Metabase vs. Lightdash: who reapproves AI answers when metric definitions change?
Teams maintaining questions in the BI interface can first strengthen Metabase metrics and verification. Teams with code review that want agents to reuse reviewed answers can evaluate Lightdash. The deciding differences are definition changes, what gets verified, and responsibility for maintaining trusted answers.
When a colleague asks “What are net sales this month?”, an AI-generated chart does not mean finance has approved how refunds are deducted. Choosing Metabase or Lightdash starts with where your team wants definitions and reviews to live. If analysts mainly maintain saved questions in the BI interface, Metabase fits that existing practice. If you already have code review and semantic-model owners, Lightdash is better suited to connecting definition changes with trusted answers. Both still need a business owner to approve the calculation; software cannot sign off on their behalf.
This analysis uses official documentation. No enterprise data was connected and no hands-on tests were performed. The comparison concerns ordinary sales and order metrics, not high-impact decisions about employees.
First decide who changes “net sales”
A Metabase metric is a reusable, standardized aggregation with a data source, formula, and dimensions. It can be built on a table, saved question, model, or another metric. A model and a metric are different: the former can organize data for querying, while the latter defines aggregation. A metric is usable only on the data source on which it was defined. Editing its definition immediately changes all questions that use it. That reduces duplicated formulas but also requires the maintainer to check the impact beforehand. Official documentation
For example, changing “order value” to “paid amount minus refunds” requires more than renaming a chart. A business owner should decide which date a refund belongs to, an analyst should change the formula, and another colleague should check refunds crossing month boundaries before weekly-report readers are notified. Otherwise, rerunning the same historical chart changes its results, and colleagues may mistake a definition change for a business change. This is a recommended division of review responsibilities, not a promise of automatic approval built into Metabase.
Lightdash can define metrics in a dbt project's meta configuration, but it also supports standalone Lightdash YAML. Teams with existing warehouse tables do not need to install dbt just for it. Native YAML describes tables, metrics, and dimensions; it does not build or transform warehouse tables, and switching the connection format does not convert files. Its MetricFlow metric translation supports only a subset, so “dbt integration” must not be presented as complete inheritance of the dbt Semantic Layer. Official documentation 1 Official documentation 2
Teams with Git review can therefore review a refund-rule change together with its description, then deploy the definition. Operations teams without those maintainers take on additional coordination around code, deployment, and warehouse tables. A definition in code is not business approval: decide who approves the calculation and who deploys it.
What the badge approves matters more than its color
Metabase content verification is performed by admins and is available on Pro and Enterprise, both cloud and self-hosted. Changing a question's, model's, or metric's query removes verification. Dashboards are different: their verification survives edits and does not extend to their questions. A badge on a weekly-report dashboard therefore does not prove that each chart has passed the latest review. Official documentation
Lightdash already offers verified charts, dashboards, and AI verified answers. AI retrieval prioritizes verified content, but verifying a dashboard likewise does not verify its charts. Editing content and re-verifying it are separate permissions. On saving changes, check whether the badge is removed or explicitly re-applied by someone authorized. The overview and detailed save instructions on the same official page do not fully agree on how the original verifier retains the badge. Before adoption, test the save prompt on the target version rather than assuming “my own edits always keep verification automatically.” Official documentation
This makes acceptance testing concrete for both products: take a net-sales chart used in a recurring meeting, change the refund filter, and check who can edit it, whether re-verification is needed, and which status report readers actually see. A badge records a person's judgment at a point in time; it cannot replace checking business rules.
Turn one satisfactory response into a reusable answer
Metabot first looks for an existing question that can answer the request, then attempts to create a chart if none fits. Generated output still needs review. In the native SQL editor it generates SQL without running it, leaving an opportunity to inspect it. That does not mean Metabot never executes queries: admins can allow internal Metabot to run queries and read results, and result rows may be sent to the AI provider. With result reading disabled, it can still build queries and charts. Official documentation 1 Official documentation 2
A Lightdash verified answer is a checked reference example for an agent, not model retraining. An Admin or Developer can verify it after reviewing the query. To also expose it as verified content and through MCP's list_verified_content, save it as a chart or dashboard first, then verify it. An answer that exists only in a conversation remains an agent example and does not appear in MCP. Removing answer verification also removes the linked verification; verifying a chart directly does not retroactively verify its originating conversation answer. Official documentation
For example, if a sales colleague asks for “net sales in East China this month,” check the region field, month boundaries, and refund date, then save the query with those conditions. Next month, check date parameters and data freshness rather than expecting AI to repeat its previous explanation verbatim. Lightdash makes this path from answer to shared content more explicit. Metabase teams should focus on maintaining saved questions and shared metrics, avoiding separate near-duplicate versions for every person.
When colleagues get different answers, distinguish three possible causes: the underlying data changed, the formula changed, or the latest question used different filters. Keep the original question, saved query, metric-version notes, and review date somewhere the team can consult to reproduce the difference. A chat screenshot alone cannot prove that the next query will use the same conditions.
Trusted answers also need a named maintainer. When the refund policy changes, that person should locate affected charts, arrange reviews, and explain when old results cease to apply, with someone covering during leave. Neither product can infer accounting or business rules the company has not decided from a natural-language question. Changing the model does not eliminate that work.
Budget for ongoing human review
Metabot is currently available on every Metabase plan. Self-hosting requires configuring your own AI provider. Cloud users can also choose the Metabase AI service, which adds token-based charges beyond the regular cloud subscription. Restricting it to verified or curated content requires Pro or Enterprise. AI availability and the ability to constrain AI to trusted content are separate purchasing conditions. Official documentation
Lightdash currently lists Cloud Pro at US$3,000 per month, without per-seat pricing, including AI agents and MCP. Open-source self-hosting is positioned as core BI; do not assume the free edition includes identical AI entitlements. Enterprise supports cloud or on-premises deployment with a separate quote. Official documentation
Small teams that already organize their questions in Metabase should usually improve formula descriptions, verification, and change notifications before deciding to migrate. Teams with code-managed metrics and a need for internal and external agents to reuse reviewed answers can prioritize evaluating Lightdash, while staffing model maintenance and review. If an old answer is misused, first remove its verification or take it out of the trusted scope, correct definitions and conditions, rerun representative queries, and only then restore it. Simply giving AI a thumbs-down is not enough. Ultimately, buy a maintenance process someone owns; the chat interface is only one part of it.
Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.
Original source:www.metabase.com