ChatGPT Code Interpreter: how the new feature revolutionizes decision-making

I open ChatGPT Code Interpreter and type, “Why were sales down overnight?”.

Less than two minutes later, my virtual analyst has sent me 3 hypotheses each of which has supporting data and rationale.

1. Competitors reduced their prices yesterday, particularly for products competing with SKU152, resulting in lower sales in that category. This contributed roughly 25% of the variation.

2. Meta ads weren't as impactful as we predicted, because one of our creative assets performed poorly – the product we were hero-ing saw reduced demand. This contributed roughly 10% of the variation.

3. Overall demand within the category was down because of the unusually hot weather. This accounts for most of the variation.

Interesting. I write back to my virtual analyst: “Based on historical data, what’s the best way of reacting to the above challenges?”

• Stick to high prices for SKU152 to prevent margin erosion. This is optimal for long-term sales.

• The weather will be optimal on Friday; increase spend across Meta and YouTube by 12% from Thursday evening through to Friday night.

• Think about creating more brand-level creative as current ratios are suboptimal based on historic meta-analysis of campaign data.

“Thanks, ChatGPT,” I write.

Then, “Please draft and send the required prompts to alter the above and message the relevant personnel across the organization based on the org structure I shared previously. Despite your guidance, please reduce prices for SKU152 this weekend as I believe our competitors will stick prices back up. Please also adjust our budgets accordingly and ensure the exec dashboard reflects the change.”

Science or fiction?

You get the picture.

You probably also recognize that this isn't a new promise. You’ve seen tons of demos promising similar outcomes over the past 10 years, from consultancies to niche SaaS providers, agencies and large tech companies.

So you’re probably skeptical that this future above will ever become a reality. Perhaps you think that getting and structuring all that data in anything close to real-time is a pipe dream. Perhaps you think that the past can never really represent the future; that we can’t make business decisions like this based on data. Perhaps you’re concerned about how we’ll know when the AIs get it wrong and start spouting rubbish.

These are, of course, challenges. But we can overcome each of them with a little bit of thought.

What’s different this time?

There are plenty of small reasons and one big one: the power of free. I don’t mean free cost, but free effort.

In general, the impact of taking effort from low to nothing drastically increases adoption (psychologist Dan Ariely and others have shown this multiple times with pricing experiments).

Cognitive natural language processing means anyone (and I mean anyone) can ask the question of Code Interpreter in basic English. Even drag-and-drop tools (like Alteryx) require an intermediary step between desire and instruction; removing that step reduces barriers to entry from low to negligible.

So adoption will be absolutely rife in a way that existing tools simply haven’t seen. Plus, we have access to more data now than ever before. 10 years ago, I was hard-pressed to find basic census data in a usable format. The data landscape is far more beautiful now.

The impacts of Code Interpreter

Put it this way: I’ve already amended my hiring plan and job descriptions for the following 6 months, because the skills we need will shift drastically. We’ll need:

• More critical thinkers, fewer data churners: we can focus more on what information we need and why, not how we get it.

• More mathematicians, fewer coders: we can focus more on hiring people who have taken the time to understand Bayesian Probability, rather than how to code specific snippets in Python

• More generalists, fewer siloed departments: anybody can type in their mother tongue; not everyone can identify when data needs to be transposed in a certain way. The centralized team will become the ‘maths team’ not the ‘data team’.

• More strategists, fewer short-term tacticians: optimizing for the short-term is a job for AI, not for people.

At a wider operational level, it will change processes, people, technology and data architecture. And even further out, at a societal level, it should change what people study, how long they study for, how many days a week people work, and where they work from. But that’s an article for another day. For now, my message is hopefully loud and clear. Adopt, adapt, and rid yourself of any complacency.