Analytics
Analytics is where you see what your assistant actually did. It records how visitors rated its answers, every call it made to an AI model, every call it made to your API, every run it did in the background and every message your guardrails blocked. Every page in the section only reads, so nothing you do there changes your setup or what your visitors see.
Use it when something went wrong in a conversation, when you want to know what a turn cost, or when you check that your guardrails and toolkits behave as intended. Open the section from Analytics in the sidebar.
Pages in this section
Section titled “Pages in this section”| Page | What it records | What you find there |
|---|---|---|
| Feedback | Thumbs-up and thumbs-down ratings visitors gave answers in the widget | The reason and comment, the AI action behind the rated answer, and a triage status and notes for your team |
| AI actions | One row per call to an AI model, and one row per conversation turn | Provider, model, status and duration of each call; models, attempts, tokens and estimated cost of each turn |
| Tool runs | One row per attempt to call your API through a toolkit | Endpoint, status, failure kind and retries; the request and response of each run |
| Agent jobs | Background runs started by a webhook, a schedule or Run now | Status, owner, what each job waits on, its steps, approvals and events |
| Violations | Visitor messages your guardrails blocked | The message, the reasons, the reply the visitor got and the policy version |
Four of these lists open a detail page for one record: Log Details, Tool Execution Log Details, Agent job and Violation details.
What you can do here
Section titled “What you can do here”- Find the answers visitors disliked. Filter ratings to thumbs-down and read the reason next to the AI action that wrote the answer. Feedback · Review and improve answer quality.
- See why a reply was slow, failed or cost more than expected. Follow a conversation turn call by call and attempt by attempt, with tokens and an estimated cost in US dollars. AI actions.
- Find out why an action in a conversation did not work. The runs show the exact request to your API, its response and whether the call was refused, timed out or never sent. Tool runs.
- Watch your automations. See which background runs are running, finished or waiting for a customer’s approval. Agent jobs.
- Check your guardrails. See what visitors tried and whether ordinary questions get blocked. Violations.
Concepts to know
Section titled “Concepts to know”Turns, calls and attempts
Section titled “Turns, calls and attempts”One visitor message is a turn. To answer it, the assistant makes one or more calls to AI models, for example one that runs a tool and one that writes the reply, or calls by sub-agents. Each call is made of one or more attempts: when a provider fails, StrikeUp retries or falls back to another model, and every try is an attempt. The Activity tab of AI actions groups attempts into calls and calls into turns.
Calls made by StrikeUp’s own stages, such as the safety check of your guardrails and the planner, are not shown on AI actions. A message that the safety check blocks appears on Violations instead.
How the records connect
Section titled “How the records connect”| From | You can reach | How |
|---|---|---|
| A rating on Feedback | The AI action that wrote the rated answer | Shown in the rating’s detail panel |
| An AI action | The tools whose results it received, by name | Search for the tool name on Tool runs |
| A conversation turn | Each call and attempt | View call in the turn dialog |
| An agent job | Its steps, approvals and events | The job’s own page |
| A violation | Nothing else | Violations record no visitor, conversation or widget |
What is kept, and what is masked
Section titled “What is kept, and what is masked”- Ratings, calls to AI models, tool runs and violations are kept indefinitely. StrikeUp does not delete them automatically.
- The only automatic deletion is agent-job events. The events of a finished agent job are deleted once they are more than 7 days old; the job, its steps and approvals stay.
- Masking applies to your API’s responses only. Your masking rules mask the responses of your API before they are stored and before the AI sees them. Visitors’ messages, tool requests, endpoint addresses, blocked messages and visitor comments are stored as they were sent. The error text of a failed tool call reaches the AI unmasked.
- Credentials are never stored with a tool run, and the operating instructions given to the AI are never shown.
Estimated cost
Section titled “Estimated cost”Costs on AI actions and Agent jobs are estimates in US dollars, worked out for each AI call when it ran from the price listed for that model at that moment. Calls without a price add nothing. The estimate is not an invoice.
Availability
Section titled “Availability”Every workspace has the section, with no plan or feature flag. Every signed-in user can open all five lists. Opening a conversation turn on AI actions and opening a tool run need the Admin role, and so does changing a rating’s triage status. Your visitors never see any of these pages. None of the pages updates on its own; use the refresh button on each.