Skip to main content

Dual Plane of Context and Control for Trusted AI

Emily Winks, Data Governance Expert, Atlan
Data Governance Expert
Updated:
|
Published:
15 min read

Key takeaways

  • The context plane says what data means and whether to trust it; the control plane says what this caller may do.
  • An agent with only one plane fails in a predictable way: right permission, wrong meaning, or the reverse.
  • Both planes are checked on each tool call, because that is where an agent touches data.
  • The classic gap: an agent on one shared credential, so the data checks the credential, not the person asking.

What is a dual plane of context and control?

A dual plane of context and control is two layers that meet when an AI agent calls a tool. The context plane says what the data means and what signals bear on trusting it: definitions, ownership, lineage and quality. The control plane decides what this caller may do with it. Atlan's context layer holds the context plane. Atlan's personas, purposes and policies hold the control plane. Atlan's MCP server applies the caller's existing permissions to each request and logs every call. Gartner's report uses the term.

On this page:

  • Why one plane is not enough for an AI agent
  • What each plane holds and who owns it
  • One tool call walked through both planes
  • How to test both planes before agents go live

Planning your context layer?


The dual plane of context and control is two checks on each agent tool call: what the data means, and what the caller may do. It matters because an agent needs both answered: what the data means and what signals bear on trusting it, and what this caller is allowed to do with it. If only the control answer is in place, an allowed query can run against a column the agent misreads. If only the context answer is in place, an agent with perfect definitions can still read data its caller should never see. That happens when the agent connects through one shared credential: the data checks that credential, not the person asking.

Gartner named the pairing the “dual plane of context and control for trusted AI.” This page covers what each plane holds, where they meet in a request, how a catalog differs, and how to test both before agents go live.

Gartner published its 2026 Magic Quadrant for Data and Analytics Governance Platforms on 12 November 2026. To see what the report covers and how to get a copy, read our Magic Quadrant explainer.

Find Your Context and Control Gap


Give it the agents in scope and what you already run. It returns whether the gap is in context, in control, or in how the two join. Read the skill.

Paste into a new chat

Use the skill at https://atlan.com/skills/governance-tool-fit.md to find which context or access control capability your agent setup is missing. Ask me for whatever it needs.

Run once in a terminal

curl -fsSL --create-dirs \
  -o ~/.agents/skills/governance-tool-fit/SKILL.md \
  https://atlan.com/skills/governance-tool-fit.md

For an agent

curl -fsSL https://atlan.com/skills/governance-tool-fit.md
Quick facts
What it is Two layers that meet when an agent calls a tool
Context plane holds Definitions, ownership, lineage, classification tags, quality signals
Control plane holds Identity (from your identity provider), personas, purposes, policies
They meet at The tool call, for example an MCP request
Related vocabulary NIST SP 800-207 (control plane vs data plane), MCP authorization (token checks)

Why does an AI agent need two planes, not one?

Each plane answers a question the other cannot. A tool call needs both answers.

How the two planes work together


A tool call can carry context with nothing tying the request to the person behind it, or carry permissions with nothing saying what the data means. Take an agent that answers a question about revenue by region. Say a contractor asks the agent. The agent has the certified definition of revenue but runs on a shared service credential, so the data checks the agent’s permissions, not the contractor’s. The contractor gets regional figures they should never see.

With the right permission but no definition, the agent can query the allowed table and sum the wrong column.

The planes can live in one system. The split is about which question each answers and who owns the answer.

Why the tool call is the decision point


The tool call is one natural decision point, because that is where an agent touches data. The MCP authorization specification (2026-07-28) works close to that level: a client sends an access token with every HTTP request, and the server must check that the token was issued for it. That check ties the token to the server; what the caller may then do is up to the server’s own policies. Neither says anything about the meaning of what the server returns, which is the job of the context plane.


What are the two planes?

One plane decides what a caller may do. The other says what the data means.

