Amazon AWS Adds an Agentic Catalog, but Human Curators Still Hold the Keys

Amazon AWS has introduced an agentic catalog experience that connects Amazon Quick with two major enterprise catalogs, despite persistent questions about AI-generated data models.

The preview supports AWS Glue Data Catalog and Databricks Unity Catalog. Data curators can describe an analytical use case in natural language, review recommended assets, and create several datasets through one guided conversation.

That workflow shifts a difficult part of business intelligence closer to the catalog. Instead of rebuilding descriptions and relationships inside every analytics tool, curators can reuse semantics already maintained upstream.

However, AWS has not removed the curator from the process. Its own documentation tells authors to review discovered tables, inferred relationships, and inherited descriptions before proceeding.

That warning defines the real story. Amazon Quick is automating the assembly of an analytical context boundary, but the quality of that boundary still depends on human judgment.

It also places Amazon AWS into a broader contest with Databricks and Microsoft. Each vendor wants governed metadata to become the foundation for conversational analytics and enterprise agents.

Amazon AWS Turns Catalog Metadata Into Quick Assets

The preview turns data discovery, dataset creation, and semantic setup into one conversational workflow.

A curator begins by connecting Amazon Quick to a supported catalog. The author then selects Explore data and describes the intended use case in ordinary language.

AWS offers a supply chain example in its agentic workflow. A curator asks for data supporting questions about carrier performance and shipping costs across distribution centers.

The agent searches available catalog metadata for relevant assets. That metadata can include table descriptions, column descriptions, data quality scores, and lineage information.

This is more specific than searching for a familiar table name. The agent is expected to match a business objective with technical assets that may use different naming conventions.

The curator reviews the recommendations before creating anything. After approval, Quick creates the selected datasets together instead of requiring a separate setup process for every table.

Those datasets use DirectQuery, which sends queries to the connected source rather than importing all data into Quick. The upstream catalog remains the declared source of truth.

The workflow then addresses relationships. Quick can inherit catalog relationships when they exist and suggest additional relationships that it infers from the available metadata.

Finally, the agent creates a multi-dataset Topic. A Topic is the semantic layer that helps Quick interpret business language and generate queries across selected datasets.

Descriptions also move downstream automatically. Table and column descriptions from the catalog become metadata on the new Quick datasets.

This inheritance matters because a column name rarely contains enough business meaning by itself. A field called revenue might represent booked revenue, recognized revenue, or a forecast.

A maintained description can preserve that distinction. Without it, an AI system must guess from names, nearby fields, and user wording.

The agentic catalog experience therefore does more than locate tables. It packages selected assets, relationships, and definitions into a context that Quick can use for analysis.

AWS describes the capability as a preview, which means its behavior and supported features remain subject to change. Preview status also makes careful evaluation essential before production dependence.

For Databricks connections, AWS explicitly says preview features are not generally recommended for critical production workloads. That qualification tempers the convenience promised by the guided workflow.

The release is still consequential because it targets a recurring bottleneck. Enterprise analytics teams often maintain valuable metadata upstream, then repeat much of that work inside reporting platforms.

Amazon AWS is betting that an agent can carry more of that context across the boundary. The curator’s job becomes selection and verification instead of repetitive asset construction.

The Real Bottleneck Is Choosing the Right Context

Finding more data is easy; defining the smallest trustworthy set for a business question remains difficult.

Large organizations can have thousands of tables spread across warehouses, lakes, operational systems, and departmental projects. A broad catalog helps organize them, but breadth creates its own problem.

An analytics agent should not receive every available table. Extra assets introduce ambiguous fields, competing definitions, irrelevant relationships, and more opportunities for incorrect query generation.

AWS calls the selected set a focused context boundary. It recommends creating only the datasets needed for the intended use case.

That advice reveals why the agentic catalog experience exists now. Conversational analytics needs narrower, better-described context than traditional catalog search usually provides.

A human analyst can inspect several similarly named tables and ask a colleague which one is authoritative. An AI agent needs those distinctions expressed through metadata, instructions, and governed relationships.

The new workflow tries to shorten the route from enterprise catalog to usable context. It searches through metadata, proposes a relevant subset, and carries approved semantics into Quick.

Consider the supply chain scenario. Delivery performance might involve shipment events, carriers, distribution centers, contracts, invoices, and calendar dimensions.

