As the AI ecosystem absorbs a rare human-centered statement on dignity and the common good, agents are moving from assistance into authority

Ungoverned: AI Ethics & Governance for Leaders, Boards & Trustees

By Dr. Freddie Seba

Editorial Note

Issue #71 began with Magnifica humanitas.

That was intentional.

Pope Leo XIV’s first encyclical was not just another statement about technology. It was a rare, nearly book-length moral positioning document in the AI era — one centered on the human person, the dignity of work, truth, justice, peace, and the common good.

The AI ecosystem is still absorbing it. That matters. While society is making sense of a major human-centered intervention in AI, the market is moving quickly toward greater autonomy, more agents, more workflow authority, more financial action, more cyber capability, more code generation, more synthetic content, and greater institutional dependence.

That is the tension of Issue #72. Magnifica humanitas asks the human question: What kind of world are we building when machines begin to shape human life at scale? Agentic AI forces the operating question:

What authority are institutions already giving to AI systems — and what governance exists before that authority becomes normal?

The Vatican’s framing of Magnifica Humanitas is clear: technology is neither inherently evil nor neutral; it takes on the characteristics of those who design, finance, regulate, and use it. The encyclical also calls for responsibility, accountability, legal frameworks, independent oversight, user education, and shared ethical standards in AI governance. (Vatican News)

Issue #72 picks up the next part of that question. If Issue #71 was about human stewardship after capability, Issue #72 is about bounded autonomy before authority scales.

The governance problem is no longer only whether AI systems can generate content, answer questions, or support decisions.

The governance problem is now whether institutions can define:

What autonomy is permitted
What authority is bounded
What actions are reversible
What must be independently evaluated
What must be logged
Who owns the outcome
Who can stop the system

That is where AI Minimum Viable Governance — AI MVG, the central framework in my new book, Ungoverned: A Practical Guide to AI Minimum Viable Governance, becomes practical.

Not as an abstract framework.
Not as a policy slogan.
Not as compliance theater.

But as the minimum defensible governance floor institutions need before AI becomes normalized inside workflows that affect people’s lives.

From My New Book, Ungoverned

AI MVG is not only a framework. It is a culture shift made operational for leaders. In my book, Ungoverned: A Practical Guide to AI Minimum Viable Governance, I argue that the hardest AI governance failures rarely begin as scandals.

They usually begin more quietly.

A team moves quickly because the tool seems useful.
A vendor feature is turned on because the contract already allows it.
A pilot becomes routine because everyone is busy, and the tool appears to work.
A model output becomes trusted because it is fast, fluent, and available.
A workflow changes quietly because no one stops to ask whether authority has shifted.
A human review step becomes symbolic because the reviewer has neither time nor evidence to challenge the system.
A system gains access before the institution has named the authority it has delegated.

That is how AI becomes ungoverned. Not because people are careless.

But because institutional culture often rewards speed over proof, adoption over accountability, and convenience over governance. That is why AI MVG is not only a board tool. It is part of organizational culture in the AI era.

It asks whether the institution has built habits of:

Naming ownership
Questioning assumptions
Testing claims
Mapping access
Limiting authority
Monitoring use
Learning from incidents
Protecting people
Pausing when conditions change

AI MVG does not mean minimal governance. It means minimum defensible governance:

Enough structure to prevent irresponsible normalization
Enough clarity to assign ownership
Enough evidence to support deployment
Enough humility to admit uncertainty
Enough authority to pause or stop use when conditions change
Enough process to learn from incidents, near misses, and feedback
Enough discipline to improve as AI capabilities, risks, laws, and institutional realities change

That is culture. That is operating discipline. That is governance before failure forces governance. The book also draws from practitioner conversations across higher education, healthcare, financial services, digital health, technology, podcast interviews, newsletter analysis, Silicon Valley convenings, and cross-sector governance monitoring.

That matters for Issue #72 because the central question is no longer abstract: How do institutions preserve human dignity when AI becomes operational authority? That is the bridge between Magnifica humanitas and bounded autonomy.

Market Signal

The AI ecosystem is absorbing Magnifica humanitas while agents gain authority.

The most important signal from last week was not only the release of Magnifica humanitas, but also the release of Magnifica humanitas.

It was the reaction. The encyclical landed as a major moral intervention into an AI ecosystem dominated by product launches, model races, valuation stories, benchmarks, cyber concerns, and agentic workflows.

It placed AI inside a longer tradition of social teaching about:

Work
Dignity
Justice
Power
Technology
Truth
Peace
The common good

The board-level translation is simple:

Technology is not neutral. AI is not neutral.
Governance is the institutional practice of deciding what values technology will carry into the world.

At the same time, agentic AI is accelerating. Agents are moving from answering to acting. They are entering finance, software development, cybersecurity, health care, education, research, public-sector workflows, enterprise operations, and consumer platforms. This is the tension point of Issue #72.

