Constitutional Polyagentic Leadership
Introducing a governance framework for the age of AI

Constitutional Polyagentic Leadership

assigns leadership authority according to demonstrated ability and applies the same limits to every leader, whether a person, a person paired with a machine, or a machine.

A framework for organizations in which carbon-based and silicon-based managers lead in concert.

Read more
What is Constitutional Polyagentic Leadership?

A new theory for sharing leadership authority among people and AI

by Merlin Dominik Jung, based on the working paper with Michael Neubert1

polyagentic / pol-ee-ay-JEN-tik / adjective
  1. describing a system in which several kinds of agents can lead;
  2. in leadership: having humans, human–AI leadership units, and autonomous AI agents hold recognized decision rights at the same time.
constitutional / adjective
  1. governed by an internal set of rules that says who may decide, who can challenge a decision, and who must respond when something goes wrong. It refers to a company's charter and has nothing to do with a national constitution.
Working definitions, Neubert & Jung (2026)

Until recently, AI was a tool. Someone asked, it answered, and a person decided. That is changing. AI systems can now be given a goal and pursue it through a sequence of actions: assigning work, committing resources, and shaping what other people do next.

Once an organization allows that, the useful question is no longer whether the AI's answer is any good. The question becomes who is allowed to decide here, and what happens when they get it wrong. That is a governance question, and existing leadership theory answers only part of it.2

Humans Human–AIunits AutonomousAI agents LINKED RIGHTS Clear rules &dignity Capability-basedauthority Calibratedautonomy Explainability &contestability Traceableaccountability
Three kinds of leaders hold connected, revisable decision rights. Five governance dimensions apply to all of them.

Constitutional Polyagentic Leadership (CPL) is a proposed rulebook for that situation. It governs an arrangement the paper calls AFTER-NEXT leadership, in which three kinds of leaders work in one connected system. CPL has five dimensions: clear rules and respectful treatment, authority based on demonstrated ability, limits on independent action that match the risk, explanations with effective appeals, and responsibility backed by decision records.

Authority should follow demonstrated ability, remain within clear limits, and change when evidence requires it. THE CENTRAL PRINCIPLE OF CPL

Three kinds of leaders

The defining feature of the arrangement is that all three hold recognized decision rights, that those rights depend on one another, and that the organization can revise them. The mere presence of AI does not qualify.

  • Humans who hold leadership authority and direct others toward shared goals.
  • Human–AI leadership units: a continuing pairing of a person and an AI component that holds decision rights together. Either side can withhold approval and request review. Consulting a chatbot regularly does not make one.
  • Autonomous AI agents that choose among actions and direct collective work inside an approved area. A system that only carries out a fixed instruction has no discretion and is not leading.
Hypothetical example · one decision, three leaders

A plant manager sets delivery priorities. A planning engineer and an AI component jointly approve production changes. A scheduling agent assigns work within the approved plan and can reject requests that exceed capacity or break operating rules, including the manager's. The manager can request changes but cannot bypass the process.

Five decision rights replace the org chart

Adding agents to an organization chart says little about who may do what. CPL maps five rights for each decision and names the holder of each. Executing an instruction is not leadership. Exercising discretion that directs other people's work is.

Initiatestart a decision process
Recommendpropose an action
Decideapprove a course of action
Executecarry it out
Overseereview, and require correction or suspension

Who governs the rules

CPL proposes a governing body of accountable human decision makers with meaningful representation for the people affected. It sets appeal routes, the procedure for changing permissions, and the power to stop harmful activity. Every AI leadership role has a named human or organizational owner who can arrange correction, and reviewers of a disputed decision are independent of it. No leader may expand its own permissions outside the approved process.

CPL treats AI leadership as an organizational role. It does not assume that AI has consciousness or legal personhood.

What CPL is, and what it is not yet

CPL is a conceptual framework. Its combined effects have not been tested empirically. The published productivity results it builds on come from AI assistance, and none of them measures autonomous AI leadership. The paper states seven propositions and a research agenda designed so that the model can be rejected as well as supported. Treat it as a design template to pilot before anyone treats it as a playbook to roll out.

