Active metadata gives AI agents a live read of enterprise data state, not a stored extract of what was true at collection time. Gartner predicts 60% of AI projects will be abandoned through 2026 due to context and data readiness gaps, not model quality. Where memory systems like Mem0 and Zep extract, embed, store, and retrieve, active metadata propagates changes through column-level lineage the moment they occur in Snowflake, dbt, or any connected system. This guide explains the architectural difference and why it determines whether your agents can be trusted.
| Attribute | Value |
|---|---|
| What It Is | An always-on metadata architecture that propagates changes through lineage automatically, giving AI agents current data context at inference time |
| Key Benefit | Eliminates the stored-extract staleness cycle; 3x improvement in text-to-SQL accuracy vs. bare-schema agents (Atlan-Snowflake joint research, 2026) |
| Best For | Enterprise AI agents querying governed data: reporting, SQL generation, compliance, incident resolution |
| Core Properties | Always-on, intelligent, action-oriented, open by default (bidirectional API) |
| vs. Memory Layers | Memory layers extract and store snapshots; active metadata maintains live state: no extraction, no embedding, no invalidation problem |
| Gartner Position | Active metadata is “the backbone for data agents and agentic AI” (Gartner 2025 MQ for Metadata Management Solutions) |
What makes metadata “active”?
Permalink to “What makes metadata “active”?”Active metadata is not a data catalog you visit. It is an always-flowing intelligence layer embedded in every tool in your data stack. Gartner defines it as “the continuous analysis of all available users, data management, and systems/infrastructure to determine the alignment between data as designed versus actual experience.” Four architectural properties separate active from passive, and each one matters for AI agents in ways that passive catalogs cannot address.
Active metadata is a way of managing metadata that uses open APIs to connect all the tools in your data stack and ferry metadata back and forth in a two-way stream. The contrast with a passive catalog is architectural, not cosmetic. A passive catalog is a snapshot repository: you go to it, query what was cataloged at last crawl, and act on what you find. Active metadata comes to you: when net_revenue is certified by the finance team, that certification flows to Looker, Slack, Snowflake, and every AI agent querying through an MCP server, automatically, without a human triggering a refresh.
Gartner formally replaced its Magic Quadrant for Metadata Management Solutions with a Market Guide for Active Metadata in August 2021, a categorical signal that the passive paradigm was over, not an incremental update. Gartner predicts organizations adopting active metadata practices will reach 30% by 2026. AI agents amplify the urgency: agents query data context programmatically and at scale, which means the human-refresh loop that passive catalogs depend on breaks completely under agent load. No one is checking the catalog between the agent’s calls.
The four characteristics function as a system, not a checklist. Each one addresses a specific failure mode that passive catalogs introduce.
Always-On
Permalink to “Always-On”An always-on metadata layer continuously collects metadata from all sources: query logs, transformation events, schema change events, usage statistics. There is no manual refresh, no scheduled batch crawl. Observation is continuous.
For AI agents, this means context is current when the agent queries, not current as of last Sunday’s batch job. The operational difference is meaningful when revenue_net_of_returns is redefined on a Tuesday and your agent answers a board-level question on Wednesday.
Intelligent
Permalink to “Intelligent”Machine learning processes the metadata stream continuously, classifying sensitive columns, surfacing anomalies, and generating recommendations. A new column matching patterns for credit_card_number automatically receives encryption policies and access restrictions. The ML detects the pattern, the system acts.
Agents inherit this classification intelligence without having to compute it themselves. When your agent touches transactions.card_number, it receives the classification already attached to the asset, not a raw schema field it must classify from context.
Action-Oriented
Permalink to “Action-Oriented”Active metadata does not stop at insight. It triggers automated workflows: alerts downstream consumers, halts pipelines on quality failure, updates access policies, opens remediation tickets.
A schema change detected in a development environment triggers: impacted dashboard owners alerted, remediation ticket opened, downstream pipeline optionally halted. Agents operating in this environment receive context that has already been acted on and validated, not raw event data they must interpret.
Open by Default
Permalink to “Open by Default”The architecture is bidirectional by design. Metadata flows from Snowflake to Looker to Slack to Jira and back to Snowflake, across every connected system. Atlan’s MCP server is the programmatic interface through which AI agents query this live state at inference time.
