How I Use AI Agents as a Data Scientist in 2025

Image by Author

# Introduction

 
As data scientists, we wear so many hats on the job that it often feels like multiple careers rolled into one. In a single workday, I have to:

• Build data pipelines with SQL and Python
• Use statistics to analyze data
• Communicate recommendations to stakeholders
• Consistently monitor product performance and generate reports
• Run experiments to help the company decide whether to launch a product

And this is just half of it.

Being a data scientist is exciting because it’s one of the most versatile fields in tech: you get exposure to so many different aspects of the business and can visualize the impact of products on everyday users.

But the downside? It feels like you are always playing catch-up.

If a product launch performs poorly, you need to figure out why — and you must do so instantly. In the meantime, if a stakeholder wants to understand the impact of launching feature A instead of feature B, you need to design an experiment quickly and explain the results to them in a way that’s easy to understand.

You can’t be too technical in your explanation, but you also can’t be too vague. You must find a middle ground that balances interpretability with analytical rigor.

By the end of a workday, it sometimes feels like I’ve just run a marathon. Only to wake up and do it all again the next day. So when I get the opportunity to automate parts of my job with AI, I take it.

Recently, I have started incorporating AI agents into my data science workflows.

This has made me more efficient at my job, and I can answer business questions with data much faster than I used to.

In this article, I will explain exactly how I use AI agents to automate parts of my data science workflow. Specifically, we will explore:

• How I typically perform a data science workflow without AI
• The steps taken to automate the workflow with AI
• The exact tools I use and how much time this has saved me

But before we get into that, let’s revisit what exactly an AI agent is and why there is so much hype around them.

# What Are AI Agents?

 
AI agents are large language model (LLM)-powered systems that can perform tasks automatically by planning and reasoning through a problem. They can be used to automate advanced workflows without explicit direction from the user.

This can look like running a single command and having an LLM execute an end-to-end workflow while making decisions and adapting its approach throughout the process. You can use this time to focus on other tasks without needing to intervene or monitor each step.

# How I Use AI Agents to Automate Experimentation in Data Science

 
Experimentation is a huge part of a data science job.

Companies like Spotify, Google, and Meta always experiment before they release a new product to understand:

• Whether the new product will provide a high return on investment and is worth the resources allocated to building it
• If the product will have a long-term positive impact on the platform
• User sentiment around this product launch

Data scientists typically perform A/B tests to determine the effectiveness of a new feature or product launch. To learn more about A/B testing in data science, you can read this guide on A/B testing.

Companies can run up to 100 experiments a week. Experiment design and analysis can be a highly repetitive process, which is why I decided to try to automate it using AI agents.

Here’s how I typically analyze the results of an experiment, a process that takes around three days to a week:

1. Build SQL pipelines to extract the A/B test data that flows in from the system
2. Query these pipelines and perform exploratory data analysis (EDA) to determine the type of statistical test to use
3. Write Python code to run statistical tests and visualize this data
4. Generate a recommendation (for example, roll out this feature to 100% of our users)
5. Present this data in the form of an Excel sheet, document, or a slide deck and explain the results to stakeholders

Steps 2 and 3 are the most time-consuming because experiment results aren’t always straightforward.

For example, when deciding whether to roll out a video ad or an image ad, we may get contradictory results. An image ad might generate more immediate purchases, leading to higher short-term revenue. However, video ads might lead to better user retention and loyalty, which means that customers make more repeat purchases. This leads to higher long-term revenue.

In this case, we need to gather more supporting data points to make a decision on whether to launch image or video ads. We might have to use different statistical techniques and perform some simulations to see which approach aligns best with our business goals.

When this process is automated with an AI agent, it removes a lot of manual intervention. We can have AI gather data and perform this deep-dive analysis for us, which removes the analytical heavy lifting that we typically do.

Here’s what the automated A/B test analysis with an AI agent looks like:

1. I use Cursor, an AI editor that can access your codebase and automatically write and edit your code.
2. Using the Model Context Protocol (MCP), Cursor gains access to the data lake where raw experiment data flows into
3. Cursor then automatically builds a pipeline to process experiment data, and accesses the data lake again to join this with other relevant data tables
4. After creating all the necessary pipelines, it performs EDA on these tables and automatically determines the best statistical technique to use to analyze the results of the A/B test
5. It runs the chosen statistical test and analyzes the output, automatically creating a comprehensive HTML report of the output in a format that is presentable to business stakeholders

The above is an end-to-end experiment automation framework with an AI agent.

Of course, once this process is completed, I review the results of the analysis and go through the steps taken by the AI agent. I have to admit that this workflow isn’t always seamless. AI does hallucinate and needs a ton of prompting and examples of prior analyses before it can come up with its own workflow. The “garbage in, garbage out” principle definitely applies here, and I spent almost a week curating examples and building prompt files to ensure that Cursor had all the relevant information needed to run this analysis.

There was a lot of back and forth and multiple iterations before the automated framework performed as expected.

Now that this AI agent works, however, I am able to dramatically reduce the amount of time spent on analyzing the results of A/B tests. While the AI agent performs this workflow, I can focus on other tasks.

This takes tasks off my plate, making me a slightly less busy data scientist. I also get to present results to stakeholders quickly, and the shorter turnaround time helps the entire product team make quicker decisions.

# Why You Must Learn AI Agents for Data Science

 
Every data professional I know has incorporated AI into their workflow in some way. There’s a top-down push for this in organizations to make quicker business decisions, launch products faster, and stay ahead of the competition. I believe that AI adoption is crucial for data scientists to stay relevant and remain competitive in this job market.

And in my experience, creating agentic workflows to automate parts of our jobs requires us to upskill. I’ve had to learn new tools and techniques like MCP configuration, AI agent prompting (which is different from typing a prompt into ChatGPT), and workflow orchestration. The initial learning curve is worth it because it saves hours once you’re able to automate parts of your job.

If you are a data scientist or an aspiring one, I recommend learning how to build AI-assisted workflows early in your career. This is quickly becoming an industry expectation rather than just a nice-to-have, and you should start positioning yourself for the near future of data roles.

To get started, you can watch this video for a step-by-step guide on how to learn agentic AI for free.
 
 

Natassha Selvaraj is a self-taught data scientist with a passion for writing. Natassha writes on everything data science-related, a true master of all data topics. You can connect with her on LinkedIn or check out her YouTube channel.