3 Innovative Ways to Use AI in Domo: From ETL Documentation to Structured Extraction

Every organization has that one person who built the critical ETL pipeline three years ago, and nobody else understands how it works. In the Domopalooza session Did You Know Domo Could Do That? 5 Innovative Solutions Customers Are Building Today, Vishakha Shenoy, director of data science and AI services at Domo, and her team walked through five innovative ways to use AI in Domo that solve exactly these kinds of problems, converting clunky, unstructured inputs into governed, analyzable outputs.

The session covered AI agents and workflows, Magic ETL automation, structured extraction from text and images, AI readiness metadata generation, and automated migration to Cloud Integrations datasets. What stood out across all five examples: each one addresses a specific pain point that data teams actually face, and each one can be replicated by practitioners regardless of their technical depth.

Here are three of the most impactful approaches from the session.

1. Let AI document your ETL logic

An AI agent can auto-document complex dataflows by cleaning up the underlying JSON, interpreting the logic, and outputting readable markdown documentation. The resulting documentation includes a summary, inputs and outputs, grain and keys, trigger settings, a detailed walkthrough of formulas, risk checks for overcomplicated logic, and a high-level dataflow diagram.

Most Domo objects for your dataflows, Magic ETL or SQL, are stored on the back end as a 3,000 to 4,000-character JSON object. It's clunky and full of noise. You can't just hand it to an agent and expect useful results. The solution presented uses a Code Engine function called "extract Dataflow" to clean up that JSON before the agent interprets it.

The output gets rendered in an app from the Appstore, making ETL logic digestible for business stakeholders who know their data but don't want to parse SQL or trace Magic ETL tile joins.

To implement this pattern youreslf, follow these steps:

1. Use the "extract Dataflow" Code Engine function to clean the JSON definition of your Dataflow
2. Configure a workflow AI agent with a detailed prompt to interpret the cleaned logic
3. Output the documentation as a markdown file
4. Render the markdown in the Appstore app designed for this purpose
5. Review the generated documentation for summary, grain and keys, risk checks, and the dataflow diagram

This approach reduces key-person risk and makes onboarding faster when someone new needs to understand existing pipelines.

2. Turn unstructured text into analyzable columns

Product reviews, survey responses, and support tickets all contain valuable signal buried in natural language. The Text Generation tile with structured output can process multi-line text and extract specific topics and sentiments as separate, analyzable columns.

In this breakout demo, product reviews were processed using a prompt focused on four topics: fit, customer support, durability, and shipping. The build used predictability settings for consistent results and structured output to generate eight fields of four topic summaries and four sentiment scores.

The prompt design matters here. As noted in the breakout, if any of the reviews (in this example) don't relate to your four topics, instruct the AI to leave them blank. This explicit instruction prevents the model from hallucinating when a review doesn't mention a target topic.

Here's how to set this use case up:

1. Drag the Text Generation tile into your ETL
2. Define your target topics in the prompt (keep it specific—four to six topics works well)
3. Set the temperature slider toward predictability for consistent outputs
4. Configure structured output fields for each topic summary and sentiment
5. Include explicit instructions to leave fields blank when topics aren't present

The result is a dashboard where you can filter by sentiment, drill into specific topics, and identify patterns across SKUs or product categories without anyone manually reading through hundreds of reviews.

3. Extract product data from images with an agent

Images stored in FileSets (now called Documents) can be processed by a workflow-triggered agent to extract fields like brand, size, price, ratings, and reviews into a structured dataset. Once you have a repository of extracted products, newly uploaded images can be compared side-by-side, and a prediction model can estimate which features influence unit sales.

The session demonstrated this with product screenshots from Amazon. The agent extracted brand name, grill size, BTU, price, number of burners, reviews, and ratings all from the image. With that data structured, a model predicted sales based on those features and quantified how much each feature contributed.

The workflow has two paths: one for processing a newly uploaded image and one for processing all existing images in a FileSet. When an image is uploaded, the agent extracts the data, calls the prediction model, appends the results to a dataset, and ends. The agent can also generate competitive analysis and publish it to a dashboard or send it via email based on workflow triggers.

To build this yourself:

1. Store product images in FileSets/Documents
2. Configure a workflow that triggers on file upload
3. Set up an agent with tool sets for image extraction and prediction model calls
4. Define the fields you want to extract (brand, price, reviews, ratings, etc.)
5. Append extracted data to a dataset for comparison and analysis
6. Layer a prediction model on top to quantify feature impact on outcomes

This pattern works for competitive intelligence, catalog management, or any scenario where you need to turn visual data into structured analysis.

Put these approaches to work

These three examples share a common thread: they take something messy and convert it into governed, analyzable outputs. The ETL documentation agent reduces key-person risk. Structured extraction from text enables slice-and-dice analysis on qualitative data. Image-to-dataset workflows turn visual information into predictive inputs.

Each approach uses AI agents and workflows, Code Engine functions, or Magic ETL capabilities that you can configure today. The patterns are replicable, and the session walked through the specific implementation steps for each.

For the full walkthrough of all five innovative ways to use AI in Domo—including AI readiness metadata generation and automated Cloud Integrations migration—watch the complete session from Domopalooza.