Sales Forecasting with AI: How ChatGPT, Claude, and Gemini Compare

Can Generative AI Actually Deliver on Sales Forecasting with AI?

Generative AI tools are being used for everything from writing emails to coding apps, but can they actually help forecast sales with real accuracy?

This case study compares how three top-tier AI models handle a sales forecasting task. ChatGPT, Claude, and Gemini are each prompted to forecast Tesla’s quarterly revenue using historical financial data to see how each model handles the data and assess the reliability of outputs.

The results reveal clear differences in reasoning, formatting, and forecast quality. It also offers a look at how these tools might (or might not) fit into your workflow for sales forecasting with AI.

The Method: Sales Forecasting with AI

For this test, we used Tesla’s publicly available quarterly revenue data. The data was uploaded in CSV format and cleaned. Missing values are either removed or replaced with zeros.

ChatGPT, Claude, and Gemini were each evaluated on the following tasks:

• Recognizing and interpreting the dataset of historical revenue performance.
• Visualizing the revenue trends in a graph.
• Using a statistical model to forecast revenue for the next 12 quarters (3 years).
• Explaining and displaying the results.

The prompts were kept consistent to allow a fair comparison. In each case, the model receives the same dataset and is guided through the same forecasting process using an autoregressive integrated moving average (ARIMA) model.

Forecasting with ChatGPT’s Data Analyst Tool

For this test, we used the Data Analyst tool available in the GPT-4o version. This analysis tool allows you to upload data files, generate data visualizations, and run forecasting code within the chat interface.

Step 1: Load and Visualize the Data

Once the Tesla revenue data was uploaded, ChatGPT recognized the columns and responded to a prompt to visualize quarterly revenue. But there was a small hiccup: the chart showed the most recent data on the left, which isn’t standard. 

A quick follow-up prompt fixed that by flipping the chart.

Step 2: Choose a Forecasting Model

Next, you could prompt ChatGPT to build a forecast. It suggested several approaches:

• Simple moving average.
• Exponential smoothing.
• SARIMA (seasonal ARIMA).
• ARIMA.

Since the data didn’t show a clear seasonal pattern, ARIMA was selected.

Step 3: Review the Forecast Output

The first ARIMA forecast removed the trend component, which made the projection less realistic. You might expect the model to retain the trend by default, but it needed more direction. 

After prompting it to include the trend again, ChatGPT updated the output, but the new forecast still didn’t align with the historical data.

ChatGPT Assessment

ChatGPT handled the mechanics well but required several rounds of prompting to produce a more usable forecast. If you’re using it in a real-world workflow, expect some back-and-forth to refine the output. It’s a capable tool, but you’ll still need to guide it carefully, especially when it comes to applying sound forecasting logic.

Forecasting with Claude

Claude 3.5 Sonnet offers a different kind of user experience. One standout feature is the Artifacts window, which displays code, charts, or files separately from your text chat. It keeps everything organized and makes working with visual outputs more intuitive. 

The same procedure used with ChatGPT was repeated for Claude.

Step 1: Load and Visualize the Data

Claude successfully processed the same Tesla revenue dataset and generated a chart using the Artifacts window. Formatting still took extra effort and adjustments, but the output looked more visually appealing than ChatGPT’s.

Step 2: Run the Forecast Using ARIMA

After confirming the model could interpret the data, Claude was prompted to generate a forecast using ARIMA with the trend component included. It responded by generating Python code and rendering a forecast in the Artifacts window.

Step 3: Evaluate the Results

Claude’s output had a critical issue: the model started the forecast from the beginning of the dataset, rather than projecting from the most recent quarter. 

As a result, the forecast appeared as a linear extension of the full dataset rather than a forward-looking projection with the revenue fluctuations you’d expect in a realistic forecast. Therefore, Claude’s forecast was linear and overly simplified, similar to what ChatGPT produced.

Claude Assessment

Claude delivered a stronger user interface and a more visually appealing chart but still struggled with forecasting logic. Like ChatGPT, it needed multiple prompts to produce something usable, and even then, the forecast required additional scrutiny.

Forecasting with Gemini

Gemini is designed to work flexibly across Google’s tools for tasks like document review and structured data analysis. While it doesn’t have a specialized data analyst tool or charting workspace, it can interpret uploaded data, generate charts, and respond to forecasting prompts within the chat.

Step 1: Load and Visualize the Data

Gemini received the same Tesla revenue data as the other models. After some initial difficulty, it displayed the chart correctly, with the most recent quarter on the right.

Step 2: Run the Forecast Using ARIMA

When asked to forecast future revenue, Gemini recommended the ARIMA model, the same as ChatGPT and Claude. One feature that stood out was its inclusion of confidence intervals, showing a lower and upper bound for each projected value. 

However, the forecast itself was flawed. Gemini placed the forecast incorrectly in the timeline and projected a sudden, unrealistic revenue drop despite no such trend in the historical data.

Step 3: Evaluate the Results

While Gemini included useful features like axis correction and confidence intervals, it produced a forecast that couldn’t be relied on without significant revision.

Gemini Assessment

Gemini handled some aspects of the task well. It correctly oriented the revenue chart and was the only tool to include confidence intervals in its forecast. 

Despite those strengths, the model produced an unrealistic revenue drop and failed to place the forecast in the correct part of the timeline. For this case study, the forecast was unreliable and would require significant corrections to be usable.

What This Case Study Reveals About Sales Forecasting with AI

After running the same sales forecasting task through ChatGPT, Claude, and Gemini, the results were clear: none of the models produced a flawless forecast, but each offered something valuable.

Here’s what stood out:

A Shared Limitation Across All Three Tools

Each model was able to run a forecast, but none showed real financial intuition or awareness of business context. Even when using the same data and model type (ARIMA), results varied depending on how well the prompt was interpreted and how much guidance was provided.

Can You Trust AI to Forecast Sales?

Generative AI tools can support sales forecasting but not without human input. Each tool needs human expertise to guide its logic, test assumptions, and review the results. ChatGPT and Claude showed the most promise but required hands-on direction to produce usable forecasts.

If you work in FP&A or any finance role that relies on sound forecasting, AI can speed up your process. It can help visualize trends, recommend model types, and even generate basic code. But it’s not a plug-and-play solution. 

As AI tools continue to evolve, the ability to ask better questions, identify errors, and connect forecasts to business realities will become even more important.

Ready to practice sales forecasting with AI? Explore CFI’s course Generative AI for Data Analysis: ChatGPT, Claude & Gemini Compared and walk through hands-on case studies just like this one.

Additional Resources

Why AI Makes Financial Modeling Skills More Valuable Than Ever

Preparing Financial Data for AI: Best Practices for Accuracy & Machine-Reading

Generative AI in Finance: Use Cases and Applications

See all Financial Modeling resources

See all AI resources