LiveRamp AI Capabilities and Best Practices
LiveRamp AI helps you find relevant data assets and build segments using natural-language prompts, all within Connect.
Note
LiveRamp’s AI-Assisted Segment Builder is currently in limited release and is available to interested customers (depending on suitability).
LiveRamp AI Capabilities
If you’ve been enabled for LiveRamp AI, the following capabilities are available on the LiveRamp AI page:
Find an asset: You can use LiveRamp AI to search your datasets, tables, views, and built segments using natural language. You can also locate recently-updated or recently-built assets, as well as review asset details.
Build a segment: You can use LiveRamp AI to build a segment from your first-party, second-party, and Data Marketplace data by describing your goals or other criteria in natural language. LiveRamp AI uses permitted asset metadata to propose segment rules, which you can review and edit. If you’ve been enabled for Data Marketplace data, LiveRamp AI can also suggest Data Marketplace segments for you to include in your segment. You can also view the size of the draft segment and split the segment into test and control groups before building the segment. Once the segment has been built, you can then activate the segment to your desired destinations.
Ask a Connect docs question: You can use LiveRamp AI to ask a question about using Connect that will be answered by referencing the Connect documentation. Answers will list the source articles that were used in responding to your question. Your specific account information will not be accessed to answer these questions, so the answers will be general and not tailored to your specific situation.
For more information on how to use these capabilities and how they work, see the sections below.
Note
The following capabilities are not currently available on the LiveRamp AI page:
Account-specific troubleshooting: LiveRamp AI cannot diagnose why an upload, segment build, match, or distribution failed, stalled, or produced an unexpected result.
Log and event investigation: LiveRamp AI cannot inspect ingestion logs, processing history, error details, job status, recent actions, or system telemetry.
Data-result validation: LiveRamp AI cannot examine raw records, file contents, field values, match rates, quality-gate results, or row-level discrepancies. The AI does not analyze underlying data or record-level values.
Account configuration and entitlement checks: LiveRamp AI cannot determine whether a product, feature, destination, permission, contract entitlement, region, or account setting is enabled.
Operational monitoring: LiveRamp AI cannot provide authoritative, real-time answers such as “Is my distribution queued?”, “When will my file finish processing?”, or “Why don’t I see this asset yet?”
Remediation and administration: LiveRamp AI cannot retry jobs, reprocess files, change permissions, update configurations, repair destination connections, or otherwise fix an account issue.
LiveRamp AI Segment Building Best Practices
When building a segment with the AI-Assisted Segment Builder, the following best practices can help you be more successful and effective:
Include the essential audience elements:
The audience or starting population.
The behavior, attribute, or interest that defines membership.
Thresholds or values, when relevant, such as time windows or date ranges.
Exclusions, such as employees or internal test accounts.
The names of any specific dataset, table, or existing segment, if known. Otherwise, use the business terminology represented in your metadata.
Make the intended logic explicit:
Use clear include and exclude language.
For complex audiences, describe which criteria must all be true and which alternatives are acceptable.
Plan complex statements before prompting. In the visual builder, rules at the same level use the same AND or OR conjunction, and rule groups can be nested up to two levels.
Build iteratively:
Start with the core audience definition, then use follow-up prompts to add or revise criteria.
If the builder offers multiple approaches or suggestions, read each summary and inspect the rules before selecting one. If neither option is appropriate, rephrase the request rather than selecting the closest result.
When a proposed rule does not match the request, edit the rule directly in the visual builder or describe the correction in a follow-up prompt.
Review and validate the draft segment before building:
Review the proposed dataset, tables, segments, fields, operators, values, inclusions, exclusions, and rationale.
Confirm that each rule expresses the intended business meaning. Pay particular attention to field types and operators; for example, a string field should not be assumed to support a numeric-style range.
Calculate the draft size after the rules are stable, and recalculate after every rule change. A single-dataset size is shown as deduplicated records; a multi-dataset size is shown in RampIDs. For more information, see “How Segment Builder Calculates Segment Size".
If you need a test/control split, configure it during the build. The AI-Assisted Segment Builder supports two parts, and a segment cannot be split after it is built.
Accessing LiveRamp AI
If your organization is enabled for LiveRamp AI, you can access it by selecting LiveRamp AI from the navigation menu in Connect.
You can also open the regular Segment Builder and enter a prompt in the AI chat field at the top of the page, then press Enter or select the up-arrow button. After you submit your initial prompt, the AI-Assisted Segment Builder opens.
How LiveRamp’s AI-Assisted Segment Builder Works
LiveRamp’s AI-Assisted Segment Builder in Connect turns natural language instructions into segment rules utilizing the LiveRamp Asset Management System.
Note
LiveRamp’s AI-Assisted Segment Builder is currently in limited release and is available to interested customers (depending on suitability).
The LiveRamp Asset Management System captures only metadata, such as table names, table descriptions, field names, field descriptions, dates, etc) and RampIDs that meet a specific criteria after a segment is built.
Note
For Snowflake activation customers, the only metadata captured is the metadata that is exposed through Snowflake directly (such as table names, table descriptions, field names, field descriptions, dates, etc.).
