---
title: "Create and manage extractors"
slug: "create-and-manage-extractors"
updated: 2026-11-12T14:20:38Z
published: 2026-11-12T14:20:38Z
canonical: "help.gong.io/create-and-manage-extractors"
---

> ## Documentation Index
> Fetch the complete documentation index at: https://help.gong.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Create and manage extractors

> **Who can use this**: Business admin
> 
> **Available on**: Gong Foundation

Create an extractor, test it, and publish it to turn a question about your customer conversations into structured data. Later, unpublish or delete an extractor you no longer need, or restart one that paused because your company ran out of Gong credits. Learn more about [how AI Data Extractor works](/v1/docs/ai-data-extractor).

## Create an extractor

**To create an extractor**:

1. From the left sidebar, click **Admin center.**
2. In the **Agent Studio** tab, hover over **AI Data Extractor** and click **Settings** to open the **AI Data Extractor management** page.
3. Click **+ Add new extractor** or ![Action Menu.svg](https://cdn.us.document360.io/2837c190-3ec8-4120-9116-6d427e774667/Images/Documentation/MoreActions.svg?sv=2026-02-06&spr=https&st=2026-11-13T00%3A31%3A20Z&se=2026-11-13T01%3A31%3A20Z&sr=c&sp=r&sig=EjQy5YSp3dA85jSWw9vnfq6vZp5lxqBPb2Dg3qAt4Wo%3D)> **Edit** to edit an existing extractor.
4. In **Field concept**, enter the question you want the extractor to answer.
5. Click **Refine** to get AI suggestions of how to improve the question so that the data extractor will be able to provide the best answers
6. Click **Replace** to replace your question with the suggested wording, or click anywhere on the screen to hide the AI suggestions.
7. In **Provide context, instructions, or examples to help the AI find the right information**, add detailed instructions to improve the quality of the answers. See [Tips for writing effective questions](/docs/create-and-manage-extractors#tips-for-writing-effective-ai-field-questions).
8. In **Target object**, select whether you want to save the answer in the Deals or Accounts object.
9. Click **Add Filters**, select the fields you want to filter by, and set a value for each filter. The available filters depend on the target object you selected.
10. In Sources, select whether you want to use calls or calls and emails to generate the answer for the extractor.
11. (Optional) Under **Narrow down further with filters**, click **Add filters**, choose the filters you want to apply to the conversations Gong analyzes, and click **Apply**.
12. Select how far back to look for data in relevant conversations.
13. In **Data type**, mark the types of answers you want to receive. You can mark multiple data types. Options are:
  1. Yes/No
  2. Text: Select whether the text should be:
    - **Short**: Answers the question with no additional explanation
    - **Detailed**: Full answer including explanations and additional information
    - **Bulleted list**: Answer in bullet format.
  3. Single-select picklist
  4. Multi-select picklist
  5. Number
  6. Date
  7. Range: The range is set from 1 to 5 by default. Adjust the range according to your requirements.
14. In **CRM mapping**, select the CRM field to update the answer in. Imported CRM fields and custom object attributes appear in the list. Text fields can be mapped to rich text fields and long text fields. If your target object is **Deals**, this field is mandatory. If your target object is **Accounts**, this field is optional.

> [!NOTE]
> Note:
> 
> It’s possible to map multiple Data Extractors to the same CRM field. In this case, the extractors all update the field and overwrite the existing value.
15. Click **Save** if you are not ready to publish the extractor and want to continue making changes. A name and description is generated automatically for the extractor. Edit them as necessary.

### Test the extractor

Test the extractor to see that you are getting accurate answers to your question before publishing.

1. Click **Test it**.
2. Select the account or deal you want to test the extractor on.
3. Check the results. The results include a value for each data type selected, together with an explanation of why the result was chosen.
4. If necessary, refine the question or context and click **Run test again**. You may have to do this several times before publishing.
5. Click **Change** to run the test on a different account or deal.
6. (Optional) In the **Credits** section, click **Calculate** to see the estimated consumption. **Note:** If you change the filters or sources afterward, click **Recalculate** to update the estimate. Learn more about [estimated credit consumption](/v1/docs/ai-data-extractor#estimated-credit-consumption).
7. Click **Publish** when you’re satisfied with the answers.
8. In the Publish extractor? window, review the estimated consumption, and click **Publish extractor**.

