The context layer for AI.
Atlan is the context layer for AI — the governed infrastructure that delivers enterprise knowledge to every model, every agent, and every team from a single source of truth.
Context doesn't come from a prompt. It comes from a pipeline.
What is a context layer?
The context layer sits between your data stack and your AI agents, translating raw metadata into the business meaning, governance, and lineage that agents need to reason correctly. Building one is straightforward; keeping it whole is the work. Drop any single property and the layer collapses back into a partial approach: a semantic layer alone, a knowledge graph, or a static catalog.
See what makes a context layer work for AI →
AI agents reason over your enterprise without context.
17% → 42% scrapping AI before production, in one year.
Atlan · State of Enterprise Data & AI 2025A production context layer requires four things.
A context layer earns the name only when four capabilities hold together at runtime — not on paper, not in a roadmap deck.
A production context layer requires four things: unified coverage across every data system, canonical business semantics, governance applied by default, and continuous synchronization with the live data estate.
Without a context layer, AI agents hallucinate, contradict each other, and act on stale or unauthorized data; with Atlan, every agent reasons from the same governed truth.
- 1Unified coverage
Every data system, BI tool, and knowledge surface in scope. Nothing left dark, nothing partial. The Enterprise Data Graph powers Atlan's Data Marketplace, the catalog and governance surface used by data teams, analysts, and stewards across the business.
- 2Canonical business semantics
One agreed model of what your terms mean, applied consistently wherever an agent or human queries them.
- 3Governance by default
Policy, authority, and access enforced at every query, every answer, every agent, not bolted on after the fact.
- 4Continuous synchronization
Stays current with the live data estate as schemas, owners, and definitions change. No stale snapshots, no drift.
Context to every agent.
Gold Layer context serves every agent through MCP, SQL, and APIs. Memory, evals, and traces feed back into the pipeline so context gets sharper with every interaction.
UPCOMING SESSION · MAY 27 · LIVE
WTF IS THE
CONTEXT LAYER?
Industry Perspective: Why is context so important?
May 27 · 12 PM ET · Live
Every vendor, analyst, and framework has landed on "context" at the same time. But why now and what does it actually mean? Sanjeev Mohan gives his unfiltered take on what's real, what's hype, and what the industry is getting wrong.
Sanjeev Mohan, SanjMo
Prukalpa Sankar, Atlan
Austin Kronz, Atlan
Already in production at AI-native enterprises.
"AI initiatives require more context than ever. Atlan's metadata lakehouse is configurable, intuitive, and able to scale to hundreds of millions of assets."
Atlan is the context layer for AI.
Let's build yours.
How does a context layer compare to a data catalog or semantic layer?
A context layer is the system that stores and delivers definitions, lineage, and governance to AI agents at inference. Catalogs, semantic layers, and knowledge graphs each handled one slice of that job. A context layer unifies all three.
vs. data catalog
A data catalog lists data assets like tables, owners, and descriptions for human discovery. It answers "what's in our stack?"
A context layer answers a different question: how should AI reason over those assets at inference? It adds decision history, runtime policy enforcement, and column-level lineage that agents query during a response. The catalog tells you a column exists. The context layer tells the agent what it means and whether the requester is authorized.
Atlan ships both: the catalog as the front door, the context layer as the engine.
vs. semantic layer
A semantic layer maps metrics to dashboards for BI tools and human analysts. Tableau, Looker, Mode all pull a "revenue" definition from one place so a chart matches the deck.
A context layer captures the full world model an agent needs: not just metric definitions, but the relationships between entities, the governance policies that apply for the user asking, and the authority record of who validated each piece. It serves that model through MCP from Atlan's Context Lakehouse to every agent in your stack, not just the BI layer.
Atlan's context layer treats the semantic layer as one input it ingests. Only the context layer can deliver definitions with the operational metadata an agent needs to cite its answer.
vs. knowledge graph
A knowledge graph captures entities and relationships: customer to order, order to product, product to inventory. The structure of what connects to what.
A context layer adds operational metadata to that structure: lineage, freshness, governed policies, decision history. The knowledge graph shows you the wiring; the context layer shows whether the wire is live, who certified it, and where it's authorized to flow.
Atlan's context layer ships the Enterprise Data Graph and the operational metadata together, served through MCP at inference. The answer arrives with the metadata baked in.
How does a context layer work for AI agents?
When an AI agent asks your data a real question, the context layer assembles definitions, lineage, policy, and provenance from the Context Lakehouse in one round-trip, then returns a grounded answer with citations.
The 7 steps in a context-layer trace
- Agent receives a question. A user asks the agent something concrete enough to require enterprise data: a metric, a timeframe, a population. The agent recognizes terms it doesn't itself define and needs governed context before it answers.