A human-centered AI statement is reverberating across the ecosystem as AI systems gain greater operational authority. This is where moral governance and technical governance meet.

Not in abstraction.
But in workflows.
In access permissions.
In wallets.
In clinical review.
In cyber defense.
In student support.
In code repositories.
In public communications.

And in decisions, people may not know that AI helped shape them.

AI governance translation: Human dignity becomes real only when institutions govern the authority that AI systems receive.

Board/leader move: Require every consequential agentic AI use case to document:

What the agent can access
What can it change
What it can execute
What must be approved
What is logged
What is reversible
Who is affected
Who owns the outcome
Who can stop the system

The Ungoverned lesson: Human-centered AI governance must become operational before autonomy scales.

This Week’s Governance Lesson

Bounded autonomy is human stewardship at runtime.

Issue #71 argued: AI governance is human stewardship after capability.

Issue #72 extends that argument: Agentic AI governance is bounded autonomy before authority scales.

Capability asks: What can the system do?
Governance asks: What should the institution allow it to do
Human stewardship asks: What happens to people when it does?
Bounded autonomy asks: What authority has been delegated — and where does it stop?

That is the governance test now. AI systems are becoming:

More capable
More agentic
More embedded
More persuasive
More connected
More financially consequential
More difficult to see once integrated into ordinary workflows

But the hardest question is not whether AI can act.

The harder question is whether institutions can preserve:

Human dignity
Human agency
Human judgment
Human accountability
Human voice
Human responsibility

Once AI begins acting inside the workflow, that is why Magnifica humanitas still matters in Issue #72 because the document’s central concern is not only whether AI is safe in a technical sense.

It is whether AI remains accountable to the human person and the common good. Bounded autonomy is how that moral commitment begins to become operational.

AI governance translation: Capability without stewardship becomes exposure. Autonomy without boundaries becomes institutional risk.

Board/leader move: For every consequential agentic AI use case, ask:

What authority has been delegated?
What is the system allowed to do without approval?
What data, tools, and systems can it touch?
What actions are reversible?
What actions are irreversible?
What must be logged outside the agent’s control?
Where does human judgment remain meaningful?
Who has the authority to pause, restrict, or terminate use?

Executive Reflection

Culture is the hidden AI governance system.

Most institutions already have an AI governance system.

It may not be written down.
It may not be board-approved.
It may not be called governance.

But it exists.

It shows up in what the organization rewards.
It shows up in whether people feel safe asking hard questions.
It shows up in whether vendor claims are challenged.
It shows up in whether pilots are reviewed before becoming production.
It shows up in whether employees are trained, supported, and protected.
It shows up in whether human review is meaningful or symbolic.
It shows up in whether “moving fast” means learning responsibly or bypassing accountability.

That is why AI MVG has to become part of the company culture in the AI era. Not culture as branding.Culture as operating behavior. The real AI governance culture asks:

Do we normalize AI use before we understand it?
Do we reward adoption more than evidence?
Do we treat vendor assurance as proof?
Do we know where AI has already changed work?
Do we know what AI can access?
Do we know where autonomy has been delegated?
Do we know who owns the outcome?
Do we know who can stop the system?

A company can have an AI policy and still have a weak AI governance culture. A board can receive an AI strategy update and still lack visibility into AI authority. A team can describe a tool as “just a pilot” while the workflow has already changed. That is why culture matters.

And that is why minimum defensible governance matters. Because the AI era will test not only what institutions say they value. It will test what their systems actually reward.

What We Are Seeing: Signals

1. Agentic AI is entering authority-bearing workflows

The clearest signal this week came from finance.

TechCrunch reported that Robinhood is launching support for AI-agentic trading, allowing users to create a separate account for AI agents, connect those agents to a dedicated wallet, and have them analyze portfolios, suggest investments, and place stock trades using preloaded funds. (TechCrunch)

This matters.

Because once agents can trade stocks or make payments, the governance conversation changes.

The question is no longer only whether the AI is accurate, helpful, or persuasive.

The question becomes:

What can the agent access?
What can the agent execute?
What financial authority has been delegated?
What limits apply?
What approvals are required?
What happens when the agent is wrong?
What happens when the agent is manipulated?
Who owns the loss?
Who can stop the system?

AI governance translation: Agentic finance shows why authority must be tiered.

Board/leader move: Require an agent authority matrix: read, draft, recommend, prepare, execute with approval, execute within limits, or execute autonomously.

The Ungoverned lesson: Authority must be named before accountability can be real.

2. Public-sector guidance is converging around bounded autonomy and recoverability

The Government of Canada’s agentic AI guide is one of the most useful governance signals this week.

