurable metrics. GPT-4 will also eventually support multimodality, allowing for combined image and text processing. Additionally, ChatGPT’s support of plugins3 allows the AI to access up-to-date information, run computations and use other third-party services to compensate for its current weaknesses, such as not being able to access the internet or struggling to evaluate simple math equations.

Business Impact

The impact of AI is ubiquitous. It can already boost the productivity of an actuary’s day-to-day tasks. Here are some examples:


AI can:

  • Help actuaries expand notes into full sentences, making it much more efficient to convert ideas into a format recipients can understand.
  • Summarize lengthy emails or email chains into concise bullet points, making it easier for actuaries to catch up on emails.
  • Help actuaries improve their writing by learning from the user’s style and offering suggestions.
  • Help explain technical concepts to nontechnical audiences.


AI can help:

  • Maintain consistent company branding and formatting in deliverables, ensuring a professional and cohesive appearance.
  • Edit a deliverable to have a consistent tone, style and verbiage, especially if several individuals worked on the same deliverable.
  • Proofread formal documents and actuarial certifications with strict requirements and ensure that all requirements are met.


AI can:

  • Help reduce spreadsheet errors by identifying labeling, formula or logical inconsistencies.
  • Explain the purpose of a specific formula or a series of formulas.
  • Generate an Excel formula or formulas given a nonformulaic prompt. An example is provided later in this article. Various AI-based Excel tools are already in development, and Microsoft has plans to incorporate AI features into Office 365.4


Whether with a client or on an internal call, AI can help:

  • Prepare discussion topics and create agendas.
  • Take meeting notes or summarize important discussion topics.
  • Identify inconsistencies if multiple callers take notes, ensuring more accurate records.
  • Provide live speech-to-text, which can assist callers in noisy areas who don’t have access to headphones, are hearing impaired or are in other challenging situations.

Chatbots and AI Assistants

Chatbots and AI assistants trained on relevant knowledge—such as emails, meeting notes and work files—can serve as helpful knowledge databases. For example, a chatbot can help train new staff on fundamental concepts and company best practices without taking time away from a manager’s workload. This allows for more efficient onboarding and continuous learning for employees.

Often, there are follow-ups to a meeting. Sometimes it’s just forwarding a document or not remembering a certain fact on the spot. AI assistants that have this information readily available can assist in real time during meetings, reducing the frequency of follow-ups.

Actuarial Example

Aside from general operational improvements, AI can help support actuaries in technical aspects of the job. Given how versatile the technology already is and how fast it’s developing, the ultimate limiting factor could be the user’s ability to steer the AI in the right direction by providing relevant and appropriate prompts.

Here is an example of how GPT-4 could help a Medicaid pricing actuary develop an analysis. This example is not to suggest that AI can supplant the expertise of an experienced actuary. Rather, it is to demonstrate that AI can help the actuary brainstorm ideas.

The responses in this example are excluded for the sake of brevity­ since the responses are hundreds of words long. Any reader with access to GPT-4 can replicate the prompting technique below and draw their own conclusion on AI’s helpfulness.

Prompt 1: I am a Medicaid pricing actuary. Help brainstorm ways to calculate the per member per month (PMPM) price impact of adding a new hospital to a network.

Response 1 (summary): GPT-4 responds with a detailed list of 11 valid considerations. All 11 were relevant and worth considering. To continue with this example, we follow up on the first item in the list, “Analyze historical utilization and cost data.”

Prompt 2: I have historical claim data at the claim line level. How would I develop the PMPM impact on children’s health insurance program (CHIP) enrollees from this data? Provide specific formulas in Microsoft Excel.

Response 2 (summary): GPT-4 responds with a 415-word step-by-step process to perform the analysis. It includes specifics such as:

  • Step 1: Organize the data. Ensure your claim line data includes key information such as enrollee ID, claim date, claim amount and service type. You may also want to include additional information, such as diagnosis codes, procedure codes and provider information.
  • Steps 2-8 (summarized): Filter the data for CHIP enrollees. Calculate the total claim cost for each enrollee. Calculate the total number of member months for CHIP enrollees. Calculate the historical PMPM cost for CHIP enrollees. Estimate the PMPM impact of adding a new hospital. Calculate the new PMPM cost for CHIP enrollees. Calculate the PMPM impact.

The response includes specific formulas in Excel and the reason for performing each calculation. The formulas and reasoning are sound, and while the steps should be obvious for any seasoned Medicaid pricing actuary, one could imagine how the AI can assist with more complicated analysis following a similar thought process.

Results from the base GPT-4 model can already increase an actuary’s productivity. However, this can only improve if the AI is specifically trained on the subject matter (e.g., a Medicaid pricing chatbot) or if the AI is directly integrated into relevant software, such as Excel, Outlook or SQL.


As with any new technology, several important risks exist. Here are a few of the risks actuaries could face when using AI.


  • Risk: For text-based AI models, the model is known to “hallucinate” when it responds with a confident and factually incorrect statement. The probabilistic nature of large language models makes them prone to hallucinations since they are predicting patterns rather than reasoning in the traditional sense.
  • Mitigation: The user should always review the AI’s output for logical consistency. Additional techniques to mitigate the impact of hallucinations include:
    • Asking the AI for similar prompts and comparing results
    • Following up with specific questions
    • Asking the AI to reflect on its response5
    • Regenerating a response with the same prompt and comparing results
    • Applying tree-of-thought prompting6

There are also various resources to learn how to communicate effectively with ChatGPT, such as

Inconsistent Results

  • Risk: GPT-4 is probabilistic. The model can produce a different output when provided with the same input. This is good for generating creative results but not beneficial for problems with deterministic answers.
  • Mitigation: The degree of variance can be adjusted via the OpenAI API or OpenAI’s playground. However, even with the lowest variance setting, the results are not necessarily deterministic. Variance in responses can be mitigated by training specific models or using other prompting techniques to increase reliability.

Evolving Technology

  • Risk: AI technology is improving at a rapid rate. As a result, today’s model could produce very different results from tomorrow’s model.
  • Mitigation: Fortunately, each successive version of GPT has shown significant improvements over the previous model. With increased focus, resources and competition in the AI space, this trend of improvement will likely continue. OpenAI also allows developers to access older models, so those endpoints could be used to offer consistent responses over time.

Data Privacy

  • Risk: Any website or application to which you provide information can store your data. A naive user might prompt an AI-based application with proprietary or confidential information, thinking that the conversation is entirely private. This risk is further exacerbated if user data is somehow leaked to the public, either intentionally or unintentionally.
  • Mitigation: Awareness is a good first step. Users should read an application’s terms and services and privacy policy before inputting sensitive information. In addition, running a self-hosted version of an AI algorithm is increasingly affordable for corporations, which could help mitigate concerns about providing data to an external party.


As a profession centered around risk management, it’s reasonable to approach AI with a healthy dose of caution. We are far from the point where AI can be injected into an established system without any human supervision. However, with the proper guidance, AI models and their implementations can provide significant value to actuaries.

The technology is still new, and it will take time to figure out how to innovate safely and effectively. Nevertheless, I am excited to see where this technology takes us in the years to come.

Jeff Yang, FSA, is an actuary, developer and educator at The Actuarial Nexus, which implements AI tools to help students prepare for actuarial exams.

Statements of fact and opinions expressed herein are those of the individual authors and are not necessarily those of the Society of Actuaries or the respective authors’ employers.

Copyright © 2023 by the Society of Actuaries, Chicago, Illinois.