AI in Action: Exploring the Use of ChatGPT in Retirement Planning

The downside for knowledge workers is that our debt is time, and we have a bottleneck problem. We have lots of ideas and opportunities and are responsible for administrative work. We must consider the best use of our time and look for ways to make things easier and more enjoyable or help us be more productive. This sets the stage for why using technology in general is important.

An Introduction to ChatGPT

I had dinner with my folks a couple of weeks ago. My parents are in their 70s. I asked them, “Do you use artificial intelligence?” My mom said, “No, I do not use AI.” I asked her, “Have you ever said, “Hey, Google?””

Sometimes, on accident, you say, “Hey, Google,\” and your phone pops up. That is AI. What about if you are working in a Word document or Outlook? Do you ever notice that as you are typing it, it sometimes completes your sentence? That is a type of AI. How about when you are sending a text? Sometimes, even before you send the text, the incoming one says, “Here are three different responses you might want to choose from.” That is artificial intelligence. How about Netflix? “Based on your watching patterns, we think you would like to watch these things.” The point is that we already use AI.

How is this different from a search engine in Google, for example?

For one, there is this functionality aspect. ChatGPT is what I am going to focus on. It is very conversational. The chat part of ChatGPT is interesting.

With Google, you search, and then that is it. Right? Whatever your second search is, it does not matter what the first term was; they are independent searches. ChatGPT, however, is a conversation. Whatever you put in after your first prompt or request is part of that conversation. It can reference itself.

From a data source perspective, it’s different than Google. Google is constantly crawling the web and, theoretically, always getting the most up-to-date information based on whatever it is that you are searching. ChatGPT is different. AIs are different. Each one is different from the other. Bing has a ChatGPT type of thing. That one references the live web.

ChatGPT is different. It is trained on an extensive set of content from the internet, and then it relies on that as its data source. For a real-time update, it depends on which version you are using. For example, if you know about current inflation as of October 2023, the default answer might be, “Hey, I am only up to date until September 2021.” Some things to be aware of.

The G in ChatGPT means “generative.\” So, what does generative mean? Unlike a passive type of AI, it generates content based on your prompts. P stands for “pre-trained,\” meaning it was based on information, based on the internet. T stands for “transformer,\” which is how it takes the inputs you give it and then transforms them back to you in the media you prefer (including audio).

How popular is ChatGPT? The popularity of tech, in general, is measured by the time it takes to get to X number of users. One hundred million users is usually a good benchmark. How long did it take for Netflix, which came out in 1999 as a DVD exchange service, to reach 100 million users? Ten years. How about Facebook? Facebook came out in 2004, and it took 4.5 years before reaching 100 million users. Instagram took two and a half years to get to 100 million users. ChatGPT came out in November of 2022. It took just two months to get to 100 million users.

How do You Use ChatGPT in Retirement Planning?

Should advisors use this in practice? Maybe. If you are in a compliance role, you will appreciate this. Does it matter if the output is accurate? If not, go ahead. You are safe to use ChatGPT.

Do you have the expertise to verify that the output is accurate? Again, if the answer is no, and this is outside your domain, then we will say it is unsafe to use. Are you able and willing to take full responsibility for any inaccuracies? It puts you, as the expert, in the role of filtering this type of information.

What are some ways that you could use this in practice? How about from a research aspect when you need to prepare for a client meeting regarding a tax law change?

How about from a customer service perspective or a content creation perspective? Sometimes, when I want to write an article or a newsletter, I have a vague idea of what I want to do, but I need help getting a first draft.

What about retirement planning recommendations? This is the big question mark. Can you use it to help generate recommendations for clients through the retirement planning process? Maybe.

What are some things that you find yourself Googling today? Tax rates, what laws have changed? We have contribution limits, RMDs, tax brackets, and retirement strategies.

What might this look like if you ask Google, “Explain how RMDs work.”? You get a ton of results. Some of those might be accurate, but some of them need to be updated. This is not unique to ChatGPT. Anytime we query a source, we need to do a fact check. The problem with this is that it can take a lot of time. Which one of these should I rely on? Is it accurate?

So, let us use an example here. If you open a search browser and type in ChatGPT, the official one is from openai.com. I am using the free version, which is version 3.5. You can pay $20.00 monthly if you want to use GPT 4.