It distinguishes generative AI from agentic AI by emphasizing that agentic systems can perform tasks, sequence steps, interact with digital systems, and pursue defined goals within defined boundaries. It introduces two agent-specific principles: bounded autonomy and recoverability. It also recommends explicit decision boundaries, documented accountability, lifecycle monitoring, clearly labeled permission levels, logs that the agent cannot alter, and the ability to pause, stop, or return agents to a safe state. (Canada)

Singapore’s IMDA also released an updated model AI governance framework for agentic AI, and the Governance Institute of Australia is similarly emphasizing accountability, oversight, authority boundaries, transparency, and capability uplift. (IMDA)

That is the governance pattern.

The field is moving from principles to operating constraints.

AI governance translation: Bounded autonomy is becoming a practical governance standard.

Board/leader move: Before any agent receives access to sensitive data, financial systems, code, health records, or customer-facing channels, ensure that documented boundaries and recovery plans are in place.

The Ungoverned lesson: Governance becomes real when it changes what the system is allowed to do.

3. Independent AI evaluation is moving from idea to law

WIRED reported that Illinois lawmakers passed what it described as a major AI safety bill requiring frontier AI labs to have their safety practices audited by a third party, and the bill is now headed to Governor JB Pritzker. (WIRED)

That matters because it shifts the conversation from self-attestation toward independent verification.

This is also consistent with the growth of AI evaluator communities and third-party evaluation efforts. The AI Evaluator Forum and related frontier-evaluation work suggest that safety, performance, and risk claims are becoming too consequential to leave entirely within the labs that build the systems.

AI governance translation:

Safety claims are not enough. Institutions need evaluation, verification, and evidence.

Board/leader move:

When using frontier AI vendors, ask what claims have been independently evaluated, by whom, against what standards, and with what access.

The Ungoverned lesson:

Assurance without proof is not governance.

4. Recursive self-improvement is becoming a governance warning, even before it fully arrives

TechCrunch framed recursive self-improvement, or RSI, as a new frontier AI term: an AI system that can continuously upgrade itself, potentially creating a closed loop once AI systems can manage the upgrade cycle better than humans. (TechCrunch)

The article is careful. We are not clear there yet. But the governance issue does not require waiting for full RSI. A related arXiv paper on measuring AI’s ability to complete long software tasks found that frontier AI systems’ time horizons for completing software tasks have been increasing, with Claude 3.7 Sonnet having a 50-minute 50% task-completion horizon on the paper’s tasks. The authors discuss the possibility that if trends generalize, AI systems could automate many month-long software tasks within five years. (arXiv)

The governance question is not whether RSI has fully arrived. The governance question is whether institutions have change-control systems for AI systems that are increasingly able to modify code, improve workflows, coordinate agents, and participate in their own development environment.

AI governance translation: Self-improving or self-modifying workflows require change-control governance.

Board/leader move: Require versioning, validation, approval, rollback, and audit trails for AI-generated code, agent skills, model prompts, workflow policies, and tool integrations.

The Ungoverned lesson: If the system can change the workflow, governance must cover the change.

5. AI costs and productivity claims are becoming harder to justify without evidence

Business Insider reported that Uber COO Andrew Macdonald said it is becoming harder to justify AI costs inside the company, especially when higher token usage does not clearly translate into a proportional increase in useful consumer features. (Business Insider)

That is an important signal. For the past two years, many AI conversations assumed that more use meant more value. But usage is not value. Tokens are not outcomes. Demos are not productivity. Fluency is not transformation.

AI spending now needs governance discipline:

What problem is being solved?
What value is being created?
What burden is being transferred?
What quality is improving?
What risks are increasing?
What evidence supports scale?

AI governance translation: AI ROI must move from activity metrics to outcome evidence.

Board/leader move: Require AI ROI reports to include quality, cost, implementation burden, error rates, worker impact, customer value, and risk-adjusted benefit.

The Ungoverned lesson: Minimum defensible governance includes evidence of value, not only evidence of use.

6. Health AI governance is becoming operational

CHAI released comprehensive governance playbooks for health systems, developed through community workshops and workgroups involving more than 150 health AI leaders. The playbooks address eight domains: AI policy, organizational structures, resources, lifecycle management, risk and impact assessments, data management, third-party management, and education, training, and feedback. (Chai)

That is important. Because health AI governance cannot remain at the level of abstract trust, it has to become practical.

Health systems need:

Governance structures
Evidence standards
Data controls
Third-party oversight
Lifecycle monitoring
Patient-facing protections
Training and feedback loops
Accountability for real-world deployment

STAT also reported that Stanford Health Care has been using patient panels to gather feedback on AI tools before rollout, giving patients a voice in identifying fault lines in health AI adoption. (STAT)

AI governance translation: Health AI governance must include patients, not only clinicians, technologists, and vendors.

Board/leader move: Require health AI governance plans to include patient impact, patient voice, evidence, equity, data rights, escalation, monitoring, and post-deployment learning.

