Preparing Data for AI ingestion

AI features in Amplitude help you move faster, uncover insights, and take action. The quality of those results depends on how easily AI can consume and understand your data. You can get the best results possible by preparing your data before it reaches Amplitude's AI. The more context you give the AI, the better it performs. When your data is well-structured, AI interprets questions the way your team would, so you get from question to insight faster. For more information about Amplitude AI, go to Agents overview.

This page describes practical steps to strengthen your data so that Amplitude can deliver accurate, reliable insights. Most of these steps also make your data easier for your team to work with. Clear naming, clean taxonomies, and thoughtful governance improve the Amplitude experience for everyone.

Clean data unlocks better AI

Preparing your data for AI isn't a one-time task. As your product evolves, so does your taxonomy. Keeping your data clean requires ongoing attention.

When you define events clearly and maintain them consistently:

• Agents can clearly analyze and provide insight that's relevant to your business needs, not how a generic model interprets raw event names.
• Teams spend less time correcting or rebuilding AI-generated analyses.
• Questions move from ask to insight to action faster, because the first answer is trustworthy.

Align your data to the following best practices for data preparation:

• Make event and property names understandable.
• Clean up ambiguity in your existing data.
• Remove or hide unwanted data.
• Provide business context for your AI.
• Establish conventions to improve data over time.

Clean data doesn't improve reporting alone. It makes AI a tool your team relies on and strengthens your entire Amplitude experience in the process.

Make event and property names understandable

AI relies on your event and property names, display names, and descriptions to understand what your data represents. The clearer those are, the more accurately AI can match questions to the right events and properties. Making your event and property names clear and using the Description field to include context is the single most effective way to prepare your data for AI.

For example, an analyst at your organization asks "How often do customers browse categories from the website's navigation?"

If your data contains missing information such as:

AI may not connect catSelectClick to the concept of browsing categories. Without information within the Display name and Description fields, AI can't guess the context for how your team uses catSelectClick. It could return a result using the wrong event or report that it can't find a match. That erodes trust in Amplitude's AI and your analyst ends up doing the work manually.

However, if your data has complete documentation such as:

The AI can match the question to the right event, return an accurate chart, and suggest follow-up analyses, such as breaking down by category. The analyst spends their time using insights rather than trying to generate them.

Aligning event and property names

Start with your most-queried events and properties. Data Assistant helps you identify which ones matter the most. Then, update that subset of your events and properties:

• Add display names in natural language: Instead of abbreviations, rename your Display names to something easier for both the AI and humans to understand. For example, catSelectClick becomes Category Selected and pgVw becomes Page Viewed.
• Write descriptions that explain intent, not implementation: Update your Descriptions to give context for why the event matters:
  • The description "Fired on click handler for nav component" describes when the event occurs, but doesn't give further context.
  • The description "Triggered when a customer selects a product category from the navigation menu" gives AI the context it needs for your environment.
• Map coded values to human-readable labels: If a property value is sku_29881, AI can't interpret the context because that value doesn't carry any inherent information. That SKU could relate to anything. Use lookup tables to map these types of values to the actual description of a product or item. For example, sku_29881 could map to Women's BrandX Running Shoe.

This matters especially for properties used in group-bys or filters.

Categorize events: Organize your taxonomy to narrow the AI's focus. Go to Plan your taxonomy to design your categorization architecture. Having a defined taxonomy also greatly helps your human colleagues.

Amplitude uses display names and descriptions to match natural-language questions to the correct events. Without them, it can either select the wrong event or return a No Result message. With them, it resolves ambiguity correctly and your team can trust the output. These same improvements make dashboards more readable and reduce back-and-forth about what an event means.

Clean up ambiguity in your existing data

Ambiguity in your data occurs when multiple events capture the same action. Or, it occurs when differences between similar events aren't obvious. That causes confusion when your dashboards and charts report similar or identical information in different ways. It affects AI when it selects events to analyze because the AI might not understand how ambiguous data overlaps.

For example, your taxonomy has two events: played song and song played. Both events capture when a user streams a song from a playlist. Because they represent the same action, you can transform them into a single event to remove the duplication.

If similarly named events capture slightly different information, for example, played song tracks songs played through a website while song played tracks songs played through an app, transform them in a slightly different manner. The optimal method is to consolidate them into a single event that contains a property that distinguishes the source.

However, if your team either doesn't have the time or capacity to make the change, Amplitude recommends adding clear descriptions to the separate events that explain the difference between the similarly named events.

For example, you have two events: played song and song played. They both capture when a user streams a song from a playlist. To clean up ambiguity, transform the events to the same played song formatting. This removes the duplicate event because they capture the exact same information. However, if the similarly named events capture slightly different information (if played song captures songs played through a website while song played captures songs played through an app), add a description to the event so that you clearly explain how the two events are different.

