Ungoverned: AI Ethics & Governance for Leaders, Boards & Trustees.
By Dr. Freddie Seba © 2026 Freddie Seba. All rights reserved.
Human-agent teams are becoming the way institutions work — and the gap between who acts and who is accountable is now the governance test boards cannot postpone.
This Issue in 60 Seconds
The thesis: AI is no longer a tool your institution uses. It is a teammate your institution works with — drafting, monitoring, transacting, diagnosing, deciding alongside humans. But teams have something AI systems do not: someone who answers for the outcome. This week’s signals — from patient surveys to central banks to Cambridge geopolitics scholars — all point at the same gap. AI is joining the team faster than institutions are deciding who is responsible for what the team does.
Five signals you need to know: Three in four workers are using AI tools their employer never approved. Patients overwhelmingly assume their health data already trains AI — and their trust is conditional. Frontier model access changed twice in one week, in both directions. Europe delayed its high-risk AI rules by up to two years. And the Bank of England warned that AI agents could amplify market stress.
One action: Before your next board meeting, run the Seba Human-Agent Accountability Map. The five questions are at the end of this issue.
One sentence to share with your board: Your institution already has human-agent teams — the only question is whether anyone is accountable for them.
The Argument
There is a phrase spreading through enterprise strategy this year that deserves more governance scrutiny than it is getting: the human-agent team.
It sounds reassuring. Humans and AI working together toward shared goals. The human brings judgment. The agent brings speed, scale, and memory. Deloitte finds that close to 75% of businesses plan to deploy AI agents by the end of 2026. Microsoft reports 500,000 agents already operating across business functions. The World Economic Forum is publishing guidance on defining roles for humans and AI on the same team.
But here is what the vocabulary of teamwork quietly obscures: teams have something AI does not—someone who answers for the outcome.
In a human team, accountability is legible. There is a manager. There is an org chart. There is a named person who owns the result. In a human-agent team, the agent executed, the human approved a workflow configured months ago, the vendor supplied the model, the platform supplied the integration — and when something goes wrong, everyone can point somewhere else.
That is not a team. That is an ungoverned system wearing the language of collaboration.
And this week, the accountability gap showed up everywhere: in what workers are doing without their employers’ knowledge, in what patients believe is happening to their data, in what central bankers fear agents could do to markets, in how governments grant and restrict model access, and in how regulators are struggling to schedule the rules. Ten signals. One gap.
From the Book
In Ungoverned, I argue that governance failures rarely arrive as scandals. They arrive as workflows.
The human-agent team is the most elegant workflow the AI era has produced. It preserves the appearance of human control while steadily expanding the scope of machine authority. The human is still in the loop — technically. But the loop has grown so large, and the agent’s actions so numerous, that the human’s presence becomes more ceremonial than operational.
AI Minimum Viable Governance asks one direct question of every human-agent system: is the human in the loop because they are genuinely reviewing and deciding — or because it makes the institution feel like someone is responsible?
Those are not the same thing. Governance requires the first. Many institutions are settling for the second.
This Week: 10 Signals, One Pattern
1. The human-agent team is becoming the dominant operating model — and the accountability chart hasn’t been drawn.
Close to 75% of businesses plan to deploy AI agents by the end of 2026, according to Deloitte’s State of AI in the Enterprise. Microsoft tracks half a million agents already at work. Some organizations are experimenting with supervisor agents that assign tasks to other agents — one gathers data, another models it, a third compiles the report — completing work end-to-end that previously required a human moving between tasks. The World Economic Forum is now publishing frameworks for defining roles across human-AI teams.
The governance gap is structural. Org charts were built for human actors with roles, reporting lines, and performance reviews. An agent has none of those. It does not get promoted, coached, or fired. It executes — at speeds and volumes no human oversight framework was designed to match. Gartner’s guidance this month points in the right direction: connect risk, data, and cybersecurity governance bodies into one unified AI governance team (organizations that do see measurably greater business impact), consolidate policies into one framework reflecting the institution’s actual risk tolerance, and adopt policy-as-code so rules are enforced automatically in the technology stack rather than aspirationally in a document.