- Claude calls Atlan via Model Context Protocol. Instead of guessing, the agent reaches out to the context layer through MCP, the open standard for delivering structured context to AI. The request specifies which terms need resolution and which user is asking.
- Definitions resolved from the ontology. Atlan returns the canonical definitions from the Context Lakehouse: "net revenue" maps to a specific column in the finance schema, "enterprise customer" maps to a tier filter on CRM accounts. These are the definitions the data team certified, not the agent's best guess.
- Lineage validated. Before the agent uses a definition, the context layer traces source columns through every transformation. If the source is stale or the lineage is broken, the agent learns about it now, not after the wrong answer ships.
- Access policy checked. The user's role and entitlements are evaluated against every column in scope. PII fields get masked, region filters get applied, and the policy decision is recorded for audit.
- Query executed against your data. With the right definitions, validated lineage, and policy in hand, the query runs against the actual data warehouse. The context layer doesn't store your data; it directs how the agent reasons over it.
- Grounded answer returned with citations. The agent returns the answer and the sources behind it: which columns it used, which definitions it relied on, which policies were enforced. That's the difference between a plausible answer and a correct one.
In a context layer, the trace is: question → MCP call → definitions resolved → lineage validated → policy checked → query executed → grounded answer returned with citations. Skip any step and the agent is back to plausible answers.
Atlan's context layer ships as four products.
An Enterprise Data Graph for unified coverage. Context Agents that draft canonical context. A Context Engineering Studio to test and version it. A Context Lakehouse that stores it on open formats. Each one builds on the catalog you already have.
Connect all your business systems and pull context across your data estate into one living graph.
AI teammates that document tacit knowledge and make your data AI-ready.
Version, test, and deploy the context your AI agents depend on, as code, in your CI.
The storage foundation purpose-built for context: a unified graph + file architecture on open formats, with vector search built in.
Common questions answered
No. A data catalog lists data assets like table names, owners, and descriptions, and is built for human discovery. A context layer is an active runtime that governs how AI reasons over those same assets at inference time, adding decision history, runtime policy enforcement, and lineage that agents query directly. The two are complementary, not interchangeable: the catalog tells you a column exists, the context layer tells the agent what that column means and whether the requester is authorized to use it. Atlan ships both: the catalog as the front door, the context layer as the engine.
Larger context windows let an agent read more tokens at once, but they don't tell the agent which definitions are certified, which lineage is current, or which policies apply to the user asking. A million-token context window full of raw documents is still a million tokens of unverified text. An agent that can't tell verified from unverified will confidently cite both. A context layer is the governed source those tokens are drawn from: it provides the certified definitions, the current lineage, and the user-scoped policies that turn raw text into a trustworthy answer. Atlan's context layer is what makes a long-window model into a production-ready agent.
Model Context Protocol (MCP) is the open standard for delivering structured context to AI agents at inference. It's the wire that connects an agent to the source of truth at the moment the agent needs it. The context layer matters because it gives MCP something governed and trustworthy to serve: verified definitions, lineage, policies, and decision history, all assembled per request and per user. Without a context layer underneath, MCP just hands agents whatever loose context happens to be available, and each agent ends up assembling its own ad-hoc version from raw data. Atlan is the source MCP serves: the Context Lakehouse holds the governed definitions, lineage, and policies that make MCP useful in production.
Atlan's context layer rolls out in four stages: Unify (connect the data estate to existing warehouses, BI tools, and business systems), Bootstrap (Context Agents draft canonical definitions, metrics, and ontology automatically), Collaborate (humans resolve and certify the context), and Activate (certified context flows from the Context Lakehouse to every agent through MCP). Most teams move through Unify and Bootstrap in days because Atlan inherits the existing estate and AI drafts the first layer of definitions. Collaborate and Activate are paced by how quickly the data team can certify and how many agents are in scope. Atlan AI Labs has measured a 5x accuracy improvement in agents grounded in this context, so the right success metric isn't time-to-launch. It's the accuracy delta on the first agent that exposes the most risk.
A semantic layer standardizes metrics for human-built dashboards: a "revenue" definition lives in one place so every chart matches. That's enough to keep BI consistent. Context grounding is the broader system that gets AI to the same standard. It pulls verified definitions, lineage, policies, and decision history into every AI answer at inference time, so an agent can cite its sources and explain why an answer is correct. The semantic layer is one input into context grounding, not a replacement for it. Atlan's context layer ingests the semantic layer into the Context Lakehouse, where governance, lineage, and authority get bound to every metric an agent needs in production.
Stay in the loop on enterprise AI
Context & Chaos is Atlan's weekly Enterprise AI newsletter. Practitioner takes on context architecture, governance, and what's working in production.
Follow our Enterprise AI newsletter