Rill gives AI and dashboard users the same governed metrics

Rill combines interactive analytics with a semantic layer that defines the measures and dimensions used by dashboards and AI assistants. Projects are described in SQL and YAML, and users can investigate metrics through the interface or ask questions conversationally. The important proposition is that an assistant should use an agreed business definition instead of inventing a new calculation from raw tables each time somebody asks why a number changed.

01 / ProductA code-defined analytics layer with several ways to ask

The Rill product site presents dashboards, a semantic layer and AI-assisted exploration as one platform. A metrics view defines what can be measured and how the data can be grouped. This is useful when terms such as revenue, impressions or active customers have precise meanings that must survive a change from a chart to a natural-language question.

The AI feature guide distinguishes project development from analysis. Coding agents can help create the SQL and YAML files, AI Chat supports conversational exploration in Rill Cloud, and the MCP server makes projects available to compatible external assistants. These are different workflows: generating a dashboard’s definition is not the same activity as querying a deployed dashboard.

The agent-skills documentation describes support for code-oriented project creation. The resulting files remain artifacts a developer can review. That is a practical advantage when business logic should be inspected in version control, although generated code still needs validation against the intended metric rather than merely successful execution.

02 / AudienceFor repeated investigation, not just a monthly slide

Rill is most relevant when teams repeatedly explore a changing operational dataset: advertising delivery, product activity, infrastructure usage or another time-based business process. An analyst may begin with an overall change, narrow to a segment and compare a different period. An AI assistant can help formulate those steps, but the data team still owns the definitions underneath them.

It is less useful when the main task is collecting and cleaning source systems that do not yet produce reliable data. A semantic layer can describe a conversion rate but cannot recover a missing conversion event. Likewise, a team that only needs a static monthly export may not benefit enough from interactive exploration to justify a new operating surface.

ThoughtSpot is a relevant comparison for conversational business analytics and how users investigate an answer. dbt is useful when the central requirement is managing transformations and analytics definitions as code. Rill should be evaluated on the combination of code-defined metrics, exploratory dashboards and AI access that the team actually needs, rather than on a generic claim to replace all BI tools.

03 / WorkflowA proposed investigation into an advertising delivery change

Suppose an advertising operations team sees a decline in completed video impressions. The proposed pilot uses a bounded dataset with campaign, publisher, device, geography and event time. Its goal is to identify the segments contributing to the decline and produce a reproducible explanation. It does not claim that Rill was tested or that an AI explanation establishes the causal reason for a campaign outcome.

Define the measures before asking the question. Completed impressions, completion rate and revenue are different metrics; their denominators and exclusions need to be explicit. Establish one timezone and a comparable reporting window. If today is incomplete, comparing it with a full previous day is a misleading default no matter how polished the chart looks.

Write descriptions for measures and dimensions in the metrics view. The AI configuration guide says this documentation is included in AI context, alongside project and view-specific instructions. Use those instructions to explain known data delays and preferred comparison logic. Keep the mathematical definition in the metric itself so a prose instruction does not become the sole source of truth.

Ask the assistant to quantify the total change, identify contributing segments and provide the filtered views supporting its explanation. Then reproduce those cuts in the dashboard. A publisher’s share of the overall decline is an arithmetic contribution, not proof that the publisher caused the change. The operations team should inspect delivery changes and source completeness before making a commercial decision.

Apply access policy before sharing the project with partners. The security guide describes general access, row filters and visibility rules for measures and dimensions, and says metrics-view policies also govern MCP and custom API requests. Test with two partner identities and an internal analyst. Each route should expose the intended subset rather than relying on the assistant to remember a restriction.

Finish by testing an ambiguous question, an incomplete date range and a metric that the project does not define. A useful response should request clarification or identify the gap instead of silently inventing an interpretation. Preserve the question, filters, metric definition and supporting result so another analyst can reproduce the answer later.

04 / PricingSeats, hosted compute and embedded access are separate lines

The pricing page lists Starter at US$20 per seat per month and Growth at US$30. Hosted project compute is additional. Starter supports up to 20 seats and three projects; Growth raises those limits and permits paid embedded analytics. Enterprise uses a negotiated annual platform license. The current table should be checked for the intended deployment rather than treating the headline seat amount as the total.

For a small illustrative Starter setup, two seats cost US$40 monthly. One project running two compute units for the page’s 720-hour example at US$0.15 per unit-hour adds US$216, giving US$256 before other applicable charges and credits. This is arithmetic using the public calculator’s assumptions, not an observed bill. Database resources, API overages, AI usage and taxes can change it.

The page’s introductory US$250 usage credit is not a recurring free allowance. Also note that Starter permits embedded previews, while Growth lists US$500 monthly for up to 100 active embedded seats. A partner-facing pilot must be scoped to the appropriate commercial route before treating an internal dashboard subscription as production embedding permission.

05 / DistinctionsOne metric definition can travel across interfaces

The meaningful distinction is continuity between code, a visual investigation and an AI question. When each surface uses the same metrics view, a team has a concrete place to correct a definition. Without that agreement, a conversational answer can differ from a dashboard because each surface made a different choice about filters, joins or denominators.

Code-defined projects also create a reviewable boundary for AI-assisted development. An agent can help produce a connector or dashboard file, but the team can inspect the proposed change and compare known answers before deployment. This is more useful than treating the generation of a plausible chart as evidence that its underlying metric is correct.

Rill’s AI context has a specific visibility boundary. The configuration guide says skill contents are available to all AI users of the project, potentially including anonymous visitors on a public project. Keep secrets and private business records out of those instructions. Row-level data policy and the visibility of explanatory text are separate matters.

06 / QuestionsVerify the explanation and the resource model

The first unresolved question is whether the assistant’s explanation remains grounded when a metric is ambiguous. Create a small set of questions with known answers and deliberate traps, such as a change caused by a partial reporting period. Judge whether the assistant exposes the filters and evidence needed to catch the trap, rather than evaluating only the fluency of its response.

The second question is capacity. Hosted compute can be a larger cost than seats, especially for always-running projects or heavy simultaneous exploration. Measure the workload that includes AI investigation, since one user question may produce several queries. A low seat count does not imply a low compute requirement.

Finally, test access through each actual consumption route. A dashboard, an embedded application and an MCP client may use different identities or tokens. Confirm that the intended attributes are propagated and that a revoked user loses access. The presence of a security feature is the starting point for a test, not evidence that a specific customer configuration is correct.

07 / DecisionStart with a metric people already argue about

Pick one operational metric whose interpretation is currently inconsistent. Define it, expose it in a dashboard and ask the same questions through AI Chat. If users can reproduce the answers and understand their access boundaries, expand to a second dataset. Rill’s strongest case is a more useful investigation process with consistent definitions, rather than merely adding a chat box to existing charts.