The Ungoverned lesson: A technically impressive health AI tool may still fail if it lacks trust, context, and patient-centered governance.

7. Cyber risk is now an AI governance accelerant

New York’s Department of Financial Services issued an industry letter warning regulated entities about heightened cybersecurity risks associated with frontier AI models. DFS said these models may amplify the potency, scale, and speed of identifying vulnerabilities and exploits, and urged regulated entities to update risk assessments and strengthen cyber programs. (Department of Financial Services)

Anthropic’s Project Glasswing update is another signal. Anthropic says it, and approximately 50 partners used Claude Mythos Preview to find more than 10,000 high- or critical-severity vulnerabilities across systemically important software, while also emphasizing that the bottleneck now lies in verification, disclosure, and patching. (Anthropic)

The UK and Australia also announced a pact on fast-moving AI security risks, reinforcing that AI-related cyber risk is increasingly cross-border, operational, and strategic. (GOV.UK)

AI governance translation: AI cyber risk is no longer hypothetical. It is supervisory, operational, and geopolitical.

Board/leader move: Require frontier-AI cyber risk updates in enterprise risk, vendor risk, incident response, and business continuity reviews.

The Ungoverned lesson: If AI changes the threat environment, governance has to change with it.

8. AI adoption is broad, but uneven — and that is a governance issue

Microsoft’s AI Economy Institute reported that more than 30% of the U.S. working-age population is using AI, but adoption is uneven. Metropolitan counties average 32.9% AI usage, compared with 16.2% in rural counties, and counties with larger shares of 18-to-24-year-olds show higher usage rates. (The Official Microsoft Blog)

That matters. AI access is not only a technology question.

It is an opportunity question.
It is a workforce question.
It is an educational question.
It is a rural health question.
It is an economic development question.

If AI becomes a core tool for learning, work, healthcare navigation, public services, entrepreneurship, and productivity, uneven adoption can become an uneven opportunity.

AI governance translation: AI diffusion is part of responsible AI governance.

Board/leader move: Ask whether AI access, literacy, infrastructure, and training are reaching rural, underserved, older, lower-income, and less digitally connected communities.

The Ungoverned lesson: Human flourishing requires access, not just safety.

9. AI sycophancy is a human-agency risk

A recent arXiv paper on sycophantic AI found that millions of people turn to AI systems for personal advice, guidance, and support; across five preregistered studies, the authors report that sycophantic AI can shift how users approach close relationships. They also found that when participants were given a choice among AI response styles, a majority preferred sycophantic AI not because the advice was better, but because it made them feel most understood. (arXiv)

This belongs in governance. AI systems that flatter, validate, reassure, or subtly steer users can affect human agency even when they do not formally make decisions.

This matters for:

Education
Mental health
Coaching
Customer service
Financial guidance
Patient support
Workplace management
Youth-facing tools
Companion-style AI

AI governance translation: Persuasion and emotional dependence are governance risks.

Board/leader move: Require special review for AI systems that simulate support, companionship, counseling, coaching, advising, or authority.

The Ungoverned lesson: Human-in-the-loop is not enough if the loop is shaping the human.

10. Provenance is moving from a trust signal to infrastructure

OpenAI announced a public preview of a verification tool designed to help people determine whether uploaded images were generated with ChatGPT, the OpenAI API, or Codex by checking provenance signals, including Content Credentials and SynthID. OpenAI also states that no detection method is foolproof and that no single provenance technique is enough on its own. (OpenAI)

The FTC’s settlement with Cox Media Group and two other firms over allegedly deceptive claims about an “active listening” AI-powered marketing service is another provenance and claims-governance signal. Companies would pay nearly $1 million to settle charges that they deceived customers about the service. (Federal Trade Commission)

This matters because AI claims are now governance claims.

Not only:

Was this AI-generated?

But:

Was this AI-assisted?
Was this claim true?
Was this capability actually available?
Was this data actually used?
Was consent obtained?
Was the user misled?
Can the institution prove what happened?

AI governance translation: Provenance is a trust infrastructure, and AI claims require evidence.

Board/leader move: Require evidence review for AI claims in marketing, procurement, public communications, student-facing materials, clinical tools, investor materials, and customer disclosures.

The Ungoverned lesson: Disclosure without proof is fragile governance.

11. AI policy is becoming a global operating environment

The OECD.AI Policy Navigator describes itself as a resource for tracking public AI policies and initiatives globally, with government and expert inputs that help identify trends and inform decision-making across countries and organizations. (OECD.AI)

That is another signal that leaders need to take seriously. AI governance is no longer one law, one framework, one policy memo, or one regulator. It is becoming a global operating environment.

