> AI agents: this is one page from Mammoth Analytics documentation. The index of all pages as Markdown is https://mammoth.io/docs/llms.txt. Append `.md` to any docs URL, or send `Accept: text/markdown`, to get Markdown.

# Find the step behind wrong pipeline output

> When a Mammoth pipeline runs without errors but the numbers are wrong: step through tasks, check order, fix filter and pivot surprises, check AI tasks.

When every task runs but the result is wrong, step through the pipeline with **Preview Data at this Task** until the output stops matching what you expect. The cause is usually task order, an expression, or a column stored as the wrong type. For tasks that show a red error, see [Fix a failing step](https://mammoth.io/docs/learn/clean-and-shape/fix-a-failing-step/).

## Step through the pipeline to find the wrong step

1. **Step through the pipeline.** Use **Preview Data at this Task** on each task and compare it with what you expect.
2. **Check the task order.** Tasks run in sequence. A Conditional Filter after a Join filters the joined output, not each dataset before the join. Reorder if needed ([Edit, reorder and copy pipeline tasks](https://mammoth.io/docs/learn/clean-and-shape/manage-tasks/)).
3. **Check your expressions.** Work out the result by hand for a few rows and compare.

## Fix Group & Pivot and Conditional Filter surprises

| Symptom | Cause | Fix |
| --- | --- | --- |
| Sum or Average won't work on a column | It holds numbers as text | Add **Convert Column Type** before **Group & Pivot** |
| **Group & Pivot** shows an empty group | The grouping column has empty values | Filter those rows out first if they don't matter, or handle the group later |
| **Conditional Filter** keeps rows you expected to remove | The conditions need all to be true, or only one | Check the And/Or logic |
| A filter on numbers behaves oddly | The numbers are stored as text | Convert the column type first |
| `status = "active"` also drops empty rows | A condition excludes empty values too | Add a condition for empty values if you want to keep them |

## Check an AI task that misses your intent

If an AI task (an **Ask in plain words** task, a **Generative AI** task, or a rule written in a function's AI bar) runs but misses your intent:

1. Read its prompt and check the output on a few rows.
2. Make the prompt more specific. Start simple and add detail step by step.
3. Compare it with a manual task and keep whichever matches your logic.

Note prompts that work so you can reuse them.

## Prevent pipeline errors

- Filter early to reduce size and complexity ([Speed up pipelines and connectors](https://mammoth.io/docs/learn/help/performance/)).
- Check data types before calculations, and handle empty values on purpose. Don't assume there are none.
- Use conditions to route problem rows into a separate flow.
- Test each task on small, known data with edge cases (empty values, duplicates, extremes), and check it with **Preview Data at this Task**.
- Test the whole pipeline on realistic volumes and source variations. Use **Draft mode** to queue tasks and review them before they take effect ([How the pipeline works](https://mammoth.io/docs/learn/clean-and-shape/how-the-pipeline-works/)).
- Add a **Data Check** between major stages as a quality gate. It can halt the pipeline and block later tasks, or warn and continue, and notify the people you choose. When a check fails, its result shows how many checks failed: fix the data and run again, or change the check threshold. See [Remove duplicate rows and add data quality checks](https://mammoth.io/docs/learn/clean-and-shape/duplicates-and-checks/).
- Compare row counts at key stages, and review the Data Quality report after changes. See [Validation](https://mammoth.io/docs/guides/advanced/validation/) for patterns.

## Related

- [Fix a failing step](https://mammoth.io/docs/learn/clean-and-shape/fix-a-failing-step/): tasks with a red error.
- [Fix a join that returns zero or too many rows](https://mammoth.io/docs/learn/clean-and-shape/fix-join-problems/): key problems.
- [Data quality report](https://mammoth.io/docs/learn/explore/data-quality/): find nulls, duplicates and type problems before they break a step.

---
Source: https://mammoth.io/docs/learn/clean-and-shape/fix-wrong-output.md · Updated: 2026-11-08