Skip to main content

Graduated Trust Models for Data Governance

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

Key takeaways

  • A graduated trust model sets access in steps from five signals, not a single yes or no.
  • The five signals are certification, sensitivity, identity, purpose and track record.
  • Business metadata such as certificates and sensitivity tags is what the decision reads about an asset.
  • AI agents move up a ladder of autonomy with a clean record and move down when it breaks.

What is a graduated trust model?

A graduated trust model sets access strictness by how far an actor and an asset can be trusted, not by a yes or no. It reads five signals at request time: certification status, sensitivity, identity, purpose and track record. Atlan's context layer carries certification and sensitivity as business metadata, and Atlan personas and tag-based purposes enforce access for people and AI agents. The outcome can be full access, masked values, read-only, approval first, or denial.

On this page:

  • The definition, with the five signals it reads
  • How business metadata drives the decision
  • Static roles versus a graduated model
  • The autonomy ladder for AI agents, and how to lose a rung

Want the report that names it?


A graduated trust model sets access strictness by how far an actor and an asset can be trusted, not by a yes or no. It reads five signals at request time: certification status, sensitivity, identity, purpose and track record. The sections below define each signal, show where Atlan’s access controls fit and where a team’s own rules take over, and follow one invented agent as it moves up and down a ladder of autonomy.

Gartner’s 2026 Magic Quadrant for Data and Analytics Governance Platforms uses the term “graduated trust models.” The five signals and the steps below are our own way of making the idea concrete. The term is unrelated to the legal “data trust,” which the Open Data Institute defines as a legal structure that provides independent stewardship of data.

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.

Audit your agent guardrails


Give it an agent, the assets it touches, and the actions it can take. It returns which signals each decision reads, which rung the agent sits on, and the gaps to close. Read the skill.

Paste into a new chat

Use the skill at https://atlan.com/skills/agent-guardrails-audit.md to audit an AI agent's guardrails against trust signals. Ask me for whatever it needs.

Run once in a terminal

curl -fsSL --create-dirs \
  -o ~/.agents/skills/agent-guardrails-audit/SKILL.md \
  https://atlan.com/skills/agent-guardrails-audit.md

For an agent

curl -fsSL https://atlan.com/skills/agent-guardrails-audit.md
Quick facts
What it is Access that moves in steps with trust in the actor, the asset and the action
Signals it reads Certification status, sensitivity, identity, purpose, track record
Output Full access, masked values, read-only, approval first, or denial
Compared with Static roles decide once at grant time; this decides per request
Who sets it up Governance teams define the tiers; the platform supplies the tags and enforces access where requests pass through it

Why does a yes-or-no access decision break down for data and AI agents?

A single allow-or-deny answer per role gets hard to defend once an AI agent reads many assets of mixed quality. Give an analyst and an agent the same role, and a draft table and a certified table can both return full access, even though only one of them should feed a board report.

If you own a data or AI platform and set access for agents, you are making this decision now. Agents differ from people in pace, scope and reversibility, and the next paragraph shows how.

An agent can repeat a mistake on a stale asset on every run before anyone looks, and some of what it does cannot be undone. The OWASP AI Agent Security Cheat Sheet asks for the minimum tools per task, per-tool scoping such as read-only versus write, and explicit approval for high-impact or irreversible actions. A finer allow or deny rule can enforce each of those asks. A graduated model folds them into one decision that also reads the state of the asset.

How much autonomy an agent has tells you what it may do; a certificate (a status such as Verified or Draft set on an asset) tells you whether the asset is fit to use. Neither alone says whether this agent may write this table. AI agent access control and agent risks and guardrails cover the surrounding ground.


What is a graduated trust model?

A graduated trust model is access that tightens or loosens in steps according to trust signals about the actor, the asset and the action.

The levers


Strictness works like a set of levers rather than a gate: what the actor sees (full or masked values), what it can do (read-only or write), and whether a person signs off first, with denial as the floor. The levers combine, and the decision is made when the request arrives, using current signals.

The five signals


Five signals do the work:

  • The certification status of the asset: is anyone vouching for it?
  • The sensitivity of the data it holds.
  • The identity of the actor, whether a person or an agent, and what that identity is allowed to be.
  • The purpose of the request: why the actor is asking, such as the task it was given.
  • The actor’s track record on past actions.

Combining the signals


One way to combine them, offered as a design choice and not a standard: the strictest signal caps the outcome. A Deprecated certificate or a sensitivity tag limits what the request’s purpose or the actor’s track record can allow, and a strong record only moves an actor within that limit. For example, a Verified table with a sensitivity tag returns masked values to an agent on any rung, however clean its record. Reversibility decides when a person must sign off, since reading a summary and deleting a record carry different risk, and OWASP places explicit approval on high-impact or irreversible actions.

Where the idea comes from


The underlying idea is older than the term. NIST SP 800-207 describes trust algorithms for zero trust architecture. A criteria-based algorithm requires qualified attributes to be met. A score-based one computes a confidence level and compares it with the resource’s threshold; if the score falls short, NIST says “the request is denied, or access privileges are reduced (e.g., read access is granted but not write access for a file).” That reduced-privilege outcome is the graduated part.

Zero trust assumes no implicit trust. A graduated model is one way to apply its score-based algorithms (which compute a confidence level instead of checking fixed criteria) to data. It is a pattern within attribute-based access control (decisions made from attributes of the actor, the asset and the request), not a replacement for it: the attributes include the actor’s track record, and the outcomes come in steps.

Gartner and this concept


The 12 November 2026 edition of Gartner’s Magic Quadrant for Data and Analytics Governance Platforms uses the term; our Magic Quadrant explainer covers the report. The passage reads:

“Use the captured business metadata to support graduated trust models for federating high-level policies to technical data assets, allowing for dynamic and adaptive policy enforcement.”

Related reading: the AI governance framework, the context layer for data governance teams and runtime governance versus pre-deployment review.


How does business metadata drive a trust decision in the context layer?

Business metadata is the information people maintain about an asset, such as its certificate and sensitivity tag. A trust decision reads it, and a context layer stores it next to the asset so the decision can find it.

What the decision reads


The certificate and sensitivity tag are business facts that someone already maintains, so the model reuses them instead of inventing a new score. For the wider practice, see data trusted for AI.

Certificates


Atlan’s docs describe four certificate statuses: Verified, Draft, Deprecated and No certificate, each set on an asset’s profile. Certificates are filterable on the Assets page (Atlan documentation).

Atlan’s access-control FAQ says access policies can target connections and tags but not certification status (Atlan documentation), so the certificate signal reaches a decision through a team’s own rule or tooling, for example by mirroring Verified and Draft into a tag. A mirrored tag can go stale, so the team has to update it whenever a certificate changes. Data certification is where that practice starts.

Personas


Personas scope what a team sees and can do, including metadata editing, data querying and domain access. Personas combine, and an explicit deny overrides a grant (Atlan documentation).

Purposes


Atlan uses the word “purpose” for something different from the request-purpose signal above. An Atlan purpose is a policy attached to an asset tag, so it controls access from the asset side. The request-purpose signal is something a team captures separately, for example in the agent’s task. A purpose scopes access by asset tag instead of by team, applies to every asset carrying the tag including future ones, and its denies always win over persona grants.

A purpose can also mask values instead of denying outright, so an analyst sees *** where a full deny would have shut the door. Approval workflows can add grants but cannot lift a deny (Atlan documentation).

Sensitive data and track record


For data tagged as sensitive, a team can configure masked values or a deny. Track record sits on the actor side: it has no built-in field in Atlan’s access policies. Tags belong to assets and a record belongs to an actor, so a team computes the record from its own agent run logs and review outcomes, outside the catalog, and applies it by changing which persona the agent’s token holds.

Where enforcement happens


Atlan’s personas and purposes are enforced where a request passes through Atlan: its interface and its MCP server, under the caller’s identity. Request purpose and track record arrive as inputs the team supplies.

A query an agent sends straight to a warehouse never passes through Atlan, so the source system’s own controls apply there. Agents reach the same tags through the AI Context Platform, so a decision can read the live tags instead of a copy, though how current they are depends on how the team maintains them, as the reference architecture describes. The tiers are a practice your team builds with these capabilities, and the team decides what each level of trust earns.


How is a graduated trust model different from static role-based access?

Static role-based access fixes the grant by job role, so a role check reads only membership. A graduated trust model reads more attributes and can step access up or down.

Roles versus attributes


A role records a job function. A trust model also weighs what is being touched, why, and how risky the action is. NIST SP 800-162 defines attribute-based access control as authorization determined by evaluating attributes of the subject, object and requested operations, and in some cases environment conditions, against policy. A graduated model builds on that idea; the masking, read-only and approve-first outcomes are design choices layered on top. In production, roles are often paired with tag-based masking. The table below compares the role check alone. A graduated model adds three things on top: the certificate, the request context and the actor’s record.

Side by side


Question Role check on its own Graduated trust model
When do the inputs change? When membership or the grant changes Whenever a signal changes
What does it read? Role membership Actor, asset, purpose and action signals
What can happen? Allow or deny Allow, mask, read-only, approve first, deny
How does it change? Someone edits the role or its members Signals change, or the actor’s record changes
What is the audit story? Who holds which role Why this request got this level

Keeping roles in place


NIST SP 800-207 supplies the counterweight: it calls for balancing security, usability and cost-effectiveness when defining trust algorithms. Roles can stay as the base layer under a graduated model, and role-based access in a context layer shows how the two combine, while embedded governance and machine-readable governance rules cover how the signals stay current.

When are static roles enough?


Static roles may be enough when the data is low in sensitivity, changes slowly and only people use it; a team can judge that a five-signal model adds audit work that buys little there. Three failure modes are worth designing against: tiers written in a spreadsheet and never enforced, no path for demoting an actor, and approval requests on every action. See the guardrail tooling comparison for enforcement options.


How does an AI agent earn more autonomy, and how does it lose it?

An agent earns autonomy in steps: it works at a lower level, shows a clean record under review, and moves up. It moves back down when the record breaks. One way to build the ladder, framed by us and not a standard, has four rungs:

  1. Observe: reads and summarizes.
  2. Propose: suggests, and a person acts.
  3. Act with approval: prepares an action that a person approves before it runs.
  4. Act with audit: the agent acts, and a person reviews afterward instead of approving each action.

Outside sources on tiers and autonomy


Two outside sources are related, though neither uses these rungs. BCG’s Trust Protocol (December 2025) tiers agents from shadow mode, where “the agent suggests and the human acts,” up to full autonomy, and reverts a supervised agent to the lower tier if its confidence score drops below a set threshold (BCG). That trigger is the agent’s confidence score, not its record; the record-based rule below is our design. In “Levels of Autonomy for AI Agents” (2025), Feng, McDonald and Zhang argue that autonomy is a deliberate design decision, separate from what the agent can do (arXiv:2506.12469).

Promotion and demotion


Promotion should need a record over a window of actions that includes the riskier tasks of the next rung, because trust needs evidence. Demotion for a serious break should need no window, because each further run on a broken record adds potential harm; a team may treat minor errors differently. Both are design reasoning. Any threshold, such as how many clean actions earn a rung, is a number a team sets for itself, not a standard.

What Atlan controls constrain


Atlan’s controls constrain what an agent can do on the lower rungs; the approval step and the promotion review stay a team process. A service-account token for an agent needs exactly one persona, and OAuth calls run under the signed-in user’s identity, so MCP “never widens what a user’s login already permits.” Read-only mode, requested through a support ticket, disables write, admin and lifecycle tools while keeping search and lineage. Admins can scope the metadata-writing tools to admins only, and changes show in an asset’s activity history (Atlan documentation). A team can start a new agent on a short rung and widen its persona as its review record builds. Why MCP matters for agents and the MCP gateway versus context layer comparison go deeper.

Example: one agent request at three trust levels


This is an illustration, not a customer result.

Picture an invented agent that answers finance questions. It starts on rung one, reading Verified, non-sensitive assets only. After a clean review period it moves to rung two, and may propose from a Draft asset only with a flag attached. A request that touches a sensitivity-tagged column returns masked values at any rung.

Then a table the agent read as Verified is deprecated. The cap now blocks new reads of it, but the agent repeats a cached answer built on it, and this team’s rules count that as a broken record. The agent drops a rung until its next review. The data’s state changed first; the agent’s record changed because it kept using the old reading.


How does Atlan approach graduated trust?

Atlan stores certificate and sensitivity tags in the context layer, and personas and purposes enforce access for people and agents. Atlan is an AI Context Platform: it holds certificate status and sensitivity tags next to the assets they describe, so the inputs are available when a request arrives. A team then writes the rules that read them.

Agents reach the same controls through Atlan’s MCP server (the Model Context Protocol endpoint that lets an agent call Atlan’s tools under the caller’s identity), which offers read-only mode and an admins-only setting for the metadata-writing tools, useful when an agent needs to begin on a short rung. Trust tiers are the practice a team builds on top. For most teams the first step is small: pick one agent, one domain and one sensitivity tag, run the agent on the lowest rung, and review the activity history before moving it up. To go wider, see implementing an enterprise context layer, the AI context stack and the data governance operating model.

Get the free 2026 Gartner Magic Quadrant for D&A Governance Platforms from Atlan

Get the Gartner report

FAQs about graduated trust models

1. What is a graduated trust model?


A graduated trust model is access that tightens or loosens in steps according to trust signals about the actor, the asset and the action. Instead of only allow or deny, the outcome can be full access, masked values, read-only, approval first, or denial.

2. How is a graduated trust model different from role-based access control?


Role-based access grants permissions by job role, and the role check on its own reads only membership. A graduated trust model also reads current signals on each request, so one actor can be treated differently on a Verified table, a Draft table and a sensitivity-tagged column.

3. How does an AI agent earn more autonomy, and what makes it lose it?


In one design (ours, not a standard), an agent moves up a ladder (observe, propose, act with approval, act with audit) after a clean record over a window of actions. It moves down when the record breaks; what counts as a break, such as relying on a deprecated asset, is a rule each team sets.

4. Does every agent action need human approval?


Not necessarily. Reversibility and sensitivity are common factors. OWASP’s AI Agent Security Cheat Sheet asks for explicit approval for high-impact or irreversible actions, and each team decides which lower-risk actions run without it.

5. Is a graduated trust model the same as zero trust?


No. NIST SP 800-207 describes zero trust as assuming no implicit trust for assets or accounts. A graduated trust model is one way to decide how much access each level of trust earns.


No. A legal data trust is a structure for sharing data under a trustee. A graduated trust model is an access control approach that sets how much access an actor gets to a data asset at request time.


Sources

Zero Trust Architecture, NIST SP 800-207 (August 2020), NIST

Guide to Attribute Based Access Control, NIST SP 800-162 (updated 2019), NIST

AI Agent Security Cheat Sheet, OWASP

How Agents Are Accelerating the Next Wave of AI Value Creation (December 2025), BCG

Levels of Autonomy for AI Agents, arXiv:2506.12469

Defining a data trust, Open Data Institute

Add certificates, Atlan documentation

User management and access control FAQ, Atlan documentation

What are personas, Atlan documentation

What are purposes, Atlan documentation

MCP security, Atlan documentation

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