Obviously AI Tutorial
Last updated: April 2026
What you'll achieve
After this tutorial, you'll have built, trained, and deployed your first predictive model. You'll be able to upload a spreadsheet (like a CSV or from Google Sheets), select the column you want to predict (e.g., 'Customer Churn'), and let Obviously AI automatically build a model. You'll learn to interpret the model's accuracy, understand key drivers, and generate predictions for new data rows. Finally, you'll know how to export these predictions and set up a simple, shareable prediction app—all without writing a single line of code. This is the fastest path from raw data to actionable business insights I've found.
Prerequisites
- •A free Obviously AI account (use the 'Start for Free' button)
- •A clean CSV or Excel file with your data (e.g., sales records, customer list)
- •A web browser (Chrome, Firefox, or Edge) with a stable internet connection
Step-by-Step Guide
Step 1: Sign Up and Set Up Your Account
Head to obviously.ai and click the prominent 'Start for Free' button. In my experience, the sign-up is refreshingly simple—just your name, email, and a password. No credit card is required for the free tier, which is a huge plus for testing. Once you confirm your email, you'll land on the 'Create New Project' screen. Don't be overwhelmed; the interface is designed to guide you linearly. I recommend immediately clicking your profile icon in the top right and exploring the 'Settings' to connect a data source like Google Sheets if you plan to use it. This upfront setup saves time later. The platform will ask for your role and use case; be honest, as it tailors some suggestions.
Use a work email. The free plan's features are generous, but some integrations preview better with a professional domain.
Step 2: Navigate the Dashboard
The dashboard is your mission control. On the left, you'll see the main navigation: Projects, Predictions, Data Apps, and Monitoring. 'Projects' is where you build models. Clicking it shows all your past and current models. The central 'Create New Project' button is your starting point. What surprised me was the 'Templates' section—don't ignore it. Templates for Customer Churn, Sales Forecasting, etc., pre-configure settings and even provide sample data. I tested these when I started, and they're brilliant for learning. The top bar has notifications and a search function. Spend two minutes clicking around; the design is intuitive. The 'Predictions' tab is where you run new data through a trained model, and 'Data Apps' is for building shareable interfaces.
Bookmark the 'Monitoring' tab. It's crucial for seeing how your model's accuracy holds up over time in production.
Step 3: Create Your First Predictive Model
Click 'Create New Project'. You'll be prompted to upload data. I always use a simple CSV: one row per item (customer, sale, day), with columns for attributes and one column for the outcome I want to predict. Drag and drop your file. The platform will analyze it. Now, the critical part: selecting your 'Prediction Column'. This is the variable you want to forecast, like 'Churned' (Yes/No) or 'Next Month Sales'. In my experience, new users often pick a vague column. Be specific. Then, choose your problem type—classification (categories) or regression (numbers). Obviously AI will suggest it. Click 'Train Model'. The platform now does its magic: cleaning data, engineering features, and testing algorithms. Go grab a coffee; this takes 2-5 minutes.
Your prediction column must have enough historical examples. For 'Churn', you need many 'Yes' and 'No' entries for the model to learn.
Step 4: Interpret Results and Refine Your Model
Once training finishes, you'll see a results dashboard. The top number is accuracy, but don't fixate on it alone. Scroll to 'Model Insights'. This is where Obviously AI shines. It shows the top factors driving your prediction. For a churn model, it might say 'Days Since Last Login' is the #1 predictor. This business insight is often more valuable than the prediction itself. Below, check the confusion matrix (for classification) to see error types. If you're unhappy with accuracy, click 'Improve Model'. Here, you can try advanced algorithms (like XGBoost), adjust features, or add more data. I tested this extensively; often, just letting it run the 'Advanced' training for a few more minutes boosts performance significantly.
Aim for interpretable insights, not just high accuracy. A 90% accurate model you can't explain is less useful than an 85% one you can.
Step 5: Deploy and Generate Predictions
Your model is useless if it sits idle. Click the 'Predict' tab. You have two main options: 'Batch Prediction' to upload a new CSV file (e.g., a list of current customers) and get predictions for all rows instantly, or 'Single Prediction' for one-off queries. I use Batch constantly. Upload your file, map the columns if needed, and click 'Predict'. In seconds, you'll get a downloadable CSV with a new column—'Predicted Churn Probability' or similar. This is pure gold for operations. The other killer feature is 'Data Apps'. Click it, create a new app linked to your model, and you get a web form or a simple table interface you can share with colleagues. They can input data and get predictions without logging in.
For Batch Predictions, ensure your new data file has the same column names and formats as your training data to avoid errors.
Step 6: Explore Integrations and Live Monitoring
Don't stop at CSV exports. Go to the 'Integrations' section under your project. You can set up a Zapier zap to send predictions to Slack, Google Sheets, or Salesforce automatically. I have a live model that predicts support ticket escalation and posts high-risk tickets to a Slack channel—it's transformative. Also, revisit the 'Monitoring' dashboard. It shows your model's accuracy over time and alerts you to 'drift'—when real-world data starts to differ from your training data, signaling it's time to retrain. This is a pro feature most no-code tools lack. Finally, explore the API documentation if you're technically inclined. It allows you to call predictions from your own software.
Model performance decays. Check the Monitoring tab monthly. A drop in accuracy means your world has changed; retrain with fresh data.
Common Mistakes to Avoid
Uploading data with a unique ID column (like 'Customer ID') as a feature. The model will fixate on it, creating a useless, overfit model. Exclude IDs.
Choosing a prediction column with too few outcomes (e.g., only 3 'Yes' for churn in 10,000 rows). The model can't learn. Ensure a balanced sample.
Ignoring the 'Date' column type. If your data is time-series, explicitly mark date columns so the platform can extract trends (month, day of week).
Forgetting to validate. Always use the platform's built-in validation split. Never train on 100% of data without a hold-out set to test real accuracy.