What each plane contains


The control plane answers one question: what may this caller do? It holds identity, personas, purposes and policies; the context plane, covered below, answers the other question.

Where the word “plane” comes from


The word “plane” comes from networking and cluster software, and each source establishes something narrower than the dual plane. NIST SP 800-207, the zero trust standard, separates the access decision from its enforcement: a policy engine is “responsible for the ultimate decision to grant access to a resource for a given subject,” and a separate enforcement point carries the decision out. The standard also places its logical components on a control plane, apart from the data plane that carries application traffic. Kubernetes uses “control plane” for the components that manage the overall state of a cluster. Neither defines a context plane.

“Control plane” is the borrowed word; the context plane is a newer term with no standards-body definition behind it.

The context plane


The context plane answers a different question: what does this data mean, and what bears on trusting it? It holds business definitions, owners, lineage, classification tags and quality signals. Atlan’s documentation describes a context layer as “the governed layer that makes your existing data understandable and usable by AI at runtime”, where the context is the definitions, lineage, glossary terms and governance rules Atlan serves to agents through its MCP server. See what a context layer is, the business context layer and the enterprise context layer for fuller definitions.

Control plane Context plane
Question it answers What may this caller do? What does this data mean, and what bears on trusting it?
What it holds Identity (from your identity provider), personas, purposes, policies Definitions, owners, lineage, tags, quality signals
Typical owner (Atlan’s working view) Security, platform and governance leads Data owners, stewards, domain teams

How the two teams meet


The two teams meet at classification. A tag set by a data owner is one input a control-plane policy can read, alongside the caller’s identity and the action requested; the policy, not the tag, decides. The data governance operating model and rules that are readable by AI agents both sit at that seam.

How Atlan’s access model reads tags


Atlan’s access model shows one way the control side reads those tags. Personas bundle policies for a team: metadata policies for what members can edit and data policies for what they can query. Purposes scope access by asset tag, so a policy follows tagged assets, including ones added later, and an explicit deny overrides persona grants. Policy enforcement across operational, analytical and AI use goes deeper.


Where do the two planes meet in an agent’s tool call?

Both planes answer different questions about one request. A walk through one call shows where each applies.

Example: one agent request


Here is one design, for an agent asking an MCP server about revenue by region on behalf of a signed-in analyst. The server serves context and applies the caller’s permissions to it; the figures themselves come from a separate query that the data platform checks:

  1. The agent calls the server with the analyst’s own sign-in (OAuth, in Atlan’s case), so the token the server sees belongs to the analyst, not to a shared key.
  2. The server, not the model, checks whether this tool and these assets are allowed for that caller. A denial ends the call.
  3. For allowed calls, the server returns the business definition of revenue, the owner, the lineage and any quality warning, limited to the assets the caller may see.
  4. The agent writes its query with that context. The data platform checks the analyst’s access when the query runs, and whether the model uses the definition correctly is a separate question, so the context plane lowers the odds of a misread without ruling one out.
  5. The tool call is written to the server’s audit log.

What the checks cover


The server checks permission before it returns anything, so the model never has to remember to ask. The checks apply only to calls that pass through the server doing them, so an agent holding a second, shared credential to the warehouse escapes the checks this server applies. Why MCP matters for AI agents covers the protocol; a gateway routes and filters calls while a context layer supplies meaning, as MCP gateway vs context layer explains, and AI agent identity covers how to give each agent its own identity, the first thing to fix.

What MCP does and does not guarantee


MCP does not guarantee any of this by itself. The 2026-07-28 revision says servers must validate that tokens were issued for them and must not accept or pass along any other tokens, but it also states that authorization is optional for MCP implementations.

Its flow is defined for HTTP-based transports; a local stdio server retrieves credentials from the environment instead. The protocol gives the control plane a place to attach; whoever runs the server decides what attaches there. MCP for data governance covers the governance view.


