🧠 Designing Rapid, Low-Cost Experiments with ChatGPT

Experiments don’t have to be complex, costly, or time-consuming. You can design rapid, low-cost tests with ChatGPT using these three simple ingredients

As a behavioural scientist, I’ve designed complex experiments across finance, retail, tech, and government. But it wasn’t until I started baking sourdough that I realised: you will never get to improve the quality of your product unless people get to taste it.

Simple, quick and practical experiments are key ingredients !

👉🏾 My three ingredients for designing experiments:

1️⃣ Start with the outcome - Ask yourself: “What does a good experiment look like?” Should I design a dynamic or diligent experiment?

2️⃣ Prioritise speed over accuracy - 1% improvements compound over time and can be far more meaningful if it can be sustained over the long-term.

3️⃣ Make it practical - Add multiple variables if it improves outcomes for the people you serve.

Sample ChatGPT Input:

My outcome: 

The outcome I want to achieve is to support the financial wellbeing of low-income customers by helping them pay their credit card balances in full and on time, preventing high-interest charges.

Sample ChatGPT Output:

Experiment Card

Concept title: Pre-Commitment Prompt

Encouraging customers to commit to full payments before they start using their card.

---------

Test duration:
1–2 weeks

---------

What is your hypothesis?
If customers make a pre-commitment to paying their full balance, they will be more likely to follow through when the bill arrives.

---------

How will you test the idea?

During account activation, ask customers:
“Would you like to commit to paying your full balance each month to avoid interest?”

Customers can select “Yes” or “Remind me later.” 

Follow up with a personalised reminder when the first bill is due.

---------
What will you measure?

1/ Percentage of customers who make the commitment.
2/ Rate of full balance payments among those who committed vs. those who didn’t.
3/ Effectiveness of the reminder message in reinforcing behaviour.