When I type in “how required minimum distributions work,\” it generates the text quickly and gives me a list of six things. Every time you do this, the results are slightly different. When you select “regenerate” at the bottom, it will display a different response.

Then, you can query from a customer service aspect. Ask, “Can you please explain what an RMD is again?”

Then, tell ChatGPT, “I am a financial planner, and a client asked me that question. Respond to my client as me so I can email them back.” The interesting thing about this is that it knows we are having an ongoing conversation.

Chat GPT responds with the subject line: Explanation of RMDs. It describes RMDs as essential to retirement planning and has six bulleted items. It finishes with, “Incorporating this in your strategy is essential. And please do not hesitate to reach out with more questions.” And it gives you a salutation.

Would you ever copy and paste that output and call it done? No, right? It is not written in your voice. This is a perfect first draft.

Next, let’s test ChatGPT’s recommendations. Let’s ask it what it thinks the best RMD strategy is. Type in a generic question. “What is the best RMD strategy?”

Have you ever had a friend who is a know-it-all? That, by the way, is ChatGPT. Whether it knows the answer or not, it will give you an answer. For this query, it provided a 10-item bullet point.

You can also use this from a content or ideation perspective. You may be trying to write a quick blog post. Write, “As a financial planner, this question comes up often. For my monthly newsletter that I send to clients, write a 250-word column for the newsletter explaining RMDs.”

It gave the article a title, and that is great. Even though we told it that we were writing a newsletter, it wrote it as an email.

What is Prompt Engineering?

Prompt engineering is the differentiator. Sometimes, people say, “I played around with it, but it did not give me the response I was looking for.” Absolutely. Part of this, like anything, is knowing how to use it.

So, what is prompt engineering? Well, it is part art and part science. And the whole purpose is that it involves crafting or adjusting the input that you provide in a way that improves the model’s response. In other words, you are trying to give it a better prompt and instruction to get a better response.

Because we are in the financial planning world, we devised an acronym called rATe.

• R is role
• A stands for ask or assignment
• T is the tone
• E is extras.

In your prompt, define the role. It could be your role but also who you are writing to. For example, I am a financial planner and writing a client newsletter.

The A is the ask or the assignment. This is the meat and potatoes you are telling it to do. Think about the output you want to get from this: is it to explain something or a bulleted list? What is the format that you want to see this as? Is it one line? Is it just a single paragraph? You can tell it to keep it to 250 words or that you want four paragraphs.

People have asked, “Well, can I just have it write a whole book?” The answer is yes, you can, but there is also a limit on the amount of text that it will generate for any given prompt.

T is tone. Like in any other setting, you decide the tone you are writing or speaking in. Is it a professional tone or with humor? Respectful versus irreverent. Enthusiastic versus manner of fact.

E is for a general set of extras that you could do. There is one that I like. Because ChatGPT is an interactive tool, you can ask, “Hey, what else would you need to know before doing it?” So, that one is an interesting one.

You can also ask for a response from a committee aspect. “I want an investment recommendation, but you are a committee of four investors.” You could say, “I want to use John Bogle, Warren Buffett, etc.” Come up with that list and say, “I want a response from each of you.” That is interesting.

Sometimes, you will go down a path and realize it is stuck. Sometimes, you have to tell it to ignore everything we have discussed up to this point. The other way to do that is to create a new session. But sometimes, you visually want to reference the information you have generated.

Returning to that original RMD example, we will implement the prompt engineering aspect to see if we can get better output. The original idea was, “What is the best RMD strategy? But ask me questions before you respond to get specific results for me.”

Here is a sample ChatGPT prompt engineering example:

“What is the best RMD strategy? But ask me questions before you respond to get specific results for me. I am 75.”
“Do you have multiple return rate costs or just one?”
“I do not know … three Traditional IRAs.”
“Have you already started taking RMDs? Or will you in the future?”
“Yes.” (I would like to see what it thinks my answer is because these are two different questions in one.)
“Do you have any specific financial goals or needs such as using for living expenses?”
“Yes, I need them for living expenses.”
“Are there other sources of retirement income that you will rely on during retirement?” “Yes, I have Social Security and a small pension.”
“What is your risk tolerance and investment strategy for the retirement accounts?”
“Moderate.”
“Are there any anticipated changes in your tax situation?”
“Yes, I think rates will increase.”
“Have you considered Roth conversions?”
“No.”
“Here is a tailored RMD strategy for you.”

