Amplitude AI agents: what they do and when to use them

What it is

Rather than one general-purpose chatbot, Amplitude documents several distinct agents scoped to specific jobs: a dashboard agent for building and explaining analyses conversationally, a website conversion agent focused on conversion and funnel questions, and a session replay agent that can summarize replay activity without a human watching every recording. A custom-agent capability lets a team define its own agent for a specific internal workflow rather than relying only on the pre-built ones.

Amplitude documents explicit AI controls and a stated privacy-and-security position governing what these agents can access and how their outputs are scoped - because an agent answering questions against product-usage and potentially revenue-bearing event data is a materially different privacy surface than a general-purpose AI assistant with no data connection.

When you need it

• A team wants faster self-serve access to charts and analyses without every question routing through whoever knows Amplitude's query interface best - a dashboard agent is built for exactly that bottleneck.
• A growth or marketing team has a recurring, narrow class of question (conversion-funnel diagnosis, specifically) that's well-suited to a purpose-built agent rather than general exploratory analysis.
• Session replay volume has grown past what any team can manually review, and a summarization agent can triage which sessions are actually worth a human's time to watch.
• An internal workflow is specific enough (a particular recurring report, a particular kind of anomaly check) that a custom agent built for that exact job outperforms a general-purpose one asked to do it ad hoc.

Implementation, done properly

1. Start with the pre-built agents scoped to your actual bottleneck rather than jumping straight to a custom agent - the dashboard, website-conversion, and session-replay agents cover common jobs, and a custom build only earns its complexity once a genuinely specific, recurring need is identified.
2. Review and configure the available AI controls before wide rollout - understand what data scope an agent can access, whether that includes PII-sensitive properties, and confirm the configuration matches your organization's actual data-handling policy, not just Amplitude's default.
3. Pilot with a small group and a narrow set of real questions before opening an agent org-wide - the fastest way to build (or lose) trust in an AI-assisted analytics workflow is the first few weeks of real usage, so treat the pilot as a genuine evaluation, not a formality.
4. For a custom agent, scope its instructions and data access as tightly as the job actually requires - a custom agent given broad access just in case carries the same governance risk as an over-scoped API key, for the same reason.
5. Set an explicit expectation with users about what these agents are good for (fast, well-scoped questions against known data) versus what still needs a human analyst (judgment calls, ambiguous causal questions, anything where the cost of a confidently wrong answer is high) - agents that quietly get treated as infallible are where trust breaks hardest.

How to verify it worked

1. Ask a pre-built agent a question with a verifiable, known answer (a metric you can also pull manually) and confirm the two match before trusting it for questions you can't independently check.
2. Test how each agent handles an ambiguous or out-of-scope question - a well-configured agent should ask for clarification or state a limitation rather than confidently answering something it can't actually determine from the available data.
3. Review the AI controls configuration directly (not just from memory of how it was set up) to confirm the data scope an agent can access still matches policy, especially after any change to your Amplitude project structure or property sensitivity classifications.
4. For a custom agent handling a recurring workflow, spot-check its output against the manual version of that same workflow periodically, not just at launch - a data or schema change elsewhere in the instance can silently degrade a custom agent's accuracy without any error being thrown.

Related: Amplitude warehouse metrics and the Data Warehouse MCP server · Amplitude MCP and AI enablement services · Session replay, defined