The AI portion of the segment builder is powered by a general‑purpose Gemini model used for inference only—we do not train on customer data. The AI only analyzes metadata (such as table/field names and descriptions, field types, date ranges, and other schema context) that exists in LiveRamp’s Asset Management System. With a “human-in-the-loop” approach, you have to review and approve the logic shown in our visual segment builder before anything is built.
For more information on how the AI-Assisted Segment Builder works, see the sections below.
For instructions on using LiveRamp’s AI-Assisted Segment Builder, see “Build a Segment”.
Data the AI Uses
The AI-Assisted Segment Builder only analyzes metadata that exists in your account LiveRamp’s Asset Management System. It does not analyze the underlying data. The AI-Assisted Segment Builder was designed to operate on structural context (the aboutness of data) rather than row‑level content.
See the table below for specific information on what data the AI-Assisted Segment Builder uses and what it does not use:
Used by the AI | Not Used by the AI |
|---|---|
Table names and table descriptions | Record‑level values (such as actual email addresses, phone numbers, or names) |
Column and field names and descriptions | Full customer profiles, purchase details tied to a person |
Data field types (string, date, numeric) and formats | Any sensitive identifiers (such custom IDs) |
Date ranges | Free‑text values from your rows |
Dataset and table tags and permissions | Enrichment or join keys in their raw values |
Metadata Used in an Example Prompt
Let’s look at an example where your segment building prompt is “Shoppers of ‘Outdoor Apparel’ with ≥3 purchases in the last 90 days; exclude employees and internal test accounts”.
Here’s an example of the metadata in your account that the AI-Assisted Segment Builder might utilize:
Tables names: “orders”, “product_category_map”, “customer_flags”
Field names and field types: “order_date” (date), “order_total” (decimal), “category_name” (string), “is_employee” (boolean)
Descriptions: "order_date" field: "UTC date order placed", “is_employee” field: "internal employee flag"
Dataset tags: “commerce”, “transactions”, “purchases”
Overall Segment Building Steps
Using the AI-Assisted Segment builder typically involves the following overall steps:
You describe the desired segment in natural language (such as, “High‑value shoppers of category X in the last 90 days; exclude employees”).
The AI-Assisted Segment Builder inspects permitted metadata from your data in the Asset Management system (schemas, field names/descriptions, types, and dataset tags).
The AI-Assisted Segment Builder creates a draft segment based on the metadata and displays those in the Visual Segment Builder, taking into account the permissions for that data.
You review the draft segment (rules and rationale) and edit before building, if necessary (editing can be done directly or you can describe your change in natural language).
You tell the AI-Assisted Segment Builder when you're ready to build the segment.
The approved rules run using your data in Connect, and creates a new segment that is registered in the Asset Management system and appears in the “Built Segments” area of the Segments page.
Once the segment has been built, you can use it to perform overlaps or to split the segment into multiple parts (for test and control, for example). You can also distribute the segment to your desired destinations. All Actions and changes are logged for review in the Activity History tab of the details page for that segment.
For detailed instructions, see “Build a Segment”.
Data Access and Security Model
The AI-Assisted Segment Builder works on the following principles:
Metadata‑only reasoning: AI-assisted segment building is permissioned to read schema metadata (names, descriptions, data types, tags, last‑updated dates). It is not permissioned to read raw rows or PII values.
Least‑privilege access: Visibility is limited to only those datasets/schemas you have explicitly configured within Connect. When datasets are removed, access is lost.
No PII egress: The AI-Assisted Segment Builder does not export or transmit your record‑level data to the AI for learning purposes. Prompts are constructed from metadata and your natural‑language request, and the Gemini inference model is used without reinforcement.
Account isolation: All reasoning and rule proposals occur within your account boundary in Connect, aligned with your role‑based access controls.
Logging and traceability: Prompts (metadata context only) are held in the user’s chat history. Creation details of the segment are held within the Segment Details in the LiveRamp Data Catalog.
Data retention: The AI-Assisted Segment Builder stores the final rules/logic, not your data. Retention of segment objects is at your direction within the platform.
Model use and training: The AI-Assisted Segment Builder uses a general‑purpose Gemini foundation model for inference only. We do not train or fine‑tune on your data, and we do not use your customer data or metadata for model training. Prompts to the model contain metadata‑only context (no PII). Product-level telemetry is collected for improvements to the user experience. For more information, see the section below.
Model Use and Training
The Gemini model used by the AI-Assisted Segment Builder works on the following principles:
Foundation model used: The AI-Assisted Segment Builder calls a general‑purpose Gemini model to translate your natural‑language prompt into proposed rules.
Inference‑only: The segment builder does not train or fine‑tune models on your datasets.
No training on your data: We do not use your customer data or metadata for model training. Gemini is used as a runtime inference service.
Metadata‑only prompts: Requests contain only permitted metadata (schema/field names and descriptions, types, tags) and your plain‑English query—never PII.
Cross‑tenant isolation: There are currently no cross‑customer learning capabilities. General telemetry data like thumbs up/down are utilized to improve the product experience.