Selecting too few assets produces incomplete answers. Selecting too many increases ambiguity and can expose fields unrelated to the managers’ actual decisions.

The curator therefore remains responsible for scope. The AI proposes an analytical package, but a person must decide whether that package reflects the organization’s operating language.

This is where semantic inheritance has practical value. A well-maintained catalog already contains definitions created by data stewards and domain experts.

Reusing those definitions reduces duplicate work and discourages another isolated semantic layer. It also gives the Quick agent more business context than raw schemas provide.

Yet inheritance cannot improve inaccurate source metadata. A vague description remains vague after Quick copies it, while an outdated definition can spread confidently into a new interface.

Lineage and quality signals help the discovery process, but they do not settle every business dispute. Two governed tables can still represent different accepted versions of the same metric.

The agent must also infer user intent from a short natural-language request. A curator asking for customer value data may mean revenue, margin, retention, or a company-specific composite score.

Natural language makes setup more accessible, but it can conceal ambiguity. A form with explicit modeling choices sometimes exposes disagreements earlier than a fluent conversation does.

The important shift is not that modeling disappears. Modeling becomes a review task performed after an agent proposes a structure.

That transition resembles other AI-assisted knowledge workflows. Tools can collect and blend context, but trustworthy output still requires control over the sources that enter the working set.

For individual research, the same principle supports careful knowledge blending. The useful unit is not every available document, but the relevant evidence for a defined question.

Amazon Quick applies that principle to governed enterprise data. Its preview will succeed only if curators can understand why each asset and relationship was recommended.

Catalog Vendors Now Compete Through the Semantic Layer

Amazon Quick is entering a contest over which platform converts governed metadata into dependable AI answers.

Databricks already treats Unity Catalog as a unified governance layer for data and AI. It applies access controls, records lineage, and exposes governed assets through several interfaces.

Databricks also supports keyword and semantic discovery for registered tables and columns. Its newer discovery tools let users browse shared assets and ask questions in natural language.

The Databricks discovery experience overlaps with part of Amazon’s proposition. Both systems use catalog context to help users find relevant governed assets.

The difference lies in the destination. Amazon Quick uses catalog results to create Quick datasets and a multi-dataset Topic for analytics and conversational questions.

AWS is therefore not replacing Unity Catalog. It is turning Unity Catalog metadata into input for an Amazon-controlled analytical experience.

That distinction makes the integration both cooperative and competitive. Databricks supplies the governed source, while Amazon Quick becomes the place where business users consume the resulting context.

The preview supports Personal Access Tokens and three-legged OAuth for Databricks connections. OAuth can also preserve individual user identity when Quick queries the upstream system.

AWS Glue Data Catalog offers a more vertically integrated route. Quick can use a service role or AWS IAM Identity Center for agentic discovery and semantic inheritance.

For customers using Lake Formation, trusted identity propagation can enforce upstream permissions for each user. The source system determines which data that user may query.

Identity propagation is optional for the agentic workflow. However, it only applies to DirectQuery datasets, according to the governance documentation.

If a dataset moves to SPICE or receives transformations, that propagated identity does not apply. Teams must then use Quick’s own row-level and column-level security controls.

That limitation matters because governance is central to the product’s value. A convenient discovery process cannot compensate for a permission model that teams misunderstand.

Microsoft is pursuing a related strategy through Fabric data agents and Power BI semantic models. Those agents connect natural-language questions with governed data sources and business definitions.

Microsoft’s guidance stresses that answer quality depends on preparing semantic models for AI. Descriptions, verified answers, and instructions help the agent interpret business language correctly.

Its semantic model guidance reinforces the same lesson as AWS’s curator warning. Conversational access works best when structured business context already exists.

The competition is not simply about which model writes better SQL. It concerns ownership of the layer between raw enterprise data and the employee asking a question.

Catalog platforms want to govern that layer. Business intelligence platforms want to consume and enrich it. Agent platforms want to reason over it and initiate actions.

Amazon AWS is linking those roles inside Quick. Its catalog preview reduces friction between governance metadata and the final analytical interface.

However, Databricks and Microsoft are also moving closer to business users. They do not intend to remain passive metadata providers while another platform owns the conversational experience.