More information

  1. Neubert, M., & Jung, M. D. (2026). After-Next Leadership: A Conceptual Framework for Constitutional Polyagentic Leadership (working paper draft). EIM – European Institute of Management, Malta.
  2. Van Quaquebeke, N., & Gerpott, F. H. (2023). The now, new, and next of digital leadership. Journal of Leadership & Organizational Studies, 30(3), 265–275. Shrestha, Y. R., Ben-Menahem, S. M., & von Krogh, G. (2019). Organizational decision-making structures in the age of artificial intelligence. California Management Review, 61(4), 66–83.
The framework

Five dimensions of CPL

Each dimension answers one practical question. They stay separate on purpose: rules can be clear and still disrespectful, and an explanation can be understandable without allowing an appeal.

Are the rules clear and respectful?

Constitutional clarity and dignity protection

People can understand the goals, decision rights, limits and review procedures, and know who may change them. Those affected are treated as persons with concerns and rights, never only as scores or production inputs.

In practice: written permissions and limits, employee voice, safeguards against dehumanizing practices.

Who is best able to decide?

Capability-based authority allocation

Decision rights go to the leader with demonstrated ability for the specific task. A title establishes responsibility. It does not prove competence for every decision, and neither does general confidence in people or in technology.

In practice: test before delegating, record why a right was assigned and what evidence would change it, check where the data came from.

How much freedom should that leader have?

Calibrated autonomy

Freedom to act matches demonstrated ability and the consequences of error. Routine, reversible decisions need little prior review. Consequential ones that are hard to repair need stronger checks.

In practice: the autonomy envelope, stating objectives, permitted resources, prohibited actions, escalation triggers and rollback.

Can others understand and challenge the decision?

Reciprocal explainability and contestability

All three kinds of leaders give reasons others can check: evidence, assumptions, uncertainty and value judgments. Affected people have a usable appeal route with an independent reviewer who can change the outcome.

In practice: the aim is accurate trust. Confidence may fall once people have seen the evidence, and that can be the right result.

Who must correct failures?

Traceable accountability

Records show who recommended, approved, executed and reviewed an action. A named owner with real resources and authority must explain and repair. Blaming the agent does not discharge that duty.

In practice: decision lineage, covering data used, AI version, permissions, overrides and completed corrections.

And how much control is too much?

Proportionate governance

A qualifier on all five. Controls should match the decision's risk and leave room for discretion and experiment. Recording consequential decisions is different from watching every exploratory activity.

In practice: protected space to test alternatives within approved limits, and periodic removal of approval steps that add nothing.

Where CPL applies

From NOW to AFTER-NEXT

Van Quaquebeke and Gerpott describe three stages of digital leadership. The working paper adds a fourth ideal type, and CPL is the proposal for governing it.

StageRole of AIWho holds authority?Main leadership question
NOWDigital tools support communication and coordination.Humans direct the work.How can digital tools help leaders?
NEWAI analyzes information and recommends actions.Humans retain final authority.How can AI improve leadership by humans?
NEXTAI carries out one or more leadership functions.Autonomous AI agents have authority in a defined area; humans may oversee them.Which leadership functions can AI perform?
AFTER-NEXTHumans, human–AI leadership units and autonomous AI agents work in one connected system.All three hold linked decision rights that can be revised.How should different leaders share and govern authority?

These are theoretical ideal types for comparing arrangements. They are not maturity levels, and no organization is required to pass through them in order.

Why it matters

AI assistance works. Combining people and AI is not automatically better.

Adding a human approval step to every AI decision can add delay and error. Removing human judgment can weaken decisions too. The design task is to work out who is better at each particular decision.

40%

less time on professional writing tasks with access to generative AI, with assessed quality up 18%.

Noy & Zhang, 2023
15%

more issues resolved per hour after an AI assistant was introduced among 5,172 customer-support workers.

Brynjolfsson et al., 2025
106

experiments reviewed: human–AI combinations beat humans alone on average, yet did worse than the better of the two acting alone.

Vaccaro et al., 2024

These studies concern AI-assisted tasks. They do not test leadership by autonomous AI agents and do not establish the benefits of CPL.

What the paper predicts

Seven propositions to test

Each is a reasoned prediction, stated so that research can support, qualify or reject it.

