DataHub connects enterprise metadata with the context AI agents need

DataHub connects information about datasets, pipelines, dashboards and business definitions into a shared metadata platform. Its commercial Cloud offer extends the open-source Core foundation with managed operations, governance and AI features. The AI use case is to give an assistant enough context to select the right data and explain its dependencies before writing a query. This blueprint uses a proposed discovery workflow and current public documentation; it does not claim a hands-on deployment or independently measured productivity gain.

01 / ProductOne company, an open-source foundation and a managed platform

The current company site presents DataHub as an AI and data context platform. Its privacy policy identifies the legal company as Acryl Data, Inc. doing business as DataHub. The current corporate domain is datahub.com. Acryl naming in older documentation therefore refers to this coverage identity, not an additional company.

DataHub Core and Cloud share an underlying metadata foundation. Core is the self-hosted open-source route, with the operator responsible for installation, configuration and upgrades. Cloud is the managed commercial offering, adding operational support and advanced features. Buyers should compare the actual editions rather than assume that every advertised Cloud capability is present in a Core installation.

The lineage product connects dependencies across data sources, pipelines, transformations and BI assets. Column-level relationships can help an engineer determine what changes when a field is renamed or a calculation is revised. The completeness of that map depends on the actual integrations and processing paths represented.

AI data management includes hosted MCP access, conversational discovery through Ask DataHub and assistance with documentation and quality checks. The broader context platform brings technical metadata, business knowledge and documentation together for agent consumers. These features help expose maintained context; they do not turn metadata into an infallible description of reality.

02 / AudienceA fit for data estates with discovery and ownership friction

DataHub is useful when engineers and analysts struggle to identify the right dataset, understand a transformation or find the person responsible for a field. The same friction affects AI assistants. An agent that can search a warehouse but cannot distinguish a certified production table from an abandoned experiment can produce a plausible answer from the wrong source.

The platform is especially relevant where data assets span several tools and teams. A shared metadata graph can connect the warehouse, transformation code and dashboard, while a business glossary records the language users actually use. A small application with one maintained schema may not need a separate metadata service before its first AI feature.

Alation and Collibra are contextual comparisons for discovery, stewardship and governance. DataHub adds a particularly clear open-source versus managed deployment choice. Evaluate the work your team will own in each case: ingestion reliability, metadata review, upgrades, access controls and support when a connector stops updating.

03 / WorkflowA proposed workflow for selecting a trustworthy churn dataset

Begin with a concrete request: an analyst wants to investigate subscription churn for a particular product and reporting period. Agree what counts as churn, whether paused accounts are included and whether the period follows billing or calendar dates. Record those terms with a named owner so an assistant can retrieve the definition rather than infer it from table names.

Ingest metadata for the relevant warehouse tables, transformation models and dashboards. Inspect the lineage chain for the selected metric, including any custom steps that automatic extraction misses. The pilot should deliberately include an old similarly named table. This tests whether the catalog’s context helps the assistant prefer the maintained asset over a superficially matching alternative.

Add ownership, descriptions and certification information where appropriate. DataHub’s governance platform describes workflows for standards, ownership and access requests. Use those records to explain why the dataset is suitable for the task. A certification label should have an accountable reviewer and a scope, rather than becoming a permanent blanket approval.

Connect a compatible assistant to the supported metadata interface. Ask it first to identify candidate assets, explain their definitions and list unresolved questions. Do not make the first task an unrestricted data query. The useful initial result is an inspectable selection decision: this dataset matches the product and period, this owner maintains it, and these upstream dependencies matter.

Have the assistant propose SQL only after the asset selection is accepted. Compare the query with the agreed business rule, especially date boundaries and joins. A correct column description does not prevent a duplicated join or an omitted exclusion. Run the proposed query through the team’s normal review and data-access controls.

Test a schema change in a non-production example. Use lineage to identify affected consumers and verify that the assistant sees the updated metadata. A stale ingestion run can leave a catalog looking complete while its relationship graph is out of date. Record ingestion freshness alongside the evaluation result so a failure can be traced to the right layer.

Include a permissions test in the proposed pilot. A user who cannot access a sensitive dataset should not receive sample values or descriptive context that reveals protected information through the assistant. Also verify that discovering a dataset does not itself grant permission to query it. Metadata visibility and source-data authorization are separate decisions.

Finish by defining how users correct an answer. A reported definition error should reach the responsible owner and become a reviewed update. If the assistant generated documentation, preserve the difference between that draft and an approved business description. The process is successful when the next user benefits from the correction without losing its provenance.

04 / PricingCompare operating responsibility as well as license cost

The Core versus Cloud page describes Core as free to run and Cloud as the managed enterprise offer. The reviewed commercial pages invite a demonstration rather than publish a universal Cloud currency tariff. That means the defensible comparison is deployment and entitlement scope, followed by a quotation for the intended environment.

For Core, budget engineering time and infrastructure even when the software is free to run. For Cloud, ask which ingestion patterns, support commitments and AI capabilities are included. A quotation should identify the scope of the managed service and any work still required from the customer.

The business case should measure the discovery problem the team actually has. Track how often users choose the wrong dataset, how long it takes to identify an owner and whether schema-change reviews become clearer. Avoid treating sponsored survey figures or broad platform claims as a forecast of your own savings.

05 / DistinctionsThe metadata graph creates useful context before a query runs

DataHub’s strongest distinction is the connection between discovery and dependency information. A keyword search can find a table; lineage and business metadata can explain why it exists and what would change if it moved. For an assistant, that context can improve the selection process before the first SQL statement is generated.

The open-source foundation also gives engineering teams a meaningful operating choice. It can be useful for organizations that need direct control and have the capacity to run the service. The managed alternative changes that responsibility and adds capabilities, but migration should be evaluated with the actual metadata model and integrations.

The AI features address maintenance as well as consumption. Suggested descriptions or quality thresholds can reduce blank fields and repetitive setup. Their value depends on review and calibration: a generated explanation based on schema and usage may overlook a business exception known only to the owner.

06 / QuestionsQuestions that keep metadata from becoming false confidence

How current and complete is the ingestion? Inspect the last successful run and compare selected assets with the live source. Unsupported transformations, dynamic SQL and manually maintained files may require extra work. An absent dependency in the graph should not be interpreted as proof that no dependency exists.

Which edition supplies each required capability? Product comparisons can simplify complex feature differences. Reopen the documentation for the installed version and confirm the exact authentication, governance and assistant behavior. Avoid basing a deployment decision on a broad statement that one edition has “security” and another does not.

What evidence supports a generated description or recommended quality check? Ask reviewers to inspect the underlying schema, sample access settings and transformation logic. The platform can propose useful metadata, but it cannot decide every business meaning from technical signals alone.

07 / DecisionStart with one discovery decision that can be audited

DataHub merits evaluation when dependable data discovery and metadata maintenance are becoming prerequisites for AI adoption. A strong pilot connects one business question to its approved definition, source, owner and lineage. The deployment choice should follow the organization’s ability to maintain that system over time.