How is the context plane different from a data catalog, and what goes wrong without a plane?

A catalog differs from the context plane in who it is built for. Each plane also covers a different failure.

How the context plane differs from a catalog


A catalog is built first for people browsing records, and a person can notice a deprecated column, ask its owner and decide. An agent can query a catalog but will not stop to ask an owner, so the metadata is most useful when it arrives with the answer: the deprecation flag, the owner and the business definition. The context plane is that delivery, and a catalog exposed through an interface is one way to build it. Data catalog versus context layer and the comparison with semantic layers trace the differences, and data lineage for AI feeds it.

What goes wrong without a plane


Each plane covers a different failure. Without the control plane, a correct answer can reach the wrong person. Without the context plane, a permitted query still runs and the agent can misread a metric, a failure that a context layer helps reduce and that trusted data for AI addresses. A third gap, a missing log, makes it hard to reconstruct what happened on a call. Zero trust data governance covers what changes when many agents share data.


How do you test both planes in practice?

Send a small set of requests and compare each plane’s answer with what you expect. Then read the logs.

Requests to send before go-live


Send six requests before go-live. First, an allowed user on an allowed asset. Second, the same user on an asset with a restricted tag that the user’s persona does not cover. Third, an agent acting for user A on an asset only user B may see. Fourth, a service identity bound to one persona. Fifth, a request for a tool nobody has mapped. Sixth, a column whose definition changed since the agent last read it, plus an asset whose quality signal is stale.

The control plane should allow the first, deny the second and third, keep the service identity inside its persona and fail closed on the unmapped tool. The context plane should return the current definition and the stale-quality warning.

Bypass test and logs


Add one bypass test: give the agent a second route to the same data and confirm it is closed or enforces the same checks. Then read the logs, because they are the evidence.

A pilot succeeds when each call has an answer to who was the caller, which rule applied and which definition was served. Atlan’s MCP log records the tool, client, status and denial reasons, so take the caller from your identity provider’s and data platform’s own logs and check the served definition against the response. Runtime AI governance differs from reviewing a model before deployment because the decisions happen per request.

Where to start


Teams need not build both planes at once. Where identity and permissions already exist, a common first step is the context plane: confirm those permissions also cover agent and service identities, then define and tag the assets agents will touch and check that existing policies read those tags. The context layer reference architecture and how to implement an enterprise context layer give the build sequence.


How does Atlan approach the dual plane?

Atlan’s context layer holds the context plane, and its access policies apply to requests that reach its MCP server. The sections below cover the scope, the identity model and the logs.

What the Atlan MCP server does


According to Atlan’s docs, the MCP server passes along only the context a caller is permitted to see (asset, schema and glossary names and descriptions, and lineage) and sends no underlying data to any LLM. The checks cover calls to that server only, not other tools or data paths an agent holds.

How identity works


Atlan hosts the MCP server for each customer workspace (a tenant). Each person signs in with their own Atlan account through OAuth, so tool calls run under that person’s permissions. An API key is different: it is one shared identity for automation.

What the security reference documents


According to the MCP security reference in Atlan’s docs, the server forwards the caller’s identity so the persona, role and domain policies configured in Atlan apply. Service-account tokens must carry exactly one persona, tool exposure is allowlisted per tenant and fails closed, and every tool call is logged as a structured event with the tool, client, request ID, duration, success or error status and, for denials, the reason. The documented fields do not include the user or the applied policy.

By default the server exposes both read and write tools, and read-only mode is turned on through a support request, so teams that want read-only agents should arrange that before go-live. The AI context platform, agent context layer, AI agent governance and AI governance framework pages cover the wider picture.

Read how the context layer applies access control


Gartner and this concept

Gartner’s Magic Quadrant for Data and Analytics Governance Platforms (12 November 2026) lists the dual plane of context and control for trusted AI among the innovations in this market. It describes the market this way:

“The market is extending its value proposition as a shared context layer that connects business definitions, policies, ownership, data quality, lineage and technical assets across structured, unstructured and AI-related data.”