  1. P1Assign authority according to demonstrated ability.Decisions improve when rights follow tested ability for the task instead of rank or a general preference for people or AI.
  2. P2Use clear rules and protect respectful treatment.Stakeholders see authority as more justified when rules are clear, consistently applied, and people have a voice.
  3. P3Set limits that match ability and risk.Matched limits reduce avoidable harm, provided someone has the power to intervene when a limit is reached.
  4. P4Provide checkable explanations and effective appeals.Explanations help people calibrate confidence; appeals that can correct decisions strengthen legitimacy.
  5. P5Connect decision records to corrective action.Records aid recovery only when named owners use them to fix problems and revise permissions.
  6. P6Distinguish informed reliance from cooperation.Trust in a particular leader and acceptance of the whole arrangement are separate pathways.
  7. P7Match the arrangement to the decision.Agents suit structured, reversible decisions under time pressure; joint units suit complex, sensitive ones; independent review earns its delay as stakes rise.
In practice

Getting started in one decision area

The paper recommends introducing the arrangement in stages and expanding only when the agreed criteria are met.

  1. 01Pick one decision area.Record a baseline before changing anything.
  2. 02Map the five rights.Who initiates, recommends, decides, executes and oversees today, in practice.
  3. 03Test ability first.Compare humans, human–AI units and agents on representative tasks, and note what evidence would withdraw a right.
  4. 04Approve the limits.Resources, escalation triggers, stop powers and a fallback process.
  5. 05Name the owner.Check that they can stop, correct or transfer the work.
  6. 06Open an appeal route.With an independent reviewer who can change outcomes.
  7. 07Involve affected employees.They know where the arrangement is likely to fail.
  8. 08Judge the whole result.Accuracy, total time from request to implementation, workload, cooperation and harmful errors together.
Resources

The working paper

After-Next Leadership: A Conceptual Framework for Constitutional Polyagentic Leadership

Michael Neubert and Merlin Dominik Jung. EIM – European Institute of Management, Malta. Working paper draft, October 2026.

The paper is a draft in progress. Details of where to read it will be posted here on publication.

Selected references
  • Baird & Maruping (2021). Delegation to and from agentic IS artifacts. MIS Quarterly.
  • Bernstein (2012). The transparency paradox. Administrative Science Quarterly.
  • Brynjolfsson, Li & Raymond (2025). Generative AI at work. Quarterly Journal of Economics.
  • Grote, Parker & Crowston (2026). Control–accountability alignment. Academy of Management Review.
  • Lee & See (2004). Trust in automation. Human Factors.
  • Noy & Zhang (2023). Productivity effects of generative AI. Science.
  • Shrestha, Ben-Menahem & von Krogh (2019). Decision-making structures in the age of AI. California Management Review.
  • Vaccaro, Almaatouq & Malone (2024). When combinations of humans and AI are useful. Nature Human Behaviour.
  • Van Quaquebeke & Gerpott (2023). The now, new, and next of digital leadership. JLOS.
About

Merlin D. Jung

Merlin D. Jung is a senior technology executive with more than 25 years of international experience in enterprise IT, most of it in leadership roles. He leads large-scale, agentic-AI-driven transformation programs for global clients and has lived and worked across Europe, the United States, South Africa and China.

A regular speaker at international industry forums, he addresses the leadership the agentic AI era demands: how organizations should lead when carbon-based and silicon-based managers hold authority side by side. His doctoral research develops a leadership framework for autonomous multi-agentic AI, and he co-authored the CPL working paper with Michael Neubert of EIM – European Institute of Management.

Merlin holds a Master's degree and a BA Honours in Sociology from the University of Port Elizabeth (now Nelson Mandela University) and is an alumnus of Harvard Business School's Advanced Management Program. He serves on the Board of the Harvard Business School Association of Germany.

Requests

Speaking and advisory

Talks, workshops and advisory conversations on leadership in the age of AI:

  • Who decides? Sharing authority among people and AI agents
  • From org chart to decision-rights map
  • Autonomy envelopes, appeals and accountable owners in practice
  • Governance that prevents harm without suppressing initiative
Email: connect@constitutionalpolyagenticleadership.com
LinkedIn: linkedin.com/in/merlinjung