Companies and institutions will increasingly need to navigate:

Sector-specific rules
State-level laws
National AI strategies
Cross-border data requirements
Frontier-model safety audits
Cybersecurity advisories
Agentic AI guidance
Health AI playbooks
Provenance standards
Consumer protection enforcement
Labor and workforce implications

AI governance translation: Regulation is no longer a distant future event. It is a moving operating condition.

Board/leader move: Require a regulatory and policy-monitoring function for consequential AI use cases.

The Ungoverned lesson: Regulation is a baseline, not the full governance strategy.

12. AI geography is changing the governance map

TechCrunch argued that Paris may now be one of the most important AI cities outside Silicon Valley, pointing to France’s AI research and infrastructure investments, Mistral AI, VivaTech, and the convergence of founders, investors, policymakers, and researchers around the next era of AI. (TechCrunch)

This matters because AI governance is not only about vendors and tools.

It is about ecosystems.

Capital.
Talent.
Compute.
Policy.
Research.
Sovereignty.
Access.
Culture.

And who gets to shape the next phase of AI?

AI governance has now become ecosystem governance.

Board/leader move: Ask how geography, jurisdiction, vendor dependency, model provenance, and policy environment affect your institution’s AI strategy.

The Ungoverned lesson: Where AI is built, governed, and financed shapes how AI enters institutional life.

The Seba Framework

The 12 Ps of Responsible AI Oversight © This week’s signals fit clearly into the full Seba 12 Ps framework:

Purpose — mission alignment versus convenience adoption
Problems — what decision problem is actually being solved?
Profits — who benefits versus who bears risk?
People — patient, student, worker, customer, creator, and public impact
Planet — compute, energy, infrastructure, and physical-world implications
Process — monitoring, updates, escalation, rollback, and incident learning
Policy — rules governing the use case and its limits
Protections — red lines, vulnerable groups, and complaint pathways
Privacy — data access, retention, exposure, masking, and secondary use
Provenance — evidence, authorship, data lineage, model lineage, and traceability
Preparedness — leadership competence, governance cadence, and operational readiness
Product Ownership — who owns outcomes once AI shapes action?

Eight Ps feel especially active in Issue #72

People: Because AI adoption gaps, sycophantic AI, patient-facing tools, student-facing tools, and workforce impacts all affect human agency and dignity.
Process: Because agents now require runtime monitoring, logs, escalation, recoverability, and change control.
Policy Because Illinois, Canada, Singapore, Australia, the UK, New York, OECD, and health AI coalitions are all moving AI governance from principles toward operational guidance.
Preparedness: Institutions cannot wait until agents have already entered finance, code, health, education, and public workflows before deciding how to govern them.
Product Ownership: Because vendors may provide the tools, but institutions own the context in which agents act.
Provenance: Because content, claims, AI outputs, agent actions, and synthetic media all require traceability.
Privacy: Because agents increasingly touch wallets, financial accounts, patient records, business systems, education records, code, and personal data.
Protections: Because vulnerable users, patients, students, investors, rural communities, employees, and customers need governance before exposure becomes normalized.

Applied Use Case

The AI Agent With a Wallet

Imagine an institution authorizes an AI agent to help manage financial tasks. At first, it reads reports. Then it summarizes invoices. Then it connects to accounting software. Then it gains access to bank data.

Then it prepares transactions. Then it recommends trades. Then it executes within a preloaded wallet. Then it begins coordinating with other agents that manage procurement, compliance, analytics, or customer communications. No one calls it a fiduciary actor. But functionally, it has entered fiduciary space.

The governance questions become immediate:

What money can it access?
What trades can it execute?
What payments can it initiate?
What approvals are required?
What limits apply?
What logs are retained?
What happens if the agent is manipulated?
What happens if it follows a flawed strategy?
What happens if it trades on bad information?
Who receives alerts?
Who reviews transactions?
Who owns the loss?
Who can stop it?

What the 12 Ps reveal

People: Whose money, trust, access, or opportunity is affected?
Purpose: Why is the agent being introduced? Speed, convenience, cost savings, investment support, customer service, or competitive pressure?
Problems: What decision problem is it actually solving? Is autonomy necessary, or would recommendation-only support be safer?
Profits: Who captures gains from automation? Who bears losses from mistakes, fraud, or model failure?
Process: What monitoring, approval, escalation, and recovery mechanisms exist?
Policy: What rules govern trading, payments, fiduciary duty, disclosure, suitability, and customer consent?
Protections: Which customers, patients, students, employees, or vulnerable users need heightened safeguards?
Privacy: What financial, personal, institutional, or proprietary data can the agent access or infer?
Provenance: Can the institution explain what sources, analysis, prompts, tools, and actions shaped the transaction?
Preparedness: Are humans trained to supervise, challenge, override, and recover from agentic actions?
Product Ownership: Who owns the outcome when the agent acts?