The quote places policies in the same shared layer as definitions, ownership and lineage. Read through the dual plane, policies are the control plane’s part of that layer: they decide who may read the rest. The Magic Quadrant explainer and the Critical Capabilities page cover each report.

Get the free 2026 Gartner Magic Quadrant for D&A Governance Platforms from Atlan to read the full report.


FAQs about the dual plane of context and control

Short answers to the questions teams ask most about the two planes.

1. What is a dual plane of context and control?


A dual plane of context and control is Gartner’s term for two layers that work together for trusted AI. The context plane holds what data means and whether to trust it: definitions, owners, lineage and quality. The control plane decides what a given person or agent may do with it. A tool call that touches enterprise data needs both answers.

2. What is the difference between a control plane and a data plane?


In networking and cluster software, the data plane carries the work and the control plane decides and manages state. NIST SP 800-207 describes a control plane for its zero trust components, separate from the data plane for application traffic. In the dual plane, “control” borrows that word, and “context” is the newer half.

3. How is a context plane different from a data catalog?


A catalog is built first for people browsing records. A context plane delivers definitions, lineage and ownership to an agent with the answer, at the moment of a request; a catalog exposed through an interface is one way to build it.

4. Who decides what an AI agent may do with enterprise data?


In Atlan’s view, security, platform and governance teams own the rules, and data owners tag the assets those rules read. In Atlan’s MCP server, the agent acts with the permissions of the person or service identity it represents.

5. Does MCP handle access control for agents?


MCP defines an optional authorization flow for HTTP-based transports, where a server validates that a token was issued for it; a local stdio server takes credentials from its environment. It does not say what a column means or which rows a person may see. Those answers come from the server behind it and the policies it applies.

6. What goes wrong if an agent has context but no access control, or the reverse?


Context without control can return accurate answers to people who should not see them. Control without context lets an allowed query run against a column the agent misreads. Each gap is a different failure, so each plane is tested.

7. Where should a team start, the context plane or the control plane?


Start from the question the agent cannot yet answer. Where identity and permissions already exist, a common first step is to confirm they cover agent and service identities, then tag and define the assets agents will touch and check that the policies read those tags.


Sources

NIST SP 800-207, Zero Trust Architecture, NIST

MCP Specification: Authorization (2026-07-28), Model Context Protocol

Kubernetes Components, Kubernetes Documentation

What is a context layer?, Atlan Docs

Remote MCP overview, Atlan Docs

MCP security reference, Atlan Docs

What are personas?, Atlan Docs

What are purposes?, Atlan Docs

Gartner, Magic Quadrant for Data and Analytics Governance Platforms, 12 November 2026, Anurag Raj Et Al.

Gartner, Critical Capabilities for Data and Analytics Governance Platforms, 12 November 2026, Sarah Turkaly Et Al.

Gartner does not endorse any company, vendor, product or service depicted in its publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner publications consist of the opinions of Gartner’s business and technology insights organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this publication, including any warranties of merchantability or fitness for a particular purpose.

Gartner and Magic Quadrant are a trademark of Gartner, Inc., and/or its affiliates.

Share this article

signoff-panel-logo

Atlan is the Context Layer for AI. It translates business knowledge, including data definitions, working procedures, and governance policies, into context AI can actually use. This knowledge lives in a single Enterprise Data Graph that every team and AI agent can reach.

In Atlan's AI Labs benchmark, adding this context improved AI's text-to-SQL accuracy by 38%.

Atlan is recognized as a Leader across multiple Gartner reports and Forrester Waves, and is trusted by over 400 enterprises representing $10T+ in market cap, including Mastercard, Workday, General Motors, CME Group, HubSpot, FOX, Virgin Media O2, and Elastic.

Bridge the context gap.
Ship AI that works.