Metabase

Someone types a question the way they would say it out loud and the built-in assistant returns an answer or a working query against your data, without needing SQL or a dashboard someone built first. An assistant that only suggests or completes SQL the user is already writing is partly; one that writes the query from a plain-English description is yes, including inside a code editor.

The assistant holds the thread, so you can say “now split that by region” and it knows what you are referring to.

The answer exposes the query or analytical steps used to derive it. Named fields belong to field citation; upstream dependencies belong to lineage.

Named tables, columns or metrics attached to the answer, such as a field list or a citation shown with it. The text of a generated query is credited to the analytical-steps row, not this one, even though it names tables. Naming only the dataset does not count.

A user can give the product a CSV or spreadsheet and use the AI to analyze its contents.

The AI can clean or reshape data and leaves transformation steps, code or another inspectable artifact behind.

The AI joins or blends at least two sources, for example two databases or a database and an uploaded file. Tables in one database are one source, and analyzing an uploaded file on its own is the file-upload row.

The AI selects and creates a chart or table suited to the question, rather than only changing a visualization the user chose.

yes: the built-in assistant breaks a change down into contributing segments or factors and shows the contribution of each, with the calculation a person can check, for example a table of segments with their share of the change. A key-driver function the assistant is documented to use counts; one only a user calls in SQL does not. Naming likely causes, or describing that a number rose or fell, without contributions is partly.

The AI creates a forecast and exposes its horizon, inputs and assumptions so the result can be challenged. A forecast that does not expose its assumptions is partly. The AI itself must produce the forecast, so a classic forecasting feature it does not drive does not count, and a horizon with a confidence range but no stated inputs is partly.

A user can define more than one what-if scenario with the AI, the changed assumptions of each scenario are stated explicitly, and the AI calculates results for each that can be compared side by side. A single what-if answer, or scenarios without stated assumptions or calculated results, is partly.

The answer exposes the age or last-refresh time of the data it used, so stale output is visible at the point of use. Freshness shown with the answer through the data source's own indicator, for example a "last refreshed" badge on the dataset the answer used, is partly.

The assistant can read the calculation behind a chart or table in BI content: its formula, measure or the query that feeds it. Reading SQL in an editor, worksheet or notebook cell that no chart in BI content uses is partly.

The assistant writes or changes the calculation behind a chart or table in BI content. Writing SQL in an editor, worksheet or notebook cell is partly.

An outside agent can retrieve, through the vendor's API, CLI, MCP server or skills, the calculation behind a specific chart or table in BI content. Downloading a whole file or workbook it must then parse is partly.

An outside agent can change the calculation behind a specific chart or table in BI content through the vendor's interface. Styling need not be writable. Replacing a whole file or workbook is partly.

The assistant can open existing BI content and work from what it shows. Analyst work only (a notebook, SQL worksheet or SQL file that people edit rather than view) is partly.

The assistant builds persistent BI content with an owner, permissions and a live governed data source. Analyst work only (a notebook, a SQL file) is partly.

An outside agent can read the contents of existing BI content (its charts, queries and filters) through the vendor's interface. Analyst work only is partly. Listing what exists is the agent-discovery row.

An outside agent can create BI content through the vendor's interface, the way it would open a pull request. Analyst work only is partly.

yes: AI-created BI content can be shared with named users or groups and each viewer sees only the data their own permissions allow (results evaluated as the viewer). Sharing that shows the author's results to everyone is partly. An asset counts as AI-created only if the AI built it, not if a person pasted AI-generated SQL into it.

Permitted users can browse or search across the workspace for saved analytical assets created by colleagues, including AI-created work.

An assistant someone configures once, with its own instructions, data and permissions, that colleagues can then reuse.

The product’s AI or an agent can rerun a defined analytical task on a schedule or trigger and deliver a fresh result or alert without a person present. An ordinary report or dashboard schedule that the AI or an agent does not run does not count, and an AI that only sets up such a schedule is partly.

yes: a built-in agent can, on its own, cause a named side effect outside the analytics content: send a message or notification, call a webhook or another system, or write data to a destination. Writing to the product's own dashboards or notebooks does not count. Generating code a person must then run is partly.

yes: a built-in agent that cannot finish a task stops and hands it to a person, with the hand-off documented (a notification, a review queue, a pause awaiting a decision). Asking the user a clarifying question during a conversation is partly. Silence is nothing found.