“Open” has a specific meaning here: agents from any framework, any orchestration layer, any LLM provider can connect. The context layer for enterprise AI is not a proprietary memory system; it is an API-accessible live metadata graph.
The stored-extract pipeline and its failure modes
Permalink to “The stored-extract pipeline and its failure modes”Every stored-extract memory system (Mem0, Zep, LangChain Memory) runs the same six-step pipeline: extract, embed, store, retrieve, inject, use. Each step introduces a failure mode. For conversational context, these are manageable. For enterprise data context, where the source of truth changes in external systems the memory layer does not observe, the pipeline carries a structural flaw that decay functions and explicit invalidation cannot fix.
| Pipeline Step | What Stored-Extract Systems Do | Failure Mode | Active Metadata Equivalent |
|---|---|---|---|
| Extract | Identifies salient facts from messages/documents; distills into compact memories (Mem0) | Lossy by design: captures what the extraction model deems salient at extraction time, not what inference will need | No extraction step; metadata IS the live state |
| Embed | Encodes extracted facts into vector or graph representations (Zep: temporal knowledge graphs) | Semantic meaning frozen at extraction time; evolves as vocabulary changes | No embedding; live structured graph, always current |
| Store | Persists extracted/embedded representation in vector DB or graph store | Snapshot of world at extraction time; no mechanism for update from external systems | Continuously updated by propagation events, not a snapshot |
| Retrieve | Similarity query at inference time to surface relevant memories | Retrieval lag: Zep testing showed correct answers appeared hours later after background graph processing | Direct query of live state; no background processing, no lag |
| Inject | Retrieved memories injected into context window alongside current observations | Stale memories injected with same weight as fresh observations; no built-in degradation signal | All context is current state; no injection of stale past |
| Invalidate | Explicit invalidation when contradicting signal received; OR decay functions (exponential recency weighting) | For enterprise data changes, the contradicting signal never reaches the memory layer. It happens in Snowflake, dbt, governance tools. | Propagation IS the invalidation; changes auto-propagate through lineage graph |
The invalidation row is where the architectural flaw becomes permanent. The invalidation problem is not a solvable UX problem for enterprise data context. Enterprise schema changes, deprecations, and ownership transfers happen in external systems. The memory layer will never receive the contradicting signal. It cannot invalidate what it does not observe.
Why the enterprise case is different. Mem0 and Zep were designed for conversational memory: dialogue history, user preferences, session continuity. In that domain, the invalidation problem is manageable because contradicting signals arrive through the conversation itself. Enterprise data context changes differently. net_revenue is recertified in January. A Fivetran table is deprecated. A column is renamed in a dbt model. None of these changes surface as a message to the agent. The memory layer has no signal. Some teams build custom webhook integrations to notify memory systems of specific schema changes, and this partially bridges the gap: Snowflake supports change data capture, dbt has hooks, and a sufficiently engineered team can wire deprecation events into a Mem0 update pipeline. But each integration must be built, maintained, and hardened against partial failure across every source system individually. At enterprise scale, this is architecturally fragile. Active metadata’s propagation-through-lineage is not a bolt-on integration; it is the core mechanism.
The result is what practitioners call “confidently wrong at scale”: agents producing precise answers based on definitions that changed months ago, with no mechanism to detect the error. A DEV Community analysis of agent memory architectures in 2026 noted this directly: “When information changes in one workflow but that update stays trapped in that workflow, other workflows keep operating on stale information, making decisions based on a reality that no longer exists.” Despite $30-40 billion invested in enterprise generative AI, 95% of organizations reported zero measurable ROI, with stale, ungoverned context identified as a primary driver.
The live-read architecture: how active metadata works instead
Permalink to “The live-read architecture: how active metadata works instead”Active metadata inverts the pipeline. Instead of extracting context and trying to keep it fresh, it maintains the live state as the system of record and delivers it to agents via live query at inference time. The architecture has four steps, not six, and it eliminates the steps where staleness enters the stored-extract pipeline: extraction, embedding, and invalidation are gone entirely. When we say “no extraction step,” the precise meaning is this: active metadata reads current operational state from source systems continuously; it does not perform the lossy semantic distillation into vector embeddings that stored-extract memory systems do. There is no frozen intermediate representation that can drift.
The four-step live-read pipeline is a direct sequence:
- Event: A change occurs anywhere in the data estate: schema update, deprecation, certification, ownership transfer, quality breach.
