Like everyone else, we’ve been thinking about how AI will shape our collective future, and we’ve been incorporating large language models like ChatGPT into our day-to-day work.
At Hoop, we meet for offsites a few times a year. A couple of months ago, the team was throwing around potential options and ideas for our next offsite, and it dawned on us that we could use ChatGPT as a robot-assist to our decision making process.
What started as a half serious suggestion turned into a full blown exercise. This is what we did and this is what we learned:
How to use AI to help with decision making:
- AI can help you set the stakes of your decision ahead of time so you know how much time and effort to spend on getting things right.
- AI can help you go BIG and generate a large list of ideas, some of which are bound to be different from the three or five options you might come up with on your own.
- AI can then help you come up with different narrowing criteria, which you can use in a systematic way to converge on key contenders from within your big list.
- Once you have a manageable number of options, present them to your team for async feedback and then, if you need to, meet to understand any areas of disagreement. After that it’s decision time!
What Kind of Decision Is This, Anyways?
We started by using ChatGPT to help orient ourselves to the stakes of the decision of where to hold our offsite. Is this something we should spend a lot of time on, or something we could decide quickly? Using some of what we learned last time around on using AI to clarify decision stakes, we determined that the choice of venue for this offsite would be a two-way door, or reversible decision.
The decision is more of a "two-way door" than a "one-way door." While the choice of location can influence the offsite's success, it is not a make-or-break decision, and there will be future opportunities to iterate and improve. It's essential to invest an appropriate amount of time in making this decision but not to the extent that it hinders progress in other critical areas.
This was a great way to start our process because it’s a good reminder that not every decision is pivotal. It also helped us later on when ChatGPT made some mistakes. We knew that the stakes were low enough that we didn’t have to spend time researching and correcting small errors.
Picking the Venue
Now that we’d considered how much time to spend on venue-selection, we leveraged GPT4 to generate a heap of options. The first 5 options an LMM outputs are probably gonna be pretty obvious (think top 3 results from Google) but if you ask it to generate 50 options, you’re going to get some things in there that you probably would not have thought of. So that’s what we did!
Here’s the prompt we used:
We need to stay in the United States for ease of travel and keeping the cost under control. Can you give me a list of 50 locations near airports that would make for a good team offsite location in late May?
In order to get those results down to a more manageable list, we considered a few different ways to cut down or think about the list. We then used GPT4 to create ranking based on our criteria, and then to create a combined ranking that put the whole puzzle together.
For the offsite, we looked at criteria including:
- Average temperature,
- Expected cost to host the offsite at each location,
- Expected travel distance based on each person’s home location,
- Ratings based on how ‘walkable’ each venue was,
- Which venues were close to bodies of water (who doesn't love an ocean or lake view??)
We did notice GPT4 making a few mistakes on these criteria, including some questionable judgments on more subjective criteria, like walkability, and some factual mistakes (pretty sure there are bodies of water in Minneapolis, 10,000 lakes is a lot of bodies of water). But here’s where our initial stakes-setting exercise helped us - we knew we didn’t have to spend a lot of time verifying the information we were getting from GPT because of the stakes of this decision. We did correct the lake thing though, I mean come on.
After all of the filtering, we ended up with 13 options. We asked GPT4 to rank our list based on total travel distance, the quality of their food scenes, and outdoor activity options, and then we asked chatGPT to create a combined ranking based on those vectors. Here is our prompt:
Combine the rankings for travel distance, estimated cost, food options, and outdoor activity options and give me a combined ranking for these 13 options.
And here is ChatGPT's ranked list:
- Chicago, Illinois
- Minneapolis, Minnesota
- Milwaukee, Wisconsin
- Washington, D.C.
- Cleveland, Ohio
- Pittsburgh, Pennsylvania
- Providence, Rhode Island
- Detroit, Michigan
- Baltimore, Maryland
- St. Louis, Missouri
- Richmond, Virginia
- Raleigh, North Carolina
- Buffalo, New York
And then as a final flourish, we removed options on the coasts to minimize jet lag, which gave us these four, final contestants.
- Chicago, Illinois
- Minneapolis, Minnesota
- Milwaukee, Wisconsin
- St. Louis, Missouri
We’d already done an offsite in Chicago, so off to Minneapolis we went. Thanks, large language model!
Facilitating the Offsite
Once we arrived in Minneapolis, we used GPT4 for a number of more superficial interactions. We used chatGPT’s new mobile app to randomly assign rooms from our airBNB, we used it to source ideas for session icebreakers, and when we went to Topgolf as a team activity, we used it to create balanced teams, including team names, based on our own assessment of each player’s golf experience (Swing Kings v. Birdie Brigade, if you were wondering 🥸).
What does it all mean??
From all of this, we learned a few things about leveraging chatGPT to plan an offsite, and to help guide a decision-making process more broadly.
First, use your model, leveraging a predefined set of prompts, to figure out how much time you should be investing in your decision making process. Is it a quick thing you can decide and reverse later if you have to? Or is it something you should spend a ton of time on because you only have one chance to get it all right. This is really helpful because later on in the process, the LLM can make mistakes - but if it’s a lower stakes decision, you can let some of these slide whereas if it were higher stakes, you might want to check the model’s work more rigorously.
Second, use GPT to brainstorm a lot of options. This is where the model can really shine and give you choices you might not have considered otherwise.
Third, use the model to help think through a set of filtering and ranking criteria, to help you winnow down your big list based on the things that matter most to you and your group.
Fourth, and finally, remember that the LLM is a tool you can use to inform divergent thinking, and to then converge on a final decision, but the model is just your partner in all of this. You should think critically as you go, and for our decision, once we had a final set of options, we turned those over to our small team to decide on.
Happy planning!