# Learn from conversations

> Understand the continuous cycle Holistics AI uses to learn from conversations, and how AI Chat Insights puts it into practice

Every conversation is a signal: when the AI misreads a dataset, lacks a metric definition, or gives an incomplete answer, that tells you exactly where your semantic layer needs work.

## AI Chat Insights

**AI Chat Insights** gives admins visibility into what the user was trying to accomplish, how the AI responded, and where gaps in context affected the quality of the answer.

Each conversation's insight is broken down into three parts:

- **Summary**: A concise overview of the conversation's topic and flow.
- **Recap**: A breakdown of the conversation's outcomes: which questions the AI handled well, and which it fell short on.
- **Suggestions**: Specific points where the AI lacked the context it needed, each with a recommended fix (for example, adding a missing metric definition, improving a dataset description, or updating stale data).

AI Chat Insights looks for the most appropriate, end-user-accessible place to make the fix: starting with a dashboard, then a dataset, then a model, and falling back to an [AI Skill](/docs/ai/skills) or [AI Context](/docs/ai/context/overview) if nothing more specific applies.

To access this feature, go to **AI Monitoring → Conversations dashboard** and click on a conversation to open its AI Chat Insights. Click **Apply** on any suggestion you want to act on. Review the proposed change in Development, adjust it if needed, and approve it.

<VideoPlayer src="https://media.holistics.io/learn-from-conversations.mp4" autoPlay={false} />

## Why this matters

AI answer quality is only as good as the context behind it. Without a way to see where the AI struggled, admins are left guessing, relying on "AI isn't good" complaints instead of evidence. AI Chat Insights closes that feedback loop.

Over time, this creates a compounding effect:

- The more users interact with the AI, the more gaps get identified.
- The more gaps get fixed, the more accurate the AI becomes. 
- The more accurate the AI becomes, the more users trust and rely on it.

## What makes our AI learning different

- **Insights become governed objects.** Learnings from conversations can turn into metrics, definitions, and documentation you can reuse.
- **Everything is directly improvable.** Because your models, datasets, and dashboards are all as-code, AI can read them directly and suggest improvements in place.
- **Customizable process.** You can adapt the insight-generation and implementation steps to fit your team's standards and review process.
- **Open to external agents.** External agents can generate insights too and persist them back to Holistics, so you can bring your own agent and still benefit from a shared, continuously improving knowledge base.

## How AI learning works

AI learning runs on a continuous 4-stage cycle:

1. **Observe**: collect signals from real usage: conversations, thumbs-up/down feedback, user interactions, and query errors.
2. **Reflect**: understand what's happening in those signals: where the AI succeeded, where it didn't, and why.
3. **Improve**: act on those findings directly, with suggested context updates you can review and apply without leaving the platform.
4. **Evaluate**: test how the AI performs against your updated context before publishing changes to production.

**AI Chat Insights** is the feature that powers the Observe, Reflect, and Improve stages: it surfaces what the AI learned from each conversation and turns it into actionable suggestions.

**Evaluate** is coming soon.