- Propagate: Active metadata’s always-on observation detects the event and traverses the lineage graph at column level, updating every downstream asset. For the high-stakes changes that matter most to agents (deprecations, certification updates, schema renames), propagation is sub-second through the event-driven lineage graph. The “15 minutes or less” figure applies to background crawl cycles for broader asset discovery — not to event-triggered propagation.
- Live in graph: The updated state is the current state of record. There is no extraction, no separate copy. The metadata IS the live state.
- Agent queries at inference time: The agent hits Atlan’s MCP server and receives the current certified, deprecated, or classified state, whatever is live right now.
The following five scenarios show how this plays out in practice.
Deprecation Propagation
Permalink to “Deprecation Propagation”A Fivetran table is flagged for migration. Active metadata sends upcoming change alerts to all downstream BI consumers; the deprecation flag propagates automatically through column-level lineage.
Any agent querying context about this table receives the deprecation flag, without a manual metadata refresh or explicit invalidation step. In a stored-extract system, the deprecation happens in Fivetran; the memory layer receives no signal; the agent continues to recommend a table that no longer exists.
CIA Classification Cascade
Permalink to “CIA Classification Cascade”A new column containing credit card numbers is added to a transactions table. Active metadata’s ML detects the pattern; CIA (confidentiality, integrity, availability) ratings propagate automatically via column-level lineage; encryption policies, access restrictions, and audit logging attach in real time.
transactions.card_number receives classification: PII and access_policy: restricted automatically. Any agent that touches this column inherits the classification. It cannot query unencrypted or unaudited data because the classification is already live in the metadata graph before the agent ever arrives.
Schema Drift Alert
Permalink to “Schema Drift Alert”A critical upstream table changes: a column renamed from revenue_gross to revenue_net_of_returns, a type changed, or a field dropped. Active metadata detects the schema drift event, identifies all downstream dashboard owners via lineage, opens remediation tickets, and optionally halts downstream pipelines.
Agents querying the schema receive the current schema. The column is revenue_net_of_returns. A stored-extract system that ran its extraction job before the rename would have given the agent revenue_gross, a field that no longer exists in the source system.
Certification Change
Permalink to “Certification Change”The finance team certifies net_revenue as the canonical revenue definition on January 15, replacing gross_revenue. The certification change propagates bidirectionally. Every connected tool, every AI agent querying via MCP, every downstream consumer now sees net_revenue as the certified definition.
An agent answering “what was revenue in Q4?” uses the post-January-15 definition, not the pre-certification snapshot that a memory extract from December would carry. Atlan’s agent context layer adds a complementary layer alongside live-read current state: decision memory, which is governed historical context rather than live-read state. It stores event trails and approval histories persistently linked to business entities, so the agent can explain not just the number but the methodology behind it — that the finance team requested the certification change on January 15. This is the WHY alongside the WHAT.
Ownership Transfer
Permalink to “Ownership Transfer”A data domain is transferred from the engineering team to the finance team following a re-org. Ownership metadata propagates through lineage. Access policies, approval workflows, and notification routing update automatically across all connected assets.
The agent routes data access requests and escalations to the correct owner, not the previous owner recorded in an extraction from three months ago. This is the ownership-transfer failure mode that never surfaces until a compliance audit reveals it.
Why staleness is an enterprise risk
Permalink to “Why staleness is an enterprise risk”Enterprise data estates change continuously. Schema migrations, table deprecations, certification updates, and ownership transfers are routine events, not exceptions. Operational AI agents require data no older than 5 minutes; analytical agents tolerate up to 1 hour. Both timescales are shorter than most stored-extract memory refresh cycles. When stale context enters agentic workflows, errors compound silently across every downstream reasoning step.
How fast enterprise data actually changes. Active metadata systems are designed to detect “schema, quality, access, or retention drift” as routine operations. Operational agents need data freshness of 5 minutes or less; analytical agents up to 1 hour. Both timescales are shorter than most extraction refresh cycles (Promethium.ai, 2026). Two latency modes operate in parallel: critical events such as deprecations, certifications, and schema changes propagate sub-second through the event-driven lineage graph; the “15 minutes or less” figure is the upper bound for background crawl cycles that cover broader asset discovery. For the high-stakes changes that matter most to agent correctness, active metadata’s propagation is effectively instantaneous — well within the 5-minute operational freshness threshold. Zep’s documented retrieval delays demonstrate the contrast: correct answers appeared hours later after background graph processing completed, in a system designed for this exact problem.
