International Planning and AI: The Right and Wrong Way to Use It

At first, it was almost comical how bad the bots were at coming up with answers to even simple questions. But they’ve gotten a lot better over time. Sometimes, it’s even possible to get a workable solution.

And yet, many so-called solutions are still very much overbuilt, not fit for purpose, and often miss important nuances that can make or break a strategy.

In our own practice, we use AI where we can. With one exception, the privacy settings of the most popular models are horrendous; we assume anything we post there is as private as a postcard.

For that reason, our ability to use it for client work is very limited.

(And although it seems to have stolen my penchant for em dashes (i.e. “—”), I don’t trust it to write for me either.)

But… it does help us with editing, running generalized scenarios, and for fact-checking.

In this week’s issue, I thought I’d share how we use it in the hopes of giving you some ideas on how to make it work for you too.

Obviously, we do not recommend implementing anything without having an experienced professional look it over. I know you know that, but still an important point.

For reference, we use a variety of models depending on the goal. That includes OpenAI’s ChatGPT, Mistral’s Vibe, Meta’s Llama, Proton’s Lumo, and occasionally Anthropic’s Claude. This piece is based on our experience with these. Other chatbots may respond differently.

I should also note that I leave memory off in most cases, meaning every new window is “factory-default” and not based on my previous interactions. If you leave memory on, these suggestions may play out differently for you. 

With that preamble, let’s jump in…

#1: Are you asking the right questions?

The quality of the answers we get seems to be based directly on the quality of the questions we ask. The problem is, there’s a bias in virtually all chatbots to answer whatever you pose without first clarifying why you’re asking or what you’re trying to do. If you ask a vague question, you might get a vague answer. Or one completely irrelevant to your needs.

The more specific the question, the better the response tends to be.

So, instead of: “What’s the best way to protect my assets?”

Consider something like:

“I’m a US citizen living in Alaska with a $2 million brokerage account and rental real estate, concerned about litigation risk. What are the tradeoffs between a domestic LLC, a domestic asset protection trust, and an international structure in my situation?”

The second question gives the bot a lot more useful information to work with.

#2: The first answer isn’t always the best one.

No matter what response I get from my initial inquiry, I rarely take it at face value. I go into “skeptical” mode and grill it on the topic and how it might apply to my situation.

Some questions I might ask include:

• What are the pros and cons of this solution? 

• What did you assume before giving this answer? 

• Who shouldn’t use this approach? 

• What’s the biggest risk or downside if I implement it as written?

• What would a critic of this approach say? 

• Are there simpler or cheaper ways to get the same result?

#3: Trust but verify.

AI “hallucinates” – confidently producing statements with no basis in reality. It’s better than it used to be, but it still happens.

When in doubt – which you should be most of the time – ask for credible sources and then verify that those sources say the same thing the bot is. You might be surprised how often the source and the bot’s claim don’t match.

Another approach is to play bots against each other – take the output of one and copy and paste it into the other, along with the request: “Please check this for factual accuracy.”

(And yes, I do say please and thank you to a computer, for the same reason Kant encourages people to be kind to animals. :))

No two models will match 100%, and when it comes to such topics, most will add a qualifier like “generally” or “typically” to hedge their bets. But there should be a lot of overlap. If there’s not, or one model corrects the other, there may be a mistake in there somewhere.

#4: GIGO

The output of a computer program can only be as good as the data fed into it. In computer circles, this is called, “garbage in, garbage out,” or GIGO.

The chatbots we use are informed by their training data taken from human-created work, warts and all. If the underlying data was flawed, out of date, or where distinct concepts were mixed up together, then the resulting data is more likely to be flawed, out of date, or conflated.

Several areas of our work suffer from GIGO, especially planning structures, residency and second passport programs, and certain advanced tax planning issues.

Trusts, for instance – many pre-2022 commentators couldn’t tell a living trust from a DAPT –  and that bad data contaminated training sets.

#5: Have a professional review before implementing

Already mentioned. Self-serving but true. At the end of the day, an AI chatbot can regurgitate its training data and whatever information it can scrape from the Internet. It has no real context. No experience. And bears no liability for whatever mistakes it makes that you run with.

Put simply, I would not rely on AI for the final word on planning advice.

Believe me, we’ve seen what happens when someone runs with an AI plan. Undoing it can be quite expensive and time-consuming.

And if it causes you to break the law, “AI made me do it” isn’t a valid excuse.

P.S. Have you run an idea through an AI chatbot and want a professional opinion before moving forward? Feel free to get in touch.