Pendo links product behavior, AI conversations and in-app guidance

Pendo combines product analytics with in-app guidance, feedback and tools for understanding conversational AI usage. Leo helps staff explore product data through questions, while Agent Analytics examines how customers use AI agents inside their products. The distinction matters: using AI to analyze software adoption and measuring an AI assistant’s actual performance require different data. This blueprint proposes an onboarding evaluation from public sources, without claiming to have tested Pendo.

01 / ProductTwo distinct AI jobs inside Pendo

The Leo documentation describes a conversational interface over Pages, Features, Track Events and account or visitor context. It supports questions about behavior and feedback, with visualizations and follow-up exploration. The documentation still labels Leo beta. Availability in a pricing bundle should not be read as a guarantee that every assistant behavior is a mature production capability.

Agent Analytics captures prompts or full conversations and groups use cases and issues. Full conversation capture can include model, tool and token metadata when supplied through the Conversations API. This gives the team a route from an aggregate issue rate to an actual exchange, rather than asking a satisfaction score to explain everything that went wrong.

02 / AudienceFor teams that own adoption and the product response

Pendo is most useful when a product organization can connect what people do to a practical intervention. A software company might find that customers repeatedly ask an assistant how to invite teammates. The relevant decision is then whether to improve the underlying interface, the assistant answer, the onboarding guide or all three. Pendo’s analytics and guidance span those conversations.

It is a less direct fit for a team seeking only model-serving infrastructure or a generic assistant for employees. Agent Analytics expects access to the browser context or agent backend. An inaccessible third-party iframe cannot be treated as an observable agent simply because it appears inside your application. Establish that implementation route before planning a conversation-level evaluation.

Amplitude is a useful comparison when broad product analytics and experimentation are the main requirements. Intercom belongs in the discussion when delivering customer assistance is the immediate job. Pendo can help understand behavior around a support or onboarding experience, while the support system handles the conversation and service workflow. Compare responsibilities and integration work rather than assuming one tool replaces the entire stack.

03 / WorkflowA proposed pilot for an onboarding assistant

Start with a single task: helping a new workspace administrator invite a colleague and assign the right role. Define completion as the intended invitation being created successfully, with a separate check for acceptance if that is part of activation. Opening a guide, sending a prompt and clicking an invitation button are intermediate actions. They should not silently become the outcome metric.

Choose the correct capture method. For a controlled application and agent backend, full conversation capture is the useful route when the team needs both prompts and responses. Browser-extension prompts-only capture answers a narrower question about what users ask. Label those datasets separately. Otherwise, an apparently low issue rate may simply reflect the absence of the response data needed to identify the issue.

Before collecting real conversations, establish redaction categories and handling of excluded visitors. Pendo’s AI documentation explains that Agent Analytics sends configured conversation data to its AI provider after the selected redactions. Use synthetic examples containing email addresses, workspace names and sensitive role descriptions to check the intended behavior. This is a proposed implementation check, not a claim that masking was tested here.

Link the observed interaction to the relevant account and product context. A user asking how to invite someone may lack administrator rights, may be on a plan without the desired role or may have already sent the invitation. Those situations need different answers. Avoid evaluating the assistant on a single idealized prompt that assumes all prerequisites are satisfied.

Inspect the emerging use cases and issues, then read representative conversations. Include successful task completions, repeated questions, unsupported requests and failed actions. Keep the original exchange with the reviewer’s explanation of the outcome. A polite closing message is weak evidence of success if the invitation was never created, while a terse response can accompany a completed task.

Use Leo to explore adoption around the task and compare cohorts. Scope the request to the application, segment and time window, and verify the resulting chart against an ordinary report. The documentation describes lenses for limiting the conversation’s scope; a new conversation starts on All data. Make deliberate scoping part of the workflow so an analyst does not accidentally compare one onboarding cohort with the entire installed base.

The proposed intervention might combine clearer permission messaging with an in-app guide at the invitation screen and a revised assistant response. Evaluate the changes against the original completion definition. Monitor support escalation and erroneous role assignments as well as usage. More assistant conversations can indicate interest, confusion or a broken interface; the product outcome distinguishes those possibilities.

04 / PricingKeep MAUs and prompt quotas in separate calculations

The pricing page lists Free, Base, Core and Ultimate with paid pricing determined by monthly active users and selected functionality. The Free plan supports up to 500 MAUs. Agent Analytics is presented with a separate prompt-volume basis, so a core product subscription and an AI conversation allowance are different dimensions of the commercial model.

The Agent Analytics guide says subscriptions without paid Agent Analytics can collect 500 prompts per calendar month. Collection stops at that limit until the monthly reset, while existing reports remain accessible. Paid entitlements follow the purchased quantity unless the contract specifies a temporary unlimited period. A pilot expected to exceed the allowance must arrange capacity before interpreting a partial month as complete observation.

Ask for a quote that specifies MAUs, prompt volume, capture method, retention and the modules required for the intervention. Core includes Session Replay in the public packaging; other capabilities vary by plan or add-on. Do not substitute a free analytics allowance for an assumption that every replay, discovery or orchestration feature is included. Novus appears as a separate open beta on the pricing page and is not required for the proposed workflow.

05 / DistinctionsWhat the combined adoption loop can reveal

Pendo’s interesting angle is the connection between product behavior, the questions users ask and help delivered inside the product. A traditional analytics report might show that invitation completion is low. Conversations can reveal that administrators misunderstand roles. A guide or interface change can address that misunderstanding at the moment it matters. The complete loop is more valuable than any one AI-generated summary.

This also supports a more disciplined discussion of agent value. Separate adoption from effectiveness, and effectiveness from cost. An agent can become popular while making the task slower. Another can reduce help requests without reducing successful completion. Record the business outcome alongside prompt counts so the team can recognize both cases rather than rewarding conversation volume itself.

The same approach helps product and customer-success teams talk about an account consistently. An apparent drop in engagement could reflect completed setup rather than churn risk. Business context and a clear definition of the expected usage pattern are necessary to interpret the event stream. AI makes that interpretation easier to request; it does not supply the missing product judgment automatically.

06 / QuestionsResolve capture, beta and access boundaries

Leo requires administrator activation and a supported data-center environment. The current guide lists US, EU, Japan and Australia environments, with additional regions planned. That is a product availability statement, not a conclusion that a particular organization’s data policy is satisfied. Check your own subscription settings and applicable arrangements before enabling the feature for sensitive workflows.

Pendo’s subscription settings guide separates AI access from the MCP server’s read and write tools. Exposing analytics to an external assistant introduces another processing context. Begin with the permissions needed for the actual analysis, and review any proposed action separately from reading a chart. A useful answer should not require granting broad write capability by default.

Conversation-level analysis also depends on what was retained. The Agent Analytics guide notes that agents configured never to store raw conversation data do not offer the Conversations views. That may be the right privacy choice, but it limits the available investigation. Design the evaluation around the chosen data boundary and resist interpreting an unavailable transcript as evidence that no issue occurred.

07 / DecisionChoose the outcome before choosing the assistant

Pendo deserves consideration when a team needs to understand product adoption and act on it within the software experience. For AI features, start with a task whose completion can be observed and an implementation that captures the evidence needed to judge it. Expand only when the combined analytics and guidance improve that outcome enough to justify their operating and commercial requirements.