This pressure benefits customers if it encourages portable descriptions, explicit relationships, and identity-aware queries. It creates risk when each platform adds proprietary semantics that do not transfer cleanly.

The preview’s strongest architectural choice is keeping the upstream catalog as the source of truth. That approach limits another uncontrolled copy of enterprise definitions.

Its long-term value will depend on whether changes continue flowing reliably. A one-time inheritance process can still create drift if later catalog updates do not reach dependent Quick assets.

How Amazon Quick Builds a Multi-Dataset Topic

The mechanism works because Quick converts catalog objects into a constrained analytical model, rather than giving an agent unrestricted access to everything.

Quick datasets represent the selected catalog assets. The agent can create several of these assets in bulk after the curator accepts its recommendations.

The workflow then connects those datasets through inherited or inferred relationships. Those relationships tell the query engine how fields from separate tables can be joined.

The resulting Topic provides a semantic container for natural-language analysis. It holds datasets, relationships, business definitions, and instructions needed to interpret user questions.

Amazon recently expanded multi-dataset Topics through a separate public preview. A Topic can contain up to 12 datasets, according to the company’s semantic layer overview.

The query engine interprets a question, identifies useful columns, follows defined relationships, and constructs the required SQL. It then returns a table or visualization.

This model addresses a limitation in the earlier Quick Sight architecture. A dataset traditionally appeared as one flattened table, and each visual could use only one dataset.

Teams often joined source tables into a large denormalized dataset during preparation. That design simplified execution but made complex domains harder to maintain.

Multi-dataset Topics let normalized datasets remain separate. The semantic layer describes how they relate, and Quick builds joins when a question needs fields from several assets.

For example, a retail Topic might connect sales, returns, customers, products, stores, and dates. A manager could ask about return rates by customer segment and product category.

No single table necessarily contains that answer. The engine must select the correct measures, follow valid relationships, and avoid multiplying rows through an incorrect join.

Catalog inheritance can reduce the setup burden for this model. When Unity Catalog already records relationships and descriptions, Quick can reuse them instead of requesting manual re-entry.

AWS Glue provides table and column descriptions, while Unity Catalog can also supply relationships. The preview’s exact capabilities differ by source and authentication method.

Descriptions are only one part of semantic accuracy. Quick’s broader enrichment model also includes synonyms, semantic types, calculated fields, exclusions, and custom instructions.

A synonym can map “headcount” to a field with a technical name. A semantic type can tell the system that a column contains currency, dates, cities, or states.

Custom instructions can encode fiscal calendars or internal definitions. These additions remain important when upstream catalog metadata lacks the detail required for a specific audience.

The agentic catalog experience accelerates initial construction, but it does not eliminate downstream refinement. Curators still need to test realistic questions and inspect generated results.

DirectQuery also creates an operational tradeoff. It keeps data and permissions closer to the source, but query speed depends on the upstream platform and generated SQL.

SPICE, Amazon’s in-memory analytics engine, can improve interactive performance for imported data. However, switching away from DirectQuery changes the identity-propagation story.

Teams must therefore balance freshness, performance, governance, and transformation needs. The preview does not collapse those choices into one universally correct configuration.

The mechanism is valuable because it automates repeatable work. It can search metadata, create representations, inherit definitions, and assemble a candidate Topic.

The harder decisions remain contextual. Curators must choose which assets belong together, which relationships are safe, and which business definitions need further clarification.

Fluent Recommendations Still Need Skeptical Review

The main risk is not an obviously broken workflow; it is a plausible recommendation that quietly encodes the wrong business meaning.

AWS explicitly instructs authors to review every recommendation. That includes discovered tables, inferred relationships, and descriptions inherited from the source catalog.

The warning is especially important for inferred relationships. A shared column name does not guarantee that two fields use the same grain, domain, or update schedule.

A customer identifier might represent an account in one table and a billing entity in another. Joining them can produce believable totals that are still wrong.

Cardinality creates another risk. An agent may identify a technically valid join but fail to anticipate duplication caused by many-to-many relationships.

These failures are difficult because the resulting dashboard can look polished. Natural-language explanations may also make an uncertain answer sound more authoritative than it is.

Curators need test questions with known results. They should compare Quick’s outputs with trusted dashboards, approved queries, and domain-owner expectations.

