Using the dashboard Agent

The dashboard Agent is a conversational interface to capabilities that already exist in the dashboard. It can inspect operations, run Explore Analyses, answer custom data questions, and prepare state-changing work for approval. Its answers link back to the relevant page so the result remains inspectable.

Inspect operations

The Agent can list and inspect organization-scoped:

• Robots;
• Extract Runs;
• Schedules;
• Enrichments;
• Imports.

Use the resource's name when possible. A page context such as "this Run" can resolve an identifier, but the Agent does not silently use an unrelated scope.

Analyze product data

The Agent has two analytical paths.

Packaged Analyses

For Price position, Price movements, Availability changes, or Assortment overlap, the Agent runs the same versioned contract as the dedicated Explore page and public API. It preserves the Analysis status, limitations, summary, coverage, freshness, and dashboard URL.

The chat response includes at most 20 evidence rows. When more evidence exists, the Agent says so and links to the Analysis page rather than presenting the preview as exhaustive.

Explore AI

For other factual questions, the Agent can query:

• Captures for raw Extract output, optionally scoped to one Run or Import;
• Items for normalized page-grain Enrich output, optionally scoped to one Enrichment;
• Explore for questions across connected Products, Variants, Listings, Offers, Reviews, and Stores.

The Agent first loads the data model for that surface, then generates and runs one bounded read query. The persisted result retains the question, generated SQL, rows, truncation state, and a link to the corresponding dashboard Query artifact on its AI page. Results contain at most 100 rows and are not pageable.

Use a packaged Analysis when its stable definition answers the question. Use Explore AI for exploratory groupings, counts, and evidence that do not have a packaged contract.

Prepare operations for approval

The Agent can prepare:

• a Robot build for a domain and country explicitly supplied by the user;
• an Extract Run using a selected existing Robot;
• a recurring Schedule using a selected existing Robot;
• a Run or Schedule that first builds a Robot when the supplied domain has no matching Robot;
• an Enrichment for an existing completed Extract Run;
• an Extract Run followed by Enrich when both budgets are explicit.

Before proposing work, it resolves the relevant existing resources and keeps the domain, country, URLs, budgets, cadence, and Enrich choice visible. It does not discover or infer a new domain that the user did not supply.

Imports are read-only in the Agent. Create or retry them from their dedicated dashboard or API surface.

Approval boundary

Preparing an operation does not execute it. The dashboard displays the exact steps and inputs, and state-changing work starts only after the user selects Approve. Rejecting or superseding a proposal does not consume Extract or Enrich credits.

Interrupted approved work can be retried from its operation state. The workflow keeps the approved inputs and avoids duplicating a completed step.

Ask precise questions

Use the product terms that match the required grain:

• Capture for one extracted or imported page record;
• Item for one normalized page-grain Enrich result;
• Product for a canonical family;
• Variant for one exact cross-store configuration;
• Listing for one Store SKU;
• Offer for one seller, price, and availability observation.

Examples:

• "Which Captures in this Run have no brand?"
• "Group these Items by category and brand."
• "Compare Nike price position for this Store in the US in USD."
• "Show current available Offer count and median price by Store and currency."
• "Prepare a 100-Capture Run for the confirmed Nike US Robot, then Enrich up to 100 Captures."

Boundaries

The Agent cannot bypass organization isolation, ClickHouse's read-only role, subscription access, execution limits, or dashboard approval. Explore AI does not expose arbitrary physical or internal tables and cannot write data.

The underlying dataset contains observed ecommerce product evidence. The Agent must qualify or reject conclusions about sales, demand, traffic, margin, market share, sentiment, or complete market coverage when those facts are not present.

See Exploring data for the four Analysis contracts and Access surfaces for the dashboard, API, export, Query, and Agent boundaries.