The cost of stale context at scale compounds in two directions. First, “confidently wrong at scale”: precise-sounding answers based on a reality that no longer exists, multiplied across thousands of queries, produces systematic error rather than individual mistake. Second, context poisoning: stale information enters agent context, and because agents build on prior context across reasoning steps, these errors compound silently. Prompt bloat from carrying stale context makes every reasoning step slower and more expensive; at enterprise scale, the workflow becomes economically nonviable. Gartner’s February 2025 analysis projected that 60% of AI projects will be abandoned through 2026 due to context and data readiness gaps, not model quality (Gartner newsroom, 2025-02-26).
Enterprise staleness is uniquely dangerous for two reasons that conversational-memory staleness does not share. High-stakes decisions (revenue calculations, compliance reports, regulatory filings) depend on current certified definitions. An agent using an outdated revenue_net_of_returns definition before a January 15 restatement produces a number that is precise but wrong by millions, with no indication that the definition changed. And enterprise changes are silent: schema changes, ownership transfers, and deprecations do not announce themselves to memory systems. The change happens in Snowflake, dbt, or a governance platform. No conversational signal arrives. The stored extract becomes permanently stale with no correction mechanism. Active metadata’s propagation architecture addresses both: changes in any connected system are observed and propagated to all downstream context consumers, including AI agents querying through context engineering pipelines, automatically.
The Gartner validation
Permalink to “The Gartner validation”Gartner’s August 2021 decision to replace its Magic Quadrant for Metadata Management Solutions with a Market Guide for Active Metadata was a categorical signal. The analyst firm was saying the passive catalog paradigm was over. By 2025, Gartner’s metadata MQ positioned active metadata as “the backbone for data agents and agentic AI,” with AI readiness elevated to a top evaluation criterion alongside “metadata anywhere” architecture.
The categorical signal. Gartner replaced its Metadata Management Solutions MQ with a Market Guide for Active Metadata in August 2021. This is not a refinement; it is a category replacement. Gartner’s formal definition: “the continuous analysis of all available users, data management, systems/infrastructure and data governance experience reports to determine the alignment and exception cases between data as designed versus actual experience.” The definition is operational and continuous, not archival and retrospective. This distinction is what separates active metadata from every passive cataloging approach that preceded it, and it maps directly to what a context graph becomes when it is live.
Gartner’s six predictions, read together, frame active metadata as foundational infrastructure for the AI era:
| Gartner Prediction | Implication |
|---|---|
| 70% reduction in time to delivery of new data assets for orgs adopting active metadata (document 4006759) | Speed multiplier for AI-driven data operations |
| 60% of AI projects abandoned through 2026 due to context/data readiness gaps (February 2025) | Context gap is the primary AI failure mode, not model quality |
| 30% of organizations will adopt active metadata practices by 2026 | Still early-mover territory; most enterprises have not yet adopted |
| 40% of enterprise apps will have task-specific AI agents by end of 2026, up from less than 5% in 2025 | Agent proliferation makes context infrastructure urgent, not optional |
| Active metadata is “the backbone for data agents and agentic AI” (2025 MQ for Metadata Management Solutions) | Infrastructure framing, not tool framing |
| Gartner D&A Summit 2026 framed context as “the new critical infrastructure” | Context layer = infrastructure category, not feature |
Every prediction reinforces the same structural argument: enterprise AI fails without governed, live context. The 60% abandonment projection is not a warning about model capability; it is a warning about context infrastructure readiness.
The MQ upgrade signal. In Gartner’s 2025 Magic Quadrant for Metadata Management Solutions, Atlan was upgraded from Visionary to Leader. Gartner’s 2025 MQ elevated active metadata, AI readiness, and “metadata anywhere” architecture to top evaluation criteria; static cataloging “carries far less weight.” The three-step narrative is: category death (passive MQ replaced), category replacement (active metadata market guide), category maturity (AI-native leadership designation). Atlan’s position in that progression reflects the metadata layer for AI argument in structural terms: the infrastructure that serves agents at inference time is where the evaluation criteria now land.