A customer can configure one agent to delegate a bounded analytical task to another named agent with explicit tools and permissions. Hidden implementation subagents do not count.

The vendor documents attaching an outside capability to its own AI. Shipping an MCP server is the opposite direction and a different row.

When a column is renamed, a table changes or a shared definition is edited, the product's AI or a built-in agent identifies the BI content affected and repairs it or flags it for an owner. yes: both finding the affected content and repairing or flagging it are documented for the AI. Impact analysis a person runs, with no AI involvement, does not count; the AI fixing one broken chart a user points it at is partly.

The assistant can see the app that exists, its code, its layout, what it renders.

The assistant creates a persistent first-class app with an owner, permissions and a live governed data source. Disposable generated HTML does not count.

Claude Code or a script can pull the app's definition out through an API.

The same interface, authored from outside the product.

A saved set of questions with expected answers, a rubric, a model judging the output, or a run you can repeat after a change, applied to the built-in assistant's answers. A thumbs-up button collects opinions and does not count.

yes: an admin or user can choose a specific named model version for the built-in assistant and it stays on that version until changed. Choosing among models or providers without a fixed version is partly. The vendor managing the model with no customer choice is no. Provider choice for data location belongs to KEY, not this. Identical answers are not required.

The product preserves the inputs, configuration, model and analytical logic needed to reproduce or audit an answer. Byte-for-byte model determinism is not required. Kept chat history with the generated query is partly; yes needs the inputs, configuration, model and logic. A record that keeps the model, the prompt and the generated query but not the data snapshot or configuration is partly.

The documentation says a correction a person makes to a built-in assistant answer changes the assistant’s later answers, and says how far it reaches (that user, a team, everyone using the data). A correction documented to affect later answers with no stated reach is partly. A feedback thumb that only rates an answer or files a ticket is not this.

When a question is already answered by certified BI content or a governed metric, the product makes the assistant use that same definition, or tells the user when its answer departs from it. yes: the documentation says the assistant is bound to or checked against certified content or governed metrics, and that a mismatch is prevented or flagged. Using the semantic layer when one exists, with no statement about certified content or mismatches, is partly (that alone is the shared-definitions read row, M_IR). A "verified" badge on some answers with no rule tying answers to certified definitions is partly.

A limit that actually blocks the built-in AI's usage once it is reached. Reporting on spend without blocking anything does not count.

Customers can measure response time on their own workload by agent, model or evaluation run. Marketing estimates and generic timeouts do not count.

Per-user or per-team attribution of AI usage. A single tenant-wide total is not this.

The product records what the built-in assistant did, the prompt or query it ran and the objects it touched, attributable to the user who asked. A general product audit log that does not distinguish assistant activity is insufficient.

A documented control that turns AI features on or off for a subset of the tenant. An account-wide on/off with no granularity is partly; a setting that only governs which model is used is MODELPIN, not this.

yes: the built-in assistant's change to governed BI content or shared definitions is held until a named person or role approves it, or a policy routes the assistant's write to an approver who is a different person or role from whoever asked. A prompt shown only to the requester is partly. A draft the same user later publishes, or a general publishing workflow the assistant is not documented to enter, is partly. A user accepting or rejecting the assistant's suggestion to their own work is not approval.

The customer controls where inference occurs. A model running in the customer's cloud, warehouse or tenant earns yes; a documented choice of region or provider endpoint, or a bring-your-own-key path where the documentation says where the provider processes the request, earns partly. Vendor-hosted inference that the documentation says the customer cannot choose or move earns no. Supplying a key does not by itself say where inference runs, and documentation that only states where processing happens, with no customer choice, belongs to the residency row. A choice that applies to only some of the product's AI features is partly.

yes: the documentation states that the built-in assistant answers under the asking user's identity, so existing row-level and column-level rules apply. Dataset-level or object-level permissions only, or row and column rules that some AI surfaces documentedly bypass, are partly.

yes: names the services that receive AI prompts or data, what they receive, and how long it is kept. Missing any one of the three is partly. A subprocessor list is not required; one published only on a separate trust or legal site is outside this documentation, so its absence here is nothing found, not no.

yes: the product documentation states the regions where AI requests are processed and where AI data is stored, for all of the product's AI features. A subprocessor or trust page counts if it is part of the product documentation. Regions for some AI features only, or processing without storage, is partly. Terms published only on a separate trust or legal site outside this documentation are not evidence, so their absence here is nothing found, not no.