The Ungoverned lesson: Human-agent teams require accountability charts, not just org charts — and policy that executes, not policy that sits in a binder.
2. Your institution already has ungoverned human-agent teams. They’re called shadow AI.
According to the BYO AI Report, 41% of workers say their employer has provided nothing — no tools, no training, no guidance — to prepare them to use AI at work. Only 19% report comprehensive employer-provided AI training. The result is predictable and measurable: more than three in four workers have used AI tools they personally found and signed up for, rather than tools their employer provided or approved, to complete work tasks.
Read that alongside the 2026 CISO AI Risk Report finding that only 16% of organizations effectively govern AI access to core business systems. The picture sharpens: the human-agent team is not a future deployment decision your board will make. It is a present reality your workforce has already built — on unapproved tools, with ungoverned data flows, invisible to your policies. Institutions that believe they have not yet deployed AI agents have, in most cases, not looked into it.
The Ungoverned lesson: Shadow AI means the accountability gap already exists at scale — governance must start by making the invisible inventory visible.
3. Patients already assume AI is training on their health data — and their trust is conditional.
The Coalition for Health AI’s new Patient Survey Report, conducted with NORC at the University of Chicago and supported by the California Health Care Foundation, should be required reading for every health system board. Seventy percent of respondents believe their personal health data is currently being used to train AI. Only 13% say they are very comfortable with AI in general. And while 64% are open to their health data being used to improve healthcare, that openness is layered with conditions: 25% only with explicit opt-in permission, 29% only if the data is anonymous and cannot be linked to them — and 35% reject it outright, regardless of safeguards. Notably, concern about health data being sold or used for profit (63%) runs higher than concern about biased treatment (55%).
Now place that beside vendor privacy policies taking effect this year — including one major AI provider’s consumer health data policy governing how health conditions, symptoms, diagnoses, biometric data, and location flow into AI systems when users connect third-party health apps. The infrastructure for the exact data flows patients are concerned about is being formalized in vendor terms — while most institutions have not reviewed which integrations their employees and patients have connected to, or whether their consent frameworks cover any of it. CHAI’s conclusion is the right one: legitimacy will depend on layered protections that combine meaningful choice, strong de-identification, and clear limits on use — not on any single consent mechanism.
The Ungoverned lesson: Patient trust in health AI is conditional and measurable — and the gap between patient expectations and institutional data practice is now documented.
4. Model access changed twice in one week — in both directions. Continuity risk is now a pattern, not an event.
Anthropic announced the redeployment of Fable 5, restoring access that had been suspended under a government directive — the story this newsletter tracked in Issues #74 through #76. The same week, OpenAI limited its GPT-5.6 rollout following a government request, stating publicly that such restrictions “shouldn’t be the norm.”
Set aside the politics entirely — this newsletter does — and note that this lesson applies equally to every frontier vendor: access can now change in either direction, on government timelines, across multiple providers, repeatedly. When access was suspended, this newsletter argued that model access belongs on the operational resilience agenda. The redeployment does not soften that conclusion. It confirms it. Institutions that treated the first suspension as a one-off vendor event now have a second data point from a second vendor in the same month. Access volatility is a recurring structural feature of depending on frontier AI — and continuity planning, substitute-model readiness, and emergency downgrade plans are the appropriate institutional response regardless of which government, vendor, or direction is involved.
The Ungoverned lesson: Model access volatility is now a documented recurring risk — continuity planning is no longer optional prudence; it is basic operational hygiene.
5. Europe delayed its high-risk AI rules by up to two years. Reading that correctly matters.
The EU’s co-legislators have agreed to delay the application of the AI Act’s high-risk provisions, which had been due to enter into force on August 2, 2026. The new dates are: December 2, 2027, for stand-alone high-risk AI systems, and August 2, 2028, for high-risk systems embedded in products.