• Continue to ensure that you meet the IRS’s minimum distribution for each of your three IRAs.
• Prioritize living expenses.
• Since you need your RMDs, planning for consistent distributions throughout the year is wise.
• Discuss partial Roth conversions with an advisor on converting a portion of your IRA to Roth. It can help manage future tax liabilities significantly if tax rates rise.

Then let us continue and say: “As a retirement advisor, that was for a client. Write an email to the client with the specific recommendations.” Again, we are trying to use prompt engineering. We are telling it who we are, who we are writing to, what we want it to do, and the format of it.

Let’s end with something funnier: Say you are a retirement planning expert. So, I am giving it a role. I am going to tell it to write a blog post about RMDs. I will ask it to write in the voice of Cookie Moster, with the audience being a group of retirement planners. Here’s what it came up with.

“Hello, hello, me friends in retirement planning! Me, Cookie Monster, and I are here to talk about something very important.

It does add a disclaimer. “This advice is for fun, not a substitute for professional financial advice. Please consult the financial expert for planning guidance.”

My point here is not to do something silly but to show how you can get creative with it and think about different ways that you would want to get the type of output.

What are the Limitations of ChatGPT?

The first is that it is a know-it-all. You could give it very little instruction, which will still generate a response. Whether or not you think it is a good one, whether it is helpful or not, it will generate something. It is just technology, so that is its job. I like to think of it as garbage in, garbage out.

Sometimes, the feedback I hear is, “Oh yes, I have used ChatGPT,\” Or “Yes, it just did not do what I thought it would.” It may be based on the type of prompt that you used. This gave you some initial ideas on the ways that you could improve the actual prompt.

Think of this as our first draft. I would never copy and paste this and go with it. We started this by talking about the time savings element for today’s knowledge worker. And this provides some initial thoughts on how to get you past that getting stuck in your head. At least, that is how I feel.

With the trapped-in-time element, remember that if you are not using the paid version, ChatGPT is up to date with current events until around September 2021. So, again, what did the stock market do today? It is going to respond, “I have no idea.” The paid version, Bing, or even Google Bard, queries the internet. Just be aware that some of it could be trapped in time or up to date; it depends on which AI platform you are using.

Hallucinations is this term where AI makes up answers. An attorney was trying to find a reference for some legal case. He said, “Provide a citation for this legal precedent.” And it came up with one. The attorney asked ChatGPT, “Are you sure that is a real law case?” And it said, “Yes, it is.”

The legal precedent ended up not being a legal precedent. A judge couldn\’t verify it. My point is to be aware; it is up to you to decide whether it is valid.

Back to our RMD example. The age ChatGPT recommended for starting RMDs was 70 1/2. It mentioned something about age 72, but it is not. secure 2.0 passed, and it changed it to 73. Right? And there is a future bump in there to 75 down the road. My point is that you are the expert for your clients. So, be the expert. Retain your expertise, and ensure that if your clients are using it, you have some idea now of the type of output it generates. And you can step in and be that buffer to say, “Look, I have used it too. Sometimes, it generates information that is not right.” So, be that expert for your clients.

Key Takeaways

The first is just how rapid the adoption has been. Twenty percent of adults living in the United States have at least tried ChatGPT or one of these so-called generative AIs. If you have not used it, you probably will. It will be like smartphones. I can remember when I got my first smartphone. And I thought I was pretty cool because I was one of the first. ChatGPT will be similar to that, where it will either be people who have never used it versus those who have used it. That gap will keep tightening.

The other is this idea that AIs, in general, whether ChatGPT or something else, streamline retirement planning operations if you know how to use them. The whole point is to free up some of the time you spend not on technical-based activities but to free you up from some of these routine things you must repeatedly do.

My example was responding to generic emails, which you could use to quickly get a good first draft. We also reviewed an example of where we told it to ask questions before responding. You could do that in a variety of situations. It could be a different retirement scenario, giving it something more specific about the client you are working with or having it generate some initial ideas.

From a risk assessment perspective, you can help it generate a more dialed-in risk profile questionnaire to come up with. From a portfolio diversification perspective, you could paste in a client’s current holdings in addition to the risk tolerance questionnaire you have generated; it could help give some ideas on portfolio changes and other long-term strategies.

I think the main thing is to not just rely on its first answer as being correct unless you, based on your human expertise, can be the buffer to say, “Yes, that is correct” or “No, it is not.”