This doesn't mean you must rename events purely for the sake of consistency. For example, if an event called played song has existed for years and is widely used, changing it to Played Song to align with title-case formatting could cause confusion if analysts aren't expecting the change. Clarity of meaning matters more than formatting. Plan to align formatting and naming conventions over time. That ensures consistency during later implementations.

Removing ambiguity in your data

• Audit for semantic duplicates: Search your taxonomy for events that describe the same user behavior with different names. Common patterns include legacy naming versus current naming, platform-specific variants (such as ios_signup instead of signup), and test events that no one removed after testing ended.
• Merge or transform where possible: If played song and song_play_event capture the same behavior, use Amplitude's merge or transformation features to roll them into a single event.
• When merging isn't feasible, add descriptions that disambiguate: Sometimes, two events that look similar capture entirely different actions. In this situation, use the Description field to state explicitly: "This event captures X. For Y, use [other event] instead." This helps AI and your team differentiate between the two events.
• Delete what's clearly unused: If an event hasn't fired in months and isn't referenced in any saved chart, you can likely delete it.

AI evaluates your full taxonomy when selecting events. Duplicate or near-duplicate events create false matches. The fewer irrelevant events the AI sorts through, the more consistently it picks the correct one.

This cleanup also helps your team. Analysts onboarding to a new project or domain can trust that the events are accurate without spending extra effort to confirm them.

Remove or hide unwanted data

If you wouldn't trust an event enough to build a dashboard around it, Amplitude AI shouldn't rely on it, either. Stale, test, and deprecated events don't just clutter your taxonomy, they introduce noise that reduces confidence in AI-generated analyses.

Removing unwanted data

• Hide or delete deprecated, stale, or test-related events: If someone created an event for a QA cycle or it's no longer instrumented, remove it from the visible taxonomy.
• For Enterprise customers: Automated Tasks removes stale events and identifies test data automatically. This shifts manual audits to a recurring, automated cleanup effort.

AI treats every visible event as a potential input when building an analysis. By removing unwanted or irrelevant data, you help it focus only on the events that matter. Fewer, higher-quality signals mean higher-confidence results. For your team, a smaller and more targeted taxonomy also leads to faster event discovery, cleaner dashboards, and less confusion when exploring data.

Provide business context for AI

Clear events and properties are only part of the picture. To be truly helpful, AI also needs to understand how your business works and how your teams define success.

When you share your revenue model, internal terminology, and how your team defines metrics such as "conversion," "activation," or "retention," Amplitude AI can interpret questions the same way your team would. That shared context ensures analyses reflect how your business operates, rather than relying entirely on raw event structure.

Providing context

Go to Project Settings > AI Controls and define:

• What your company does and who your users are: This helps AI interpret questions in the right domain. An "order" means something different for an ecommerce company, a restaurant delivery app, or a healthcare platform.
• Core metrics and how they're defined: What counts as activation? Conversion? Retention? A qualified user? Be specific. Include the events, thresholds, and time windows that your team uses.
• Internal terminology: If your team uses terms such as "workspace," "campaign," or "deal," define what those mean to your team. AI encounters these terms in user questions and maps them to the correct events and properties.
• Analysis preferences: Answer the following types of questions for the AI:
  • Do you typically look at weekly or monthly granularity?
  • Do you exclude internal users from analyses?
  • Do you have standard segments (such as free, paid, or by region)?

Business context acts as a foundation for every AI-powered interaction in Amplitude. From the Global Agent in Chat to other AI features built on your data, context about how you use Amplitude matters. Without it, AI is a capable analyst with no onboarding. With it, AI behaves like a team member who's been briefed on how your business works, what you care about, and how you measure it.

For more on configuring this context, go to AI Context.

Establish conventions to improve data over time

The steps above address the data you have now. The following section focuses on protecting your data structures as your implementation grows.

It's easier to instrument events correctly from the start of your Amplitude journey than to consolidate, merge, or clean them up later. AI performs best when your data patterns stay consistent and naming is predictable. Clear conventions reduce rework, prevent duplication, and help maintain data quality as you add new teams, features, and use cases. Aligning to these best practices also makes it easier for new team members and AI to understand your data model.

Establishing conventions

• Define naming conventions and document them in Amplitude Data's settings: Choose a format (for example, Title Case for events and snake_case for properties) and enforce it during implementation.
• Use the tracking plan proactively: Add event definitions before instrumentation begins. The tracking plan is your single source of truth and it's easier to name something correctly at the beginning of your effort than to merge or rename it later.
• Include naming standards in implementation reviews: Treat taxonomy hygiene the same way you treat code review. Catch issues before they ship.

Consistent naming reduces the number of disambiguation decisions AI has to make. When patterns are predictable, AI can confidently match questions to events, even for events it hasn't seen queried before. For your team, conventions also reduce rework, prevent duplication, and make it easier for new team members to understand your data model without a guided tour.