The wrong reading of this delay is relief — governance can wait. The right reading is more sobering: the delay reflects how difficult it is to build the compliance infrastructure, for regulators and institutions alike. The requirements themselves — documented human oversight, bias monitoring, explainability, incident reporting — have not been softened. They have been rescheduled. Meanwhile, in the United States, the Senate is reviving a push for AI labelling requirements, and state-level activity continues to accelerate. The regulatory environment is not pausing. It is fragmenting — which for multi-jurisdiction institutions is harder to govern, not easier. Institutions that use the EU delay to defer building governance infrastructure will face the same requirements later, on more timelines, in more jurisdictions, with less lead time.
The Ungoverned lesson: The EU delay rescheduled the requirements — it did not reduce them. Fragmented timelines make governance readiness more valuable, not less.
6. Agents are going off-script in health systems — and institutions are improvising the response.
Health systems are discovering that AI agents in clinical and operational workflows do not always behave as configured — flagging what they were not asked to flag, declining tasks they were authorised to perform, producing outputs inconsistent with their deployment context. As Becker’s reports, health systems are figuring this out in real time — managing deviations without formal incident response frameworks designed for non-human actors.
This is the accountability gap at its most concrete. The question is not whether agents will occasionally go off-script; they will. The question is whether the institution has a response framework ready when they do: a named owner, a defined escalation path, a documented incident process, a mechanism for tracing what the agent did and why. Most health systems are building that framework reactively — after the first deviation — rather than proactively, before deployment. In clinical environments, that sequencing is backward.
The Ungoverned lesson: Off-script agent behavior requires incident response frameworks designed before deployment — improvised governance is not governance.
7. The Bank of England warned that AI agents could amplify market stress. Financial governance is becoming systemic.
Sarah Breeden, Deputy Governor of the Bank of England, warned this week that AI agents could amplify market stress. The concern is structural: agents executing trades, managing liquidity, and responding to market signals at machine speed can act in correlated ways during stress events — reinforcing rather than dampening volatility at speeds human circuit breakers were not designed to catch.
This extends the argument this newsletter made in Issue #75 about agentic finance: what began as an institutional governance question — who authorized the agent, what limits apply, what is reversible — is becoming a systemic one. When many institutions deploy similar agents responding to similar signals, individual governance is necessary but insufficient. The regulatory perimeter will follow. Boards of financial institutions should expect supervisory attention to agentic trading and treasury operations to intensify — and should be able to demonstrate agent-level authorization, limits, and kill-switch capability before being asked.
The Ungoverned lesson: Agentic finance has graduated from institutional risk to systemic risk — supervisory expectations will follow.
8. A major vendor’s terms update takes effect July 30 — and it redraws a line in the AI supply chain.
Effective July 30 — in under four weeks — Google’s updated Terms of Service explicitly prohibit using AI-generated content from Google services to develop machine learning models or related AI technology. The terms also formally ban jailbreaking, adversarial prompting, and prompt injection outside Google’s official testing programs.
Two questions for your institution before the effective date. First: does any internal AI development, fine-tuning, or data preparation process use content generated through Google services? If so, that practice may become a terms violation on July 30. Second, and more durable: does your institution have a scheduled process for reviewing vendor terms updates before they take effect — or does it discover them afterward? The pattern matters more than the instance — and the pattern is industry-wide, not specific to any one company. AI vendors are increasingly defining the boundaries of permissible use in terms of service rather than waiting for regulation. Institutions not reading those terms on a review cycle are accepting risk they cannot see.
The Ungoverned lesson: Vendor terms of service are now AI supply-chain governance documents — put them on a review calendar.
