Guide to trusting Mixpanel AI answers

docs.json

@@ -630,6 +630,7 @@
                         ]
                       },
                       "guides/guides-by-use-case/empower-your-team/headless",
+                      "trust-ai-answers",
                       "guides/guides-by-use-case/empower-your-team/close-strategy-execution-gap",
                       "guides/guides-by-use-case/empower-your-team/see-replays"
                     ]

trust-ai-answers.mdx

@@ -0,0 +1,107 @@
+---
+title: "Set up Mixpanel AI for accurate answers"
+sidebarTitle: "Trust AI Answers"
+---
+
+Any answer AI gives you is only as good as the data behind it. So how do you know you can trust what Mixpanel’s AI tells you? This guide walks through what actually shapes an accurate answer and the process to validate AI insights before you act on them.
+
+Mixpanel's AI shows up in a few different ways — [Agent](/guides/guides-by-use-case/empower-your-team/mixpanel-agent), [MCP](/guides/guides-by-use-case/empower-your-team/mcp), and [Headless](/guides/guides-by-use-case/empower-your-team/headless) — ranging from a chat-style assistant inside the product to a connection an outside AI tool can query directly. This guide applies across all of them and assumes the reader is familiar with [Mixpanel reports](/docs/reports) and [event tracking concepts](/docs/what-is-mixpanel#concepts).
+
+## What determines answer accuracy
+
+Good data begets good answers. Two admin-controlled settings shape how accurate an AI response is, putting your team in the driver’s seat.
+
+#### Business Context
+
+Ask Mixpanel Agent how activation is trending, and it needs to know what activation means at your company before it can answer well. That's what [**Business Context**](/docs/business-context) is for. It covers things like:
+
+- Company background
+- Customer segments
+- Key dashboards
+- Team norms
+
+Every Mixpanel AI tool reads your Business Context automatically, so filling it in helps every answer that follows.
+
+<Info>
+  If your organization has both org-level and project-level Business Context, Mixpanel AI will use both of them together. Should there be a conflict between the two, project context takes higher priority.
+</Info>
+
+#### Verified Mode
+
+A team with five different event names for the same signup flow has a decision to make before letting AI loose on it: which one actually matters. [**Verified Mode**](/docs/mixpanel-agent#verified-mode) makes that decision explicit, guiding AI toward events, metrics, behaviors, and cohorts your team has reviewed in [**Lexicon**](/docs/data-governance/lexicon).
+
+When you query with Verified Mode enabled, Mixpanel prioritizes verified equivalents when they exist. When Business Context and a verified event point in different directions, the Agent weighs both rather than following a fixed order, and will often tell you there were two options and why it chose the one it did.
+
+[**AI-Powered Data Governance**](/docs/data-governance/ai-powered-data-governance) can help speed up Lexicon cleanup. It finds near-duplicate events and properties automatically, and lets you merge them in bulk once approved.
+
+<Tip>
+  Before turning on Verified Mode, check whether the events in your most-used queries are already verified in Lexicon. If your highest-traffic events aren't verified yet, verify those first so that turning on Verified Mode will immediately prioritize the data your team relies on most.
+</Tip>
+
+## Scope what the AI tool can do
+
+It’s one thing to trust an AI’s answer. It’s another to trust its authority to act.
+
+#### Control your own read-write access
+
+Most MCP-compatible AI tools, such as Claude, let you set permissions per connector. Tools are typically grouped as read-only or write and delete, and each group can be set to:
+
+- Always allow
+- Require approval
+- Block
+
+Scoping your Mixpanel MCP connection to read-only — so that it can only answer questions and never change anything — lives entirely on the AI tool's side, independent of anything configured in Mixpanel.
+
+#### If you allow write access
+
+When you allow write access, MCP may be able to do things like:
+
+- Create boards
+- Edit metrics
+- Set up experiments
+
+That means a change can happen without a person clicking through the UI themselves. For a full list of available tools, see [what's possible through MCP](/docs/mcp).
+
+<Tip>
+  If Lexicon cleanup has been sitting on someone's backlog, MCP can help clear it fast. Ask the Mixpanel MCP to generate descriptions for events that don't have any, or to surface likely duplicate event names for review. Preview the proposed changes before committing them at scale.
+</Tip>
+
+## Check the answer in front of you
+
+Business Context and Verified Mode lower the odds of a wrong answer. They don't guarantee the specific one in front of you is right. Always do a quick check on any specific answer before acting on it.
+
+#### Spot-check the mapping
+
+Always open a report or investigate a finding that an AI tool returned, and confirm that the finding matches the data.
+
+For example, when building manually in Mixpanel, you'd easily notice a cohort that comes back incorrectly with zero users because of a stale or renamed property. An AI-generated answer deserves the same scrutiny, not automatic trust just because AI produced it.
+
+<Warning>
+  Artificial intelligence can behave in ways that are inaccurate, incomplete, or unreliable despite guardrails. AI-generated insights in Mixpanel are designed to inform human decision-making, not replace it. Users remain in control of how insights are interpreted and acted upon.
+</Warning>
+
+#### Use RCA Agent's confidence and rating
+
+[RCA Agent](/guides/guides-by-use-case/empower-your-team/mixpanel-agent#rca-agent) surfaces a confidence level for each finding, signaling how much to trust a given finding. You can also rate the RCA board's accuracy yourself after reviewing it, which improves future results.
+
+#### Look for rendered charts in MCP
+
+Look for MCP responses that render an actual interactive Mixpanel report, rather than the AI describing or redrawing the data itself. The interactive Mixpanel report is the same report you'd see inside Mixpanel directly, not the AI's interpretation of it.
+
+## Common questions
+
+A few questions come up often enough to answer directly, outside of configuration or spot-checking.
+
+- **Does Mixpanel train on customer data?** No. For that, along with [GDPR](https://mixpanel.com/legal/mixpanel-gdpr/), [CCPA](https://mixpanel.com/legal/mixpanel-ccpa/), and similar compliance questions, see the [AI compliance page](https://mixpanel.com/legal/ai-compliance/).
+- **What if an Agent answer is consistently off?** The feedback buttons inside the Agent experience enable you to provide feedback to Mixpanel's AI team. That's a way to improve the product over time, not a fix for the answer in front of you right now — for that, spot-check it using the section above.
+- **Which AI model powers this?** See the [Mixpanel Agent FAQ](/docs/mixpanel-agent#faq) for information on the current model that powers the Mixpanel Agent. [Mixpanel MCP](/docs/mcp) doesn't run on a set model; it connects to whichever AI tool you're already using, including Claude, ChatGPT, Gemini, Cursor, or Notion.
+
+## Key takeaways
+
+- Business Context and Verified Mode are the settings you control that shape answer accuracy most directly.
+- You control whether your AI tool's connection to Mixpanel MCP is read-only or if it can write, through that tool's own connector permissions, independent of Mixpanel's settings.
+- Spot-checking a specific answer, using RCA Agent's confidence and rating, or looking for a rendered chart in MCP catches what configuration alone won't.
+
+## Next step
+
+Before utilizing Mixpanel’s AI tools, write [Business Context](/guides/guides-by-workflow/write-business-context). Mixpanel’s AI tools read it automatically, so it improves every answer that follows, not just the one in front of you.