How Atlan connects active metadata to AI agents
Permalink to “How Atlan connects active metadata to AI agents”Atlan’s active metadata architecture connects to AI agents through three mechanisms: an MCP server that serves live metadata at inference time, Context Studio for building and testing agent context pipelines, and decision memory that stores event trails and methodology changes alongside current state, so agents explain not just what the data says but why it changed.
The challenge
Traditional governance tools were designed for human discovery workflows — a data analyst who can check a catalog, spot a deprecation flag, and adjust their query. AI agents can’t browse for context they don’t know to look for:
- Agents receive context at inference time only — if it was extracted three weeks ago, they operate on a three-week-old view
- Schema changes, deprecations, and ownership transfers happen in external systems the memory layer never observes
- Governance tools requiring human interpretation as an intermediary step don’t scale to agentic workflows
The context layer for enterprise AI must serve agents directly — not through a human intermediary.
Atlan’s approach
Atlan’s agent context layer provides two distinct capabilities that work together:
- MCP server (live-read): Exposes the live metadata graph as a queryable interface. Agents receive current certifications, deprecation flags, ownership, column-level classifications, and business glossary definitions at the moment of query — not from a stored extract
- Context Studio: Enables teams to define what metadata gets surfaced to which agent, under what conditions, and with what confidence signals
- Decision memory (the WHY): Stores event trails, approval histories, incident timelines, and metric definition changes — linked to business entities. When an agent explains revenue dropped in February, decision memory surfaces that the
revenuecalculation changed on January 15 at the finance team’s request
What the research shows:
- Snowflake internal research: adding an ontology and metadata layer → 20% improvement in agent answer accuracy, 39% reduction in tool calls
- Atlan-Snowflake joint research (145 queries): 3x improvement in text-to-SQL accuracy when agents are grounded in rich metadata vs. bare schemas
The outcome
- CME Group: 18 million data assets and 1,300+ glossary terms cataloged in year one — the scale at which live-read context becomes architecturally necessary
- Atlan customers: 50–70% faster incident resolution using active lineage vs. manual investigation
- Compliance teams: 40–50% reduction in compliance process time
The metadata layer for AI isn’t an additional tool — it’s the infrastructure that makes agent reasoning trustworthy. Combining live-read active metadata with governed decision history separates agents that produce answers from agents that produce trusted answers.
See how CME Group achieved governance at speed with Atlan’s active metadata approach.
Wrapping up
Permalink to “Wrapping up”The stored-extract pipeline is not flawed engineering. It is the right architecture for conversational memory. But for enterprise data context, it carries a structural problem the pipeline cannot escape: the invalidation mechanism requires a signal the memory layer will never receive, because enterprise data changes happen in systems the memory layer does not observe. Active metadata solves this at the architectural level, not by refreshing faster, but by eliminating the extraction and staleness cycle entirely. Propagation replaces extraction. Live state replaces snapshots. The Atlan context layer realizes this in practice: a unified, live, governed metadata graph that agents query at inference time, serving current certified state without background processing delays.
As AI agents proliferate across enterprise workflows, with Gartner projecting 40% of enterprise apps with task-specific agents by end of 2026, context infrastructure becomes as foundational as compute infrastructure. The market frames this as “give AI agents better memory.” The more precise frame: enterprise AI agents do not need better memory for data context. They need a system that is already current when they query it. Active metadata is that system. The teams that treat active metadata management as a live operational layer, not a documentation artifact, will build agents that produce trusted answers at scale.
FAQs about active metadata and AI agent memory
Permalink to “FAQs about active metadata and AI agent memory”1. What is active metadata and how is it different from regular metadata?
Permalink to “1. What is active metadata and how is it different from regular metadata?”Active metadata is a continuously-updated, action-oriented metadata layer that propagates changes through every connected system automatically. Regular metadata sits in a catalog you query manually: a snapshot of what was cataloged at last crawl. Active metadata uses bidirectional APIs and always-on observation to maintain live state, matching Gartner’s definition: continuous analysis to determine alignment between data as designed versus actual experience.
2. Why do AI agents fail when they use stored memory for enterprise data?
Permalink to “2. Why do AI agents fail when they use stored memory for enterprise data?”Stored-extract memory systems face the invalidation problem: they can only update when they receive a contradicting signal, but enterprise data changes happen in Snowflake, dbt, and governance platforms, which are external systems the memory layer does not observe. Schema changes, deprecations, and ownership transfers never surface as conversational signals. The stored extract becomes permanently stale, and agents compound these errors silently across every reasoning step.