9. Cambridge scholars just showed that the AI metaphor your board uses is itself a governance decision.
A new anthology from the AI & Geopolitics Project at the University of Cambridge’s Bennett School of Public Policy — Reimagining the AI Arms Race, with essays from Verity Harding, Sir Lawrence Freedman, Professor Dame Diane Coyle, Professor Dame Wendy Hall, former Japanese Foreign Minister Kono Taro, and the Paris AI Summit envoys — makes an argument boards should hear, because it applies far beyond geopolitics. The framing of AI as a zero-sum race, Harding argues, is a self-fulfilling prophecy: the threat of a simplistic narrative is not that it is true, but that believing it makes it so. If leaders believe they are in a race with no finish line, the only rule becomes “don’t slow down” — and every safeguard must justify itself against the fear of losing.
Translate that to the institutional level and the lesson lands squarely on your boardroom. If leadership frames AI as a race — against competitors, against peers, against disruption — governance gets framed as drag, speed becomes the metric, and dependency accumulates unexamined. Baroness Martha Lane Fox’s essay in the same collection supplies the counter-model: aviation, where safety culture is inseparable from innovation, and where risk management is a precondition for delivering benefit — not a brake on it. Flying is the safest way to travel because of its governance, not despite it. The question for your board is not whether to move fast or slow. It is which metaphor is silently setting your risk appetite — and whether anyone has ever examined that choice deliberately.
The Ungoverned lesson: Your institution’s AI narrative is a governance input — boards should choose it consciously, because it determines how every other control is weighed.
10. Global governance architecture is forming — with or without your attention.
The United Nations’ Independent International Scientific Panel on AI released its preliminary report. The UN Global Dialogue on AI Governance is underway — early steps toward international assessment machinery for AI, loosely modeled on the climate governance experience. Meanwhile, the Cambridge anthology’s research arm documents something the race narrative obscures: across nine international bodies and over a thousand policy documents, forums representing very different political systems are converging on functionally equivalent governance priorities — accountability, workforce adaptation, data protection — not through negotiation, but because they face the same problems.
For institutional leaders, the practical takeaway is not that a global AI treaty is imminent. It is that governance baselines are hardening across jurisdictions simultaneously — UN panels, EU implementation, US legislative activity, sectoral certification like CHAI’s in health — and the institutions that build a defensible governance floor now will find themselves already compliant with most of what emerges. Those who wait for regulatory clarity will find it arrives as a patchwork, on someone else’s timeline.
The Ungoverned lesson: Governance convergence means building to principle now is cheaper than retrofitting to patchwork later.
The Common Thread
Every signal this week exposes the same gap: the distance between who acts and who answers.
Workers act with tools no one approved. Agents act at speeds no one reviews. Patients’ data moves through integrations no one inventoried. Markets absorb agent behavior no supervisor designed for. Model access shifts on timelines no institution controls. And the vocabulary of teamwork — collaborative, augmented, human-in-the-loop — makes all of it sound governed when much of it is merely accompanied.
The human-agent team is not a bad model. It may be the best model available. Human-plus-AI outperforms either alone across domains, from domain to domain. But a team is not defined by who shows up. It is defined by who is responsible. And right now, in most institutions, the honest answer to “who answers for what the human-agent team does?” is: it depends; we’d have to check; probably the committee.
That answer will not survive the first serious failure. The work of governance is to replace it — with a name, a scope, a review standard, and a stop mechanism — before that failure arrives.
The Seba 12 Ps: What Issue #77 Activates
The Seba 12 Ps of Responsible AI Oversight © — twelve dimensions for anchoring and analyzing the practical aspects of responsible governance — frame every issue of this newsletter. Five are especially active this week.
People — because 41% of workers have been given nothing to prepare them for AI, and because patient trust in health AI is conditional, measurable, and currently unearned by most institutional data practices.
Process — because off-script agents, continuous loops, and machine-speed markets all require incident response, monitoring, and stop authority designed before deployment, not improvised after.
Provenance — because health data flowing through third-party integrations, vendor terms redrawing supply-chain rules, and shadow AI data flows all demand that institutions can trace what moved where, under whose authority.
Preparedness — because the metaphor a board uses for AI silently sets its risk appetite, and choosing that narrative deliberately is itself a governance capability.
