Maximize your productivity with AI Assistants

Estimated Reading Time: 6 minutes

When I first investigated the world of AI, I didn’t expect much. At most, I thought I’d find a few time-savers, maybe some novelty tricks. But I was wrong. What I found instead was something transformative – a set of tools that quietly reshaped how I think, how I plan, how I create. These aren’t just gimmicks. AI has two major capabilities that I do not have. Ability to draw from a vast storehouse of human knowledge and the capacity to draw connections from disparate parts of that storehouse.

Have you signed up to our weekly job alerts on Actuary List? We post 50 new handpicked jobs every week that match your expertise. To ensure you don’t miss out, sign up here. Here are a few examples of new jobs this week:

What are AI Assistants

Then I stumbled onto AI Assistants. On first look, they seem too simple to be worth much. All they are is a capability to save some instructions (a universal prompt that applies every time the Assistant is used). And in addition, you can include some files with data that is also available every time the Assistant is used. Different platforms add in different additional capabilities, but those two features are the most important and are common to all platforms that support assistants. The AI Assistants are a feature on several platforms. They are called CustomGPT on ChatGPT, Gems on Gemini, Projects on Claude and Bots on Poe.

But just adding those two capabilities opens a world of possibilities. Things that I found helpful to do with AI become easily repeatable. And as an old programmer, I immediately started creating a series of AI Assistants that I think of as small programs.  

Here are three examples of AI Assistants that I have written and used:

Prompt Completer - If you have tried to learn best practices for prompting, you will have noticed an annoyingly long list of issues that you are told to address in every prompt.

InsDigitalStratgyBot - Assists insurance companies in developing and refining their digital strategies. It provides tailored advice based on specific company data or offers generic recommendations for the insurance industry at large.

AI ProblemSim - provides a business situation where an actuary can play the role of a Chief Actuary who must work with another senior officer to develop a solution to a new company problem. Practicing communication and persuasion.

Types of Prompts

These AI Assistants are built using several different types of prompts. In addition, some rely on some specific additional data usually in a file.

The types of prompts include:

● Single Shot Prompts - which means simply that the AI Assistant will be developed to answer one question. Additional questions would not be anticipated though they are likely to be allowed. The Prompt Completer described below is a Single Shot AI Assistant that needs you to provide a subject or question. Any single question that you might ask a LLM will do. Such as “Write a report on my project to update the assumptions for the UL model”

● Multi Shot Prompts - This AI Assistant would be expecting multiple questions and answers in a session. In fact, it may be instructed to ask some questions as well as answering them. The responses to the questions asked are used to enrich the answers provided. One example of Multi Shot would be PlanRiskMgtBot that asks a series of questions about your risk management program to develop a suggestion for plans for next year. Those plans will be delivered piecemeal and allow for adjustments from the user at each step.

● Chained Prompts - When the process is best broken into known steps a chained prompt provides robust responses. Can be combined with Multi Shot prompts to create a limited interactive process. An example of chained prompt with six prompts separated by~:

1. List three new and innovative types of insurance products that could help individuals or businesses manage financial risks related to climate change. ~

2. Briefly describe the customer need each addresses. ~Pick the one from your list that would be the most attractive to customers in Texas. ~Explain why it would be attractive in Texas, ~Describe its key features, including target market, types of coverage, exclusions. ~Tell how pricing might be determined.

3.  This prompt produces a nice 900 word report.

● Simulated Dialog - The AI can be told to act a role with only general rules for the dialog. The AI Assistant can mimic a conversation with a human. The conversation will go back and forth just like a conversation with a person and no two conversations will be the same. In a simulated dialog, the situation for the conversation is usually specified as well. So TomPeterBot will discuss a specified business strategy scenario with the user using ideas from Tom Peters writings.

This is an example of a simulation dialog between the user (in dark blue) and a Sim character, Dr. Nia Langford.

These prompts can be combined with different types of data files to create complex systems. The simplest of which is called Retrieval Augmented Generation (RAG). In a RAG situation, the AI Assistant will be able to utilize specialized information from files that have been added by the developer. RAG is AI with a smart filing cabinet full of information that the developer has provided.

The most amazing example of a RAG Assistant that I have created is InsStrategyBot. It consisted of a very simple prompt “Answer questions about Insurance Company Strategy using the information in your knowledge base.” The knowledge base is the name for the RAG files that the developer provides. In this case those files were extracts of a dozen insurer 10ks. The AI Assistant gives responses to strategy questions that consist of seven or eight bullet points that are real live examples found in the 10ks. Not a dissertation, but a great start for doing some competitor research. It was this single shot RAG Assistant that was the lightbulb moment about AI Assistants for me.

The final additional function that I have used with my AI Assistants is the Evaluation. The evaluation capabilities of the AI Assistant could be directed at an uploaded document, or they could be used with a Simulation Assistant to grade the user performance in a dialog. The evaluation can be assigned to a character. Imagine asking for the evaluation of your decisions in a simulation by Warren Buffet, for example.

Use Cases

Let’s go back to those three examples now and explain how they would use these parts to create their effects.

The Prompt Completer would be a single shot prompt. It would ask you for the subject of your query. It would have a system prompt built in that defines a particular type of situation in terms of the reason for the question, the audience, the format and the style. For example, you could create an AI Assistant like this to produce material for your boss in a consistent look and tenor.

The InsDigitalStrategyBot is a chained prompt which would allow input of one or more files that describe the company’s current digital strategy and that also uses a multi shot approach to allow the user to approve the Assistant’s suggestion for different aspects of digital strategy as it is being developed, rather than all at once after the entire things has been developed.

The AIProblemSim uses the Simulated Dialog type prompt with multiple characters being defined for different stages of the simulation including an evaluation step with a defined coach providing feedback. There is a defined problem that the user would be working on with a simulated counterpart who might be a help or a hindrance. Allowing the user to practice different types of situations that they are likely to encounter on the job and get targeted feedback on their performance. All in private.

Once you get used to assembling these AI Assistants, they become as second nature as spreadsheets. I have averaged creating close to two new Assistants per week. Think about it, do you have any idea how many spreadsheets you made in the past year? You could become that blase about AI Assistants. And those Assistants could become just as vital to your workflow.

What’s Next

I have since gone several additional directions with my AI Assistants. As a risk management actuary, scenarios are very important to my work and I found that an AI Assistant can be very helpful creating detailed descriptions of scenarios that can be used for stress testing or for planning, based upon relatively lean specifications. I have also created an AI Assistant that will tell me how to attack a problem using a variety of thinking styles that can result in very different conclusions. And recently, I took the Evaluation idea to the next level, creating an AI Assistant with four different personas who are instructed to criticise my work. And the toughest of the four is an avatar of me!

AI assistants have catapulted me from feeling threatened by obsolescence to running with the leaders of the pack. And did I mention, all of this was done in plain English!

Dave Ingram is currently the editor of the Actuarial Intelligence Bulletin and a member of the SOA Board. He can be reached at [email protected]