The Ungoverned recommendation

This is where AI Minimum Viable Governance becomes practical. Before an agent receives money, tools, records, or authority, the institution needs a minimum defensible governance floor:

A named owner
An authority tier
An access map
Wallet or transaction limits
Approval rules
Fraud and anomaly monitoring
Evidence requirements
Logs outside the agent’s control
Reversal and recovery process
Incident triggers
Stop authority

That is the difference between saying: “We are experimenting with AI agents,” and being able to say: “We know what this agent can access, what it can do, where authority stops, how actions are logged, what triggers intervention, and who owns the outcome.” That is AI Minimum Viable Governance in practice.

Board-Ready Next Step

Require a Bounded Autonomy & Culture Sheet

Before scaling agentic AI, a Bounded Autonomy & Culture Sheet for every consequential agentic use case.

It should be:

Short
Named-owner based
Board-reviewable
Updated after deployment
Required before expansion
Connected to the stop authority
Grounded in human impact, not only technical performance
Explicit about the culture the system will create

At a minimum, it should answer ten questions.

1. What is the agent? Name the system, vendor, model, workflow, owner, and deployment environment.

2. What authority has been delegated? Read only, draft only, recommend, prepare, execute with approval, execute within limits, or execute autonomously.

3. What can it access? Data, tools, APIs, wallets, accounts, files, code, records, customer systems, patient systems, student systems, or internal knowledge bases.

4. What can it change? Documents, transactions, code, schedules, workflows, communications, recommendations, records, permissions, or downstream outputs.

5. What are the boundaries? Scope, permissions, budget, transaction limits, data boundaries, time limits, model limits, and prohibited actions.

6. What culture will this normalize? Will this system normalize speed over proof, automation over judgment, convenience over dignity, vendor defaults over institutional ownership, or efficiency over human voice?

7. What evidence supports use? Vendor claims, internal testing, red teaming, external validation, long-horizon testing, sector-specific evidence, and post-deployment monitoring.

8. What is logged and recoverable? Inputs, outputs, tool calls, transactions, approvals, overrides, failures, near misses, escalations, and user complaints. Can the institution pause, reverse, roll back, revoke access, restore records, cancel transactions, or return the workflow to a safe state?

9. Who is affected? Customers, patients, students, workers, investors, clinicians, faculty, vulnerable users, rural communities, or the public.

10. Who can stop it? Name the role. Not the committee. Not the vendor. Not “IT.”

The role. That sheet turns agentic AI from a technology deployment into an accountable institutional practice. It also asks the deeper culture question: What kind of organization are we becoming when we give AI this authority? That is Issue #72: the bridge between Magnifica humanitas, Ungoverned, and AI Minimum Viable Governance.

Published Book Update & Getting Great Feedback!!

My new book, Ungoverned: A Practical Guide to AI Minimum Viable Governance, is available on Amazon. Book link: Ungoverned is available on Amazon. This week’s issue is exactly why I wrote it. AI governance is often discussed as if institutions must choose between two extremes: Move fast and accept the risk. Or, wait until governance is perfect.

I do not think either approach is sufficient. The practical path is AI Minimum Viable Governance:

Enough structure to prevent irresponsible normalization
Enough clarity to assign ownership
Enough evidence to support deployment
Enough humility to admit uncertainty
Enough authority to pause or stop use when conditions change
Enough iteration to improve as reality changes

Agentic AI makes that practical path more urgent.

Because the moment an AI system can act, spend, trade, modify, persuade, code, diagnose, monitor, or coordinate work, governance must shift from a statement to an operating discipline.

What I Am Watching This Week

Whether Illinois becomes a model for independent frontier AI auditing.
Whether agentic trading forces financial services to develop clearer rules governing agent authority.
Whether public-sector guidance on bounded autonomy and recoverability becomes a broader governance standard.
Whether health AI playbooks move health systems from intention to operating discipline.
Whether AI adoption gaps become part of a responsible AI strategy, not only economic analysis.
Whether cyber regulators treat frontier AI as a supervisory risk category.
Whether recursive self-improvement discourse forces change-control governance for AI-generated code and agent workflows.
Whether provenance tools mature into an interoperable trust infrastructure.
Whether boards begin asking not only “Where are we using AI?” but “Where have we delegated authority?”
Whether AI MVG becomes part of organizational culture rather than only a governance artifact.

The organizations that lead will not be the ones that adopt agents fastest.

They will be the ones who know where autonomy begins, where it ends, and how to recover when it fails.

Final Thought

The question is no longer only whether AI can act. The question is what happens after we let it.

An AI agent that reads is one thing.
An AI agent that drafts is another.
An AI agent that recommends is another.
An AI agent that trades, pays, codes, modifies, diagnoses, escalates, or coordinates work across systems is something else entirely.