The review should cover negative cases as well. A reliable Topic must know when the selected context cannot answer a question safely.

Metadata quality remains another pressure point. Catalogs often contain incomplete descriptions, stale ownership records, and inconsistent naming across business units.

Semantic inheritance preserves existing work, but it also preserves existing defects. Automation increases the speed at which both good and bad context travels.

Data quality scores can help rank assets, yet a score rarely captures every semantic concern. Freshness, completeness, and validity do not prove that a table answers the intended question.

Lineage can show where data originated and how it moved. It does not necessarily explain why finance and sales use different definitions for the same label.

Preview maturity adds uncertainty. AWS has not published independent accuracy benchmarks for the complete catalog-to-Topic workflow in the reviewed materials.

There is also no public evidence showing how much curator time the feature saves across different catalog sizes. Claims about faster delivery should therefore remain company claims.

Cross-platform behavior deserves similar caution. Unity Catalog metadata can be rich, but organizations configure and maintain it differently.

A well-governed Databricks environment offers more useful input than a catalog containing mostly technical schemas. The integration cannot create missing institutional knowledge from nothing.

Permissions require deliberate testing. Teams should verify results under several user identities rather than assuming catalog connectivity guarantees correct enforcement.

DirectQuery is necessary for identity propagation, but identity propagation itself is optional. Administrators must understand which control plane protects each dataset.

The agent’s suggested context should also remain inspectable after creation. Curators need a clear record of selected assets, inherited definitions, inferred relationships, and manual changes.

Without that visibility, troubleshooting becomes harder. A wrong answer might originate in source data, catalog metadata, relationship inference, Topic configuration, or generated SQL.

This does not make the preview unworkable. It defines the evaluation standard that enterprise buyers should apply.

A useful pilot should focus on one bounded business domain with established reference answers. The team can then measure setup effort, correction rates, and answer consistency.

The strongest outcome would not be zero human involvement. It would be faster assembly while preserving clear ownership and an auditable review path.

What to Watch as the Preview Expands

Three signals will show whether Amazon Quick is becoming a trusted semantic consumer or merely another place to repair metadata.

The first signal is the quality of production feedback from AWS Glue and Databricks customers. Teams should watch how often curators accept recommendations without substantial correction.

High acceptance rates across well-governed catalogs would support AWS’s mechanism. Frequent table substitutions or relationship repairs would weaken the automation claim.

Raw acceptance alone is insufficient. Customers also need stable answers when several users phrase the same business question differently.

The second signal is how AWS handles catalog changes after initial creation. Semantic inheritance has lasting value only when teams can manage updates without silent drift.

AWS should clarify whether changed descriptions, relationships, quality signals, and lineage flow into existing Quick assets. It should also explain how conflicts are surfaced.

Reliable synchronization would strengthen the upstream source-of-truth model. Manual re-imports would return much of the maintenance burden that the workflow promises to reduce.

The third signal is competitive response from Databricks and Microsoft. Both already combine governed data, semantic context, and natural-language interaction.

Databricks can deepen its own path from Unity Catalog discovery to business analysis. Microsoft can tighten connections among Fabric, Power BI models, and data agents.

If those platforms improve cross-system portability, customers gain more freedom over the consumption layer. If semantics remain proprietary, switching costs will rise.

General availability will provide another practical checkpoint within these signals. Buyers should look for broader catalog support, documented limits, administrative controls, and measurable reliability.

Amazon AWS has identified the right enterprise problem. AI analytics cannot depend on schema names and unrestricted catalog access.

The preview also uses a sensible boundary. The agent proposes assets and relationships, while the curator approves what becomes part of the analytical context.

Now AWS must show that this division of labor survives real catalog complexity. A fluent setup conversation is helpful, but trustworthy analytics requires repeatable verification.

Teams evaluating Amazon Quick should choose one domain with mature metadata and known answers. They should record every correction made during dataset and Topic creation.

They should then test permissions, relationship behavior, query consistency, and catalog updates. That evidence will reveal whether the workflow reduces modeling work or simply moves it.

The broader question is relevant beyond business intelligence. Every enterprise agent needs a controlled bridge between a user’s language and the organization’s actual information.

Amazon AWS now offers one version of that bridge. The next few months should show whether curators can cross it faster without surrendering the judgment that keeps enterprise data credible.