A statement that customer data is not used to train AI models by default. An opt-out you have to find is weaker and should be partly.

yes: an administrator can enforce, across the workspace or tenant, a limit on which data the assistant may use, separate from user permissions. A scope the author of each agent sets, or a scope the documentation says is not a security boundary, is partly.

yes: an outside agent can act under an identity or token scoped narrower than the human's own access (read-only, limited objects or limited tools). Acting as the full human identity is partly. Shared broad credentials do not count.

yes: an outside agent can follow a number to its upstream sources and transformations through a documented lineage interface. Reading the SQL of a view or model and inferring its sources is partly. A link from an answer back to its underlying source is partly. A description that names no upstream source or transformation does not count.

The product records actions taken by an outside agent, tied to its delegated identity and affected analytical objects. A general user activity log is insufficient.

yes: an outside agent's change to governed BI content or shared definitions is held until a named person or role approves it, or an MCP or agent policy routes the agent's write to an approver who is a different person or role from whoever started the agent. A prompt shown only to the requester is partly. A draft the same user later publishes, or a general Git pull-request workflow not documented for agent changes, is partly.

A reviewer can see a before/after diff of changes an outside agent made to a governed analytical asset. A generic publish or change diff is partly; yes needs the documentation to show it covers changes made by an outside agent.

A change made by an outside agent to a governed analytical asset can be reverted or restored.

A vendor-supported MCP server gives outside agents access to the customer’s analytical objects or data, with documented authentication and no customer-built middleware. A vendor-supported server the customer operates, or one in preview, earns partly. A server built by the community or by customers, or one that only searches documentation, does not count.

A documented way to enumerate objects, not just fetch one by id. Reading a named dashboard is a different row; this is finding out the dashboard is there.

A named credential type for programmatic access. Whose permissions apply is a different row; this is how the agent gets in at all.

A vendor-published skill, plugin or instruction pack (for example a SKILL.md or an agent plugin) that an outside coding agent loads to use the product well. Documentation of an MCP server alone is the MCP server row, not this.

A vendor-supported command-line tool whose documentation addresses use by AI or coding agents, including non-interactive authentication. A general CLI with no agent guidance is partly.

A documented admin control over agent or MCP access, separate from ordinary user permissions. An account-wide on/off is partly.

A limit on agent-driven usage (rate limit, quota, spend or compute cap) documented for the agent interface and separate from ordinary user limits. A general API rate limit that applies to agents is partly.

Somewhere to write down what the schema does not say: which customers count as active, which of the three revenue tables is authoritative, what finance means by bookings. The AI reads it before answering.

The business context lives in files you can diff, review and roll back.

Context attached to a workspace, team, audience or model rather than the whole tenant.

A named validation, preview, staging or test step for context. Version control alone is C5.

The specific instruction, description or rule surfaced with the answer.

Context ingested from the warehouse or a modelling tool, not retyped in the product.

yes: a product setting that removes a table, field or answer from the AI's reach, or marks it deprecated so the AI avoids it, specific to the AI. An instruction in the AI's context telling it not to use something is partly. Ordinary data permissions are AIRESTRICT or GOVBIND, not this.

yes: the documentation names who owns or approves business context (a role or person) and describes a review or approval step for changes to it. A named owner role with no review step, or a review workflow with no named owner, is partly. Edit permissions alone are not this.

When your definitions live in dbt or a warehouse semantic view, the assistant uses them.

The assistant authors something that lands below the BI tool, a transform, a model, a view.

yes: an outside agent can read warehouse-level logic (view definitions, dbt models, semantic views) through the product. Table and column names or types alone are not logic and earn nothing found.

An agent outside the product can author warehouse logic through it.

The product consumes definitions you already maintain in dbt, Cube or Snowflake semantic views, instead of requiring you to rebuild them here.

The assistant answers using the metrics, dimensions and joins your team defined in the shared layer.

yes: the built-in assistant can create or edit metrics, dimensions or joins in the shared definition layer. Writing descriptions or documentation for existing definitions only is partly.

An agent can see what a field means, its description and business definition, rather than guessing from a column called REV_AMT.

The semantic layer is writable from outside the product.

Work modelled in the BI tool can be written back out so it lives in the warehouse rather than in the product.

yes: the product documents a direct path from a calculation written for one chart or analysis to a maintained shared definition, without a developer rewriting it. A path that needs the calculation re-expressed by a developer (for example in a modeling language) is partly. Reusing logic that already exists as a shared definition, macro or snippet does not count.