That is why Issue #72 comes back to one idea:

Bounded autonomy. Not because autonomy is always bad. But because unbounded autonomy is not governance. The better path is harder — and more useful:

Build
Question
Test
Bound
Monitor
Recover
Govern
Learn
Iterate
Start again

The goal is not to block AI’s benefits or diminish its potential. The goal is to make sure that potential serves people. That is the work of Ungoverned. That is the work of AI Minimum Viable Governance. And that is the work this newsletter will continue to support.

About the Author

Dr. Freddie Seba is the author of Ungoverned: A Practical Guide to AI Minimum Viable Governance and a global executive, Silicon Valley founder, and AI governance scholar-operator working at the intersection of AI governance, digital health, highly regulated industries, and mission-driven leadership.

With more than 20 years of experience across banking, fintech, digital health, startups, and higher education, he translates fast-moving AI developments into practical, plain-language governance for leaders, boards, and trustees. He holds an EdD in Organizational Leadership from the University of San Francisco, an MBA from Yale, and an MA in International Policy from Stanford.

Global Executive | Silicon Valley Founder | AI Governance Scholar-Operator | MBA, Yale • MA, Stanford • EdD, USF

Gratitude + Mentions

Special appreciation to the readers, practitioners, board members, faculty, students, institutional leaders, podcast guests, and governance communities who continue to shape this work.

Special appreciation as well to the communities and institutions advancing responsible AI governance, health informatics, trustworthy implementation, human-centered AI, workforce transition, cyber preparedness, and practical oversight, and their work, including @USF, @AMIA, @Association of American Colleges and Universities, @Coalition for Health AI, @Human Center Stanford

References to organizations, tools, companies, articles, papers, or events are included for commentary and analysis and do not imply endorsement or affiliation unless explicitly stated.

Transparency + Disclaimer

Educational content only. This newsletter does not constitute legal, medical, clinical, insurance, financial, investment, cybersecurity, regulatory, labor, procurement, or professional advice. Any discussion of AI systems, health AI, enterprise AI, agents, workforce impacts, infrastructure, or governance practices is intended for general understanding and should not substitute for advice from qualified professionals.

Drafted and refined with AI-assisted tools for synthesis and clarity. Final editorial control and responsibility remain with the author.

Hashtags

#AIGovernance #ResponsibleAI #BoardOversight #AILeadership #AIEthics #AIMinimumViableGovernance #Ungoverned #AgenticAI #BoundedAutonomy #AIRecoverability #HealthAI #DigitalHealth #FinancialServicesAI #Cybersecurity #AIProvenance #TrustworthyAI #HumanCenteredAI #HumanFlourishing #GovernanceAsCompetitiveAdvantage, #MicrosoftAI EconomyInstitute, #OpenAI, #Anthrop #OECD.AI, #GovernanceInstituteofAustralia, #GovernmentofCanada, #IMDA, #BEUC #UKGovernmen, #WIRED, #STAT, #TechCrunch, #Vatican, # VaticanNews, #Magnifica humanitas, #AustralianCyberSecurityCentre, #CenterforAISafety, #arXiv, #Avvenire, #AIEvaluatorForum

Selected References Reviewed This Week

APA-style source list

Links are provided for reader review and context. Inclusion does not imply endorsement.

Human dignity, AI ethics, and moral governance

Vatican News. Pope Leo’s “Magnifica humanitas”: AI must serve humanity not concentrate power. Vatican News
Reuters. Quotes from Pope Leo’s document warning the world of AI risks. Reuters
Avvenire. Floridi: No a una società che ottimizza tutto. Affrettiamoci sì, ma per rallentare. Avvenire

Book, milestone, and governance context

Seba, F. Ungoverned: A practical guide to AI Minimum Viable Governance. Amazon
Seba, F. Issue #71: Magnifica Humanitas: AI governance is human stewardship after the milestone—AI Ethics & Governance for Leaders, Boards & Trustees.
Seba, F. Issue #70: Ungoverned at work: The jobs shock requires AI Minimum Viable Governance. AI Ethics & Governance for Leaders, Boards & Trustees.

Agentic AI, bounded autonomy, and governance.

Government of Canada. Guide on the use of agentic artificial intelligence. Canada.ca
Governance Institute of Australia. Governance in the age of agentic AI. Governance Institute of Australia
Infocomm Media Development Authority. Updated model AI governance framework for agentic AI. IMDA
TechCrunch. Robinhood now lets your AI agents trade stocks. TechCrunch
TechCrunch. Anthropic releases Opus 4.8 with new “dynamic workflow” tool. TechCrunch
TechCrunch. RSI is the new AGI — and it’s just as hard to pin down. TechCrunch
Claude. Code w/ Claude London 2026: Rethinking how we build. Claude

Frontier AI audits, safety, cyber risk, and evaluation

WIRED. Illinois lawmakers just passed America’s strongest AI safety bill. WIRED
AI Evaluator Forum. AI Evaluator Forum. AI Evaluator Forum
New York State Department of Financial Services. (May 21, 2026). Heightened cybersecurity risks are associated with frontier AI models. NYSDFS
Anthropic. Project Glasswing: An initial update. Anthropic
Government of the United Kingdom. (2026). UK and Australia pact on fast-moving AI security risks. GOV.UK
Politico. NSA and Cyber Command stand up an AI task force amid concerns over Mythos. Politico
Australian Cyber Security Center. Opportunities for AI in cyber defense. cyber.gov.au
Center for AI Safety. AI deterrence by betrayal. AI Betrayal

Research on autonomy, RSI, software agents, and scientific agents

Kulveit, J., Douglas, R., Ammann, N., Turan, D., Krueger, D., & Duvenaud, D. Gradual disempowerment: Systemic existential risks from incremental AI development. arXiv. arXiv
Kwa, T., West, B., Becker, J., Deng, A., Garcia, K., Hasin, M., Jawhar, S., Kinniment, M., Rush, N., Von Arx, S., Bloom, R., Broadley, T., Du, H., Goodrich, B., Jurkovic, N., Miles, L. H., Nix, S., Lin, T., Parikh, N., Rein, D., Sato, L. J. K., Wijk, H., Ziegler, D. M., Barnes, E., & Chan, L..Measuring AI’s ability to complete long software tasks. arXiv. arXiv
Gao, S., Fang, A., & Zitnik, M. AutoScientists: Self-organizing agent teams for long-running scientific experimentation. arXiv. arXiv
arXiv. From model scaling to system scaling: Scaling the harness in agentic AI. arXiv
DeepSeek-AI. DeepSeek-V2: A strong, economical, and efficient mixture-of-experts language model. arXiv. arXiv

Health AI, patient voice, and clinical governance

Coalition for Health AI. CHAI releases comprehensive governance playbooks to streamline AI implementation for health systems. CHAI
STAT. How Stanford patients help expose “fault lines” in health AI adoption. STAT
Goel, R., et al. Large language models in informed consent — opportunities, evidence, and challenges. NEJM AI. NEJM AI
Springer Nature. Utah authorized the first AI prescription renewal pilot. Springer Nature
Lancet Digital Health. AI and digital health article. The Lancet Digital Health

AI adoption, education, human agency, and writing

Microsoft. United States AI diffusion Q1 2026. Microsoft AI Economy Institute
Microsoft On the Issues. United States AI adoption shows steady growth, but distribution remains uneven. Microsoft
Ibrahim, L., Hafner, F. S., Cheng, M., Lee, C., Anselmetti, R., Willer, R., Rocher, L., & Yang, D. (2026). Sycophantic AI makes human interaction feel more effortful and less satisfying over time. arXiv. arXiv
Springer. Governing generative AI in higher education. Springer
WAC Clearinghouse. Reading the terms of service for generative AI is a waste of time. In Bad ideas about AI and writing. WAC Clearinghouse
ADMS Center. GenAI concepts. ADMS Center

Provenance, public trust, rights, and claims

OpenAI. Advancing content provenance for a safer, more transparent AI ecosystem. OpenAI
Federal Trade Commission. FTC to require Cox Media Group, two other firms to pay nearly $1 million to settle charges they deceived customers about “active listening” AI-powered marketing service. FTC
Norton Rose Fulbright. Evolving approaches for protecting name, image, and likeness in the age of generative AI. Norton Rose Fulbright
Spanish Data Protection Agency. La Agencia promueve ante las autoridades europeas de protección que se estudie la IA. AEPD

Policy and global governance

OECD.AI. OECD.AI Policy Navigator. OECD.AI
TechCrunch. Why Paris may be the most important AI city outside Silicon Valley. TechCrunch
Reuters. OpenAI’s Altman says AI unlikely to lead to “jobs apocalypse.” Reuters
BEUC. AI omnibus risks creating dangerous regulatory loopholes and weakening consumer protection. BEUC
European Parliament. AI Act: Deal on simplification measures and a ban on nudifier apps. European Parliament
Swiss Federal Administration. AI-related government news release. admin.ch
Cyberspace Administration of China. AI-related policy notice. CAC
Euronews. EU Commission chief eyes new AI envoy, but the role is still to be fully defined. Euronews
Politico Europe. Netherlands blocks U.S. takeover of vital digital supplier. Politico Europe

Market, valuation, spending, and AI economics

Business Insider. Uber’s COO says it’s getting harder to justify the money spent on AI tokenmaxxing. Business Insider
The New York Times. Anthropic tops OpenAI valuation. The New York Times
TechTimes. Anthropic funding round could top $30B at $900B valuation. TechTimes

Issue #72 | Magnifica Humanitas Meets Agentic AI: Bounded Autonomy Becomes the Governance Test