Product Ownership — because the entire accountability gap comes down to this: vendors provide the system, but institutions own the outcome — and ownership must have a name.
The Board-Ready Action
The Seba Human-Agent Accountability Map
Before your next board meeting, ask your leadership team to answer five questions about every consequential human-agent workflow in your institution — including the ones nobody officially deployed.
1. Who answers for the outcome? Name the role — not the vendor, not the committee, not “IT.” The human accountable when the human-agent team produces a consequential result, right or wrong.
2. What is the full inventory — including shadow AI? Which AI tools are your people actually using, approved or not? What data flows through them? An accountability map of only the sanctioned tools is a map of the minority.
3. What does meaningful human review actually require? Define what a reviewer must understand, challenge, and verify before AI output becomes institutional action. An approval click is not review. Specify what is.
4. What happens when the agent goes off-script — and what happens when access changes? Is there an incident response process for agent deviations, with a named escalation path? And is there a continuity plan if a model is restricted, a vendor changes terms, or costs spike — tested, not theoretical?
5. Does the team perform equitably — and would the people it serves accept how it works? Is performance monitored across languages, communities, and populations? And could you explain your data practices to the patients, students, or customers involved — knowing, as the CHAI survey shows, that their trust is conditional and their assumptions are already formed?
If your leadership team cannot answer all five, your institution has human-agent teams without human accountability. Closing that gap is not a technology project. It is a leadership one.
Save this. Share it with your board.
What I Am Watching
Whether the shadow AI numbers force boards to treat workforce AI readiness as a governance obligation rather than an HR nicety — 41% receiving nothing is not an adoption gap, it is an accountability gap.
Whether CHAI’s patient survey becomes the reference dataset for health AI consent design — and whether health systems close the measured gap between patient expectations and institutional data practice before regulators or journalists measure it for them.
Whether the Bank of England’s warning marks the start of supervisory frameworks for agentic finance — and which jurisdiction moves first.
Whether the UN’s scientific panel process matures into an IPCC-style assessment body for AI, and what that would mean for institutional reporting expectations.
On the desk for Issue #78: Three major new datasets are under review — new enterprise AI adoption research, fresh industry usage data, and the latest AI incident tracking from MIT. The evidence base for how AI actually behaves inside institutions is getting richer by the month. What that evidence shows — and what boards should do with it — is where this newsletter goes next.
Closing Thought
Teams are not defined by who shows up. They are defined by who answers.
A human-agent team in which no human can explain what the agent did, why it did it, or who authorized it is not a team with an AI member. It is an institution that has delegated authority without retaining accountability — and called it collaboration.
Something will eventually go wrong. Not because AI is reckless, but because every system that acts consequentially in the real world — human or otherwise — eventually produces an outcome no one intended. When that day comes, the question your institution will face is not whether humans were in the loop. It is whether the humans in the loop were actually responsible — or simply present.
AI can be a teammate. It cannot be accountable. That responsibility remains human — and it must have a name, before it is needed.
Gratitude & Acknowledgements
This issue draws on public-interest research made possible by organizations doing the slow, unglamorous work that responsible AI governance depends on. Special appreciation this week to the Coalition for Health AI (CHAI), NORC at the University of Illinois Chicago, and the California Health Care Foundation for the Patient Survey Report — patient-voice data of exactly the kind health AI governance has lacked; to the AI & Geopolitics Project at the University of Cambridge’s Bennett School of Public Policy for the Reimagining the AI Arms Race anthology; and to the Cloud Security Alliance for its agent governance research.
Ongoing gratitude to the communities that shape this work week after week: AMIA (American Medical Informatics Association), the University of San Francisco (USF), the University of Illinois Chicago (UIC), and the readers, board members, clinicians, faculty, and institutional leaders who send signals, ask hard questions, and put these frameworks to work.
This issue also draws on public research from the United Nations Independent International Scientific Panel on AI, NIST, and the European Parliament — public-interest institutions whose open publication of governance research makes independent analysis like this newsletter possible.
Selected