3. What is the difference between a memory layer and a context layer for AI agents?
Permalink to “3. What is the difference between a memory layer and a context layer for AI agents?”Memory layers like Mem0 and Zep capture conversational history and session continuity, optimized for user-interaction context across dialogue turns. The context layer distinction is architectural: context layers maintain live enterprise data context (asset definitions, certifications, deprecations, policies, schema states). Both are needed for complete agent context; they serve different dimensions of what agents must know to reason correctly.
4. Can Mem0 or Zep handle enterprise data governance for AI agents?
Permalink to “4. Can Mem0 or Zep handle enterprise data governance for AI agents?”No, but that is not their design intent. Mem0 and Zep are purpose-built for conversational memory persistence: user preferences, dialogue history, session continuity. They have no mechanism to observe schema changes, certification updates, or deprecation events in external data systems. For enterprise data context, an agent context layer with propagation-based active metadata architecture is the appropriate layer.
5. How does active metadata propagate changes to AI agents automatically?
Permalink to “5. How does active metadata propagate changes to AI agents automatically?”Active metadata uses column-level lineage traversal. When a change event occurs anywhere in the data estate (schema update, certification, deprecation, ownership transfer), always-on observation detects it and pushes updated context through the lineage graph to every downstream asset. AI agents querying via the MCP server receive the propagated state at inference time, with no extraction job or cache invalidation required.
6. What happens when an AI agent uses a deprecated data asset?
Permalink to “6. What happens when an AI agent uses a deprecated data asset?”Without active metadata: the agent receives no deprecation signal and continues using the asset, producing results based on data that may be incomplete, migrated, or incorrect. With active metadata: the deprecation flag propagates automatically through column-level lineage the moment the deprecation is recorded. The agent receives the flag at query time and can route to the replacement asset instead.
7. How does Atlan provide real-time context to AI agents?
Permalink to “7. How does Atlan provide real-time context to AI agents?”Atlan’s MCP server exposes the live metadata graph as a queryable API. Agents request context at inference time and receive current certifications, deprecation flags, column-level classifications, business glossary definitions, ownership metadata, and decision trails, whatever is live in the graph at the moment of the query. No extraction job, no cache to invalidate. Joint Atlan-Snowflake research across 145 queries demonstrated a 3x improvement in text-to-SQL accuracy when agents use this live metadata layer versus bare schemas.
8. What does Gartner say about active metadata for agentic AI in 2026?
Permalink to “8. What does Gartner say about active metadata for agentic AI in 2026?”Gartner’s 2025 MQ for Metadata Management Solutions positions active metadata as “the backbone for data agents and agentic AI.” Gartner’s D&A Summit 2026 framed context as “the new critical infrastructure.” Gartner predicts 40% of enterprise apps will have task-specific AI agents by end of 2026, up from less than 5% in 2025, and that 60% of AI projects will be abandoned through 2026 due to context and data readiness gaps, not model quality.
References
Permalink to “References”- Gartner Market Guide for Active Metadata Management, document 4006759. https://www.gartner.com/en/documents/4004082
- Gartner newsroom: “Lack of AI-Ready Data Puts AI Projects at Risk,” February 2025. https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
- Snowflake: “Agent Context Layer for Trustworthy Data Agents,” 2026. https://www.snowflake.com/en/blog/agent-context-layer-trustworthy-data-agents/
- DEV Community: “Your AI Agent’s Memory Is Broken. Here Are 4 Architectures Racing to Fix It,” 2026. https://dev.to/ai_agent_digest/your-ai-agents-memory-is-broken-here-are-4-architectures-racing-to-fix-it-55j1
- The New Stack: “Is Agentic Metadata the Next Infrastructure Layer?,” January 2026. https://thenewstack.io/is-agentic-metadata-the-next-infrastructure-layer/
- Anthropic: “Effective Context Engineering for AI Agents.” https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
- Atlan-Snowflake joint research: text-to-SQL accuracy study, 145 queries, 2026. https://atlan.com/know/snowflake-intelligence-atlan-partner-talk-to-data/
Share this article