## Unpublish or delete an extractor

Unpublish an extractor when you want it to stop updating when new relevant conversations occur, but keep the data it has already generated in Gong. Existing data in Gong remains available, and data previously synced to your CRM remains in the CRM.

Deleting an extractor removes the extractor and its extracted data from Gong. Data previously synced to your CRM remains in the CRM. Delete an extractor when you no longer need the extractor or its existing data in Gong.

If you want to stop future updates while keeping existing data in Gong, unpublish the extractor instead.

**To unpublish or delete an extractor**:

1. From the left sidebar, click **Admin center** and select **Agent Studio**.
2. Hover over AI Data Extractor, and click **Settings**.
3. Locate the extractor you want to unpublish or delete.
4. Click the **More actions** ![Action Menu.svg](https://cdn.us.document360.io/2837c190-3ec8-4120-9116-6d427e774667/Images/Documentation/MoreActions.svg?sv=2026-02-06&spr=https&st=2026-11-13T00%3A31%3A20Z&se=2026-11-13T01%3A31%3A20Z&sr=c&sp=r&sig=EjQy5YSp3dA85jSWw9vnfq6vZp5lxqBPb2Dg3qAt4Wo%3D) icon, and choose one of the following:
  1. **Unpublish**: Review the confirmation dialog and click **Unpublish**.
  2. **Delete**: Review the confirmation dialog and click **Delete**.

![Table displaying publication status, last modified dates, and action options for items.](https://cdn.us.document360.io/2837c190-3ec8-4120-9116-6d427e774667/Images/Documentation/data extractor.png?sv=2026-02-06&spr=https&st=2026-11-13T00%3A31%3A20Z&se=2026-11-13T01%3A31%3A20Z&sr=c&sp=r&sig=EjQy5YSp3dA85jSWw9vnfq6vZp5lxqBPb2Dg3qAt4Wo%3D)

> [!WARNING]
> Warning
> 
> If you delete the extractor, the action removes its extracted data from Gong and can’t be undone. If you want to stop future updates while keeping existing data in Gong, unpublish the extractor instead.

## Restart a paused extractor

A paused extractor stays paused until you restart it. After your company has Gong credits again, restart it so it resumes running.

**To restart a paused extractor**:

1. From the left sidebar, click **Admin center**.
2. In the **Agent Studio** tab, hover over **AI Data Extractor** and click **Settings**.
3. On the paused extractor, click the **More actions** ![Action Menu.svg](https://cdn.us.document360.io/2837c190-3ec8-4120-9116-6d427e774667/Images/Documentation/MoreActions.svg?sv=2026-02-06&spr=https&st=2026-11-13T00%3A31%3A20Z&se=2026-11-13T01%3A31%3A20Z&sr=c&sp=r&sig=EjQy5YSp3dA85jSWw9vnfq6vZp5lxqBPb2Dg3qAt4Wo%3D) icon, and select **Edit**.
4. Click **Publish**.
5. In the Publish extractor? window, click **Publish extractor**.

## Tips for writing effective questions

AI Data Extractor is designed to extract specific, concise answers to questions that should have an objective correct answer. The types of questions you ask and the guidance that you provide directly affect the quality of the results.

### General

- **Use natural language. Ask, don’t prompt**: Write the way you’d speak to a new colleague with no prior context. Example: “Which competitors did the customer mention during this deal?”
- **Be specific**: Ask clear questions about the conversation. Example: “Did the customer request a proof of concept or trial?”
- **State your intention**: Explain the goal behind the question. The same question can have different intended outcomes. Example: “Which competitors were mentioned? Focus only on companies that were considered as an alternative option for conversation intelligence"
- **Clarify terms**: Define terms that are ambiguous or jargon.

### Questions

- Ask one question per extractor. Results will be less accurate if the question is compound or attempts to answer multiple things at once.
- Only ask questions that can be answered from the content of a conversation. Questions that rely largely on metadata won’t be answered accurately. Example: “What stage is the deal currently in?” Stage is a CRM field. It isn’t mentioned in the conversation, the AI can’t infer it reliably.

### Provide context, instructions, and examples

**Use examples to guide Gong AI**

Include one or more clear examples in your instructions. Examples show the model exactly how you expect the answer to look and help it stay consistent. Providing examples reduces variability and improves accuracy.

**For single and multi-select picklist fields**

- Provide guidance and definitions on when each option should be chosen, particularly if the options use company/industry specific language that may not be specifically found in a conversation. This helps the model differentiate between similar options and understand nuances that may not be specifically obvious.
- When providing extra context, include common variations or alternate phrasing that might refer to a specific picklist option.
- If there is an “Other” option you can map a text field for this extractor and then collect those outputs there. The text field is updated even when the answer isn’t Other, so your reporting logic should only consider the text field if the picklist is set to “Other”. You won’t be able to constrain the Text field output to a single/limited word output so you may not be able to aggregate this field but it can be used to identify new options to add to the picklist

**For multi-select picklist fields**

- Explain whether the extractor should select all applicable options or only the most relevant options.
- Clarify whether options are genuinely distinct or can be selected together. Overlapping definitions can lead to inconsistent selections.
- Provide examples that show how to select multiple values when several options are mentioned in the same conversation.

**For numeric fields**

- Clearly specify the type of value you are looking for, such as number of seats or headcount, percentages, or time units.
- If the conversation includes a range, explain how the Data Extractor should handle it. For example, indicate whether it should capture the minimum value, maximum value, or an average.

**For date fields**

- Clearly specify the type of date you want, such as a contract start date, renewal date, decision deadline, or target go-live date.
- Explain how to handle relative dates mentioned in the conversation, such as “next quarter” or “end of the month.” Specify whether it should convert them to an exact calendar date and which reference date to use, for example the call date.
- If multiple dates are mentioned, explain which one to capture. For example, choose the earliest confirmed date, the latest agreed date, or the final committed date.
- Clarify how to handle tentative or conditional dates, such as “hopefully in March” or “if legal approves by Friday.” Indicate whether to capture them or ignore anything not confirmed.

**For Boolean (Yes/No) fields**

- Include guidance for when each option should be chosen
- Provide the appropriate default answer
  - For questions seeking confirmation of a completed action or a positive state for example, "Was the payment successful?", the default is false.
  - For questions seeking confirmation of an ongoing status or the absence of a negative state, for example, "Is the user still eligible?", the default is true.
- The model will likely not adhere to any guidance given on output formatting

**For range fields**

- Clearly describe what the minimum and maximum values represent. For example, specify whether you want a budget range, expected number of seats, or implementation timeline.
- Specify the type of value you are looking for, such as currency, number of seats or headcount, percentages, or time units.
- Explain how to handle ranges expressed in different ways. For example, indicate whether "between 100 and 200," "100–200," and "at least 100" should all be treated as valid ranges.
- If multiple ranges are mentioned, explain which one to capture. For example, choose the final agreed range rather than an earlier estimate.
- Explain what to do if only one value is mentioned. For example, specify whether the value should be used for both the minimum and maximum, or whether no range should be returned.
- If you're using a range to score or rank something, define the criteria for each score. For example, explain what qualifies as a 1, 2, 3, 4, or 5 so the AI can apply the full scoring range consistently.

A team member who manages the business entities in Gong, such as deal and forecast boards. [Find my Business admin](/v1/docs/find-your-technical-or-business-admin).

Gong’s core platform for gaining advanced visibility into your revenue operations. Also the prerequisite license needed for Gong’s add-on packages. Access to features depends on your [company’s plan and your assigned seat](/v1/docs/plans-and-seats).