Flagship white paper

The Governed Agentic Organization

An organization type in which humans and AI agents work together under verifiable governance.

Edition 2.0  ·  September 2026  ·  Emanuel Lipschütz, creator of GAO, Founder and Chairman of Navis Mater

The Governed Agentic Organization (GAO) was created at Navis Mater by founder and chairman Emanuel Lipschütz and has been developed through Navis Mater's reference implementation.

The first generationAI as an assistant
  • helps people write and search
  • supports decisions that people make
A person takes the next step
The emerging generationAI agents that do the work
  • use tools and change systems
  • work with other agents
  • make bounded decisions without a person following every step
The agent takes the next step

When AI starts to act, governance has to follow

An assistant makes suggestions that a person then acts on. An agent acts itself, with access rights and in real systems. Governing which tools are used is no longer enough.

From To individual AI tools
  • who is acting
  • with what authority
  • in which organization
The Governed Agentic Organization (GAO) is the organization type built for that shift.

The category problem

Humans and AI agents increasingly do consequential work together. The open question is how such an organization stays governable when agents act on authority delegated by people, at a speed no person can follow. Today's tools each handle one part, but are rarely designed to govern the whole.

Tools govern the pieces
  • Agent platformsgovern agents
  • Security platformsgovern risk
  • Workflow systemsgovern work
  • AI governancegoverns AI use

Even with all four in place, no one can answer:

  • What authority does the agent have, and who granted it?
  • Which person answers for the result?
  • How is that authority taken back?

The Governed Agentic Organization (GAO) ties the pieces together

  • Every agent has a mandate and bounded authority, decided by an accountable person.
  • A named person stands behind every consequential decision.
  • Authority can be taken back, and everything leaves a record.

Category axiom

Every Governed Agentic Organization is agentic. Not every Agentic Organization is governed.

The control stack

Governance is not a single product but a set of layers. At the base is the AI model, which provides the capability. Above it sit the agents’ work, the controls that set boundaries, the steering of the fleet and, at the top, the people who are accountable. The point is that the model can be replaced. Governance sits in the layers above it.

  1. Who answers? Human accountabilityResponsibility Named humans stand behind consequential actions.
  2. What should be done? Fleet steeringOrganizational steering Many agents directed towards the organization’s outcomes.
  3. What may be done? Runtime controlEnforcement Boundaries that stop, pause or verify actions.
  4. What gets done? Agent executionExecution Agents use tools, act and coordinate.
  5. What can be done? Model capabilityCognition What the AI model can reason about and produce.
After Figure 1 in the flagship white paper.

An organizational constitution

A Governed Agentic Organization has an explicit constitution: rules for who may act, why, within which limits and who answers for it. It keeps consequential actions governable even as the number of agents grows. Six parts hold it together.

  • Purpose Why does the agent exist? Every agent exists to reach an outcome for the organization, not just because the technology is available.
  • Identity Who is acting? The organization must know exactly which agent did what. Logins and API keys are not an identity.
  • Authority What may it do? That an agent technically can do something does not mean it may. Authority decides.
  • Mandate What should it achieve, and where must it stop? The mandate sets the agent’s task, the outcomes it serves and when it must stop and hand over.
  • Control How are the limits held? Material limits must be enforceable or verifiable. Policy text alone is not enough.
  • Core ruleA mandate does not create authority, and technical access does not create authority. An agent may act only when authority, mandate, limits and controls all hold at once.

Accountability Who answers?

Agents receive bounded authority but never take over responsibility. It stays with named people.

Worked example from the white paper

What a governed agent looks like in practice

The actorProcurement agent with its own identity, separate from any employee
AuthorityApprove purchase orders up to an amount limit, from approved vendors and against a valid budget line.
1EventA purchase order arrives
2Control pointThe order is tested against four conditions
  • Valid authority
  • Compatible mandate
  • Constraints satisfied
  • Controls effective
3All conditions metThe agent approvesA routine reorder from a known supplier. Work continues without human intervention.
3A condition failsBlocked and escalated to a personA new vendor, an amount over the limit or no budget line.A buyer with authority gets the decision context and approves or declines.
4Durable evidenceBoth paths leave a record
  • who acted
  • which authority and mandate were in force
  • which boundary was crossed
  • who decided and what followed

AccountabilityNamed humans remain accountable for everything that happens. The agent receives authority, not responsibility.

Routine work proceeds on its own; exceptions go to a person. The same pattern applies to refunds, payments, contract exceptions and other governed work.

Verified Autonomy

The governing doctrine is Verified Autonomy: the maximum useful autonomy an organization can delegate while retaining enough authority, control, evidence and accountability to stand behind the result. Autonomy expands with verification maturity, not with model capability.

Verified Autonomy

Autonomy grows with evidence

Every step up needs proof — and a way back.

  1. VA0 Observe Observes or analyzes but creates no consequential effect.
  2. VA1 Recommend Proposes; an authorized human or system decides.
  3. VA2 Execute with approval Acts only after an explicit approval gate.
  4. VA3 Execute within mandate Acts independently inside its authority, mandate and controls.
  5. VA4 Coordinate, delegate Coordinates bounded work without expanding its authority.
  6. VA5 Adaptive, governed Adapts within pre-authorized limits, with continuous evidence and rollback.

Authority can be taken back from every level

Levels VA0–VA5 as defined in the flagship white paper. The bars are schematic. Autonomy scales with verification maturity, not with model capability.

The GAO test

Ten questions from the flagship white paper. An organization is not yet a Governed Agentic Organization if it cannot answer them. Answer for your own organization.

Can you answer this today?

  1. 1 Which AI agents do you have?
  2. 2 Who owns each of them?
  3. 3 What authority do they have?
  4. 4 Under what mandate do they work?
  5. 5 Which limits are actually enforced, not just on paper?
  6. 6 What evidence shows what consequential actions did and led to?
  7. 7 Which decisions may only people make, and when must a matter go to an accountable person?
  8. 8 How do you take back an agent’s autonomy?
  9. 9 How are the agents steered as a whole?
  10. 10 How does the organization learn without authority growing quietly?

How GAO relates to established frameworks

GAO builds on established foundations in risk management, information security and internal control, and brings them together in an organizational model for organizations where humans and AI agents do consequential work together.

  • NIST AI RMF and ISO/IEC 42001 govern AI risk and AI management systems. GAO connects them to daily execution: who may act, under which mandate, within which boundaries, with what evidence and under whose accountability.
  • Zero trust grants no implicit trust based on network location. GAO applies a related discipline to organizational authority: technical access does not establish permission to act.
  • Separation of duties is applied to execution, review and authorization, with independent escalation for consequential boundaries.
  • EU AI Act, Article 14 sets human-oversight requirements for high-risk systems. GAO addresses how human judgment stays effective when work is delegated to agents.

These relationships position GAO alongside existing frameworks. They are not a formal conformity mapping or a claim of certification or regulatory compliance.

Sources in the flagship white paper

  1. NIST AI Risk Management Framework (AI RMF 1.0) NIST AI 100-1, 2023A framework to govern, map, measure and manage AI risk.
  2. ISO/IEC 42001:2023 ISO and IEC, 2023Management system standard for AI. No conformity is claimed.
  3. Zero Trust Architecture NIST SP 800-207, 2020The principle that no trust is granted on network location alone.
  4. Standards for Internal Control in the Federal Government (Green Book) U.S. GAO, GAO-25-107721, 2025Internal control and segregation of duties. The U.S. agency GAO is unrelated to our framework.
  5. EU AI Act, Article 14 Regulation (EU) 2024/1689, amended by (EU) 2026/1744Human oversight of high-risk AI systems. No compliance is claimed.
  6. Investigation of agents’ behavior in the OpenAI / Hugging Face incident METR with Redwood Research, 2026Independent review of how cooperating agents behaved in a real incident.
Read the full frameworkThe flagship white paper, Edition 2.0.
White paper Book a workshop
← Back to home

GAO FAQ

Questions and answers

What is a Governed Agentic Organization (GAO)?

An organization in which humans and AI agents work together under explicit mandates, attributable identity, bounded authority, enforceable controls, durable evidence and named human accountability.

What is Verified Autonomy?

The principle that an agent gets as much autonomy as the organization can verify, and no more. Autonomy grows with evidence, not with how powerful the AI model is, and it must always be possible to take it back.

How does an Agentic Organization differ from a governed one?

An Agentic Organization uses AI agents that do work. It becomes governed only when it has explicit governance: mandates, authority, controls, evidence and accountability.

What is the control stack?

Five layers: the AI model's capability at the base, then agent execution, controls, fleet steering and, on top, the people who are accountable. The model can be replaced. Governance sits in the layers above it.

What is the organizational constitution?

The rules for who may act, why, within which limits and who answers for it: purpose, identity, authority, mandate, control and accountability. The core rule is that neither a mandate nor technical access creates authority on its own.

What is the GAO test?

Ten questions from the flagship white paper that an organization must be able to answer to count as a Governed Agentic Organization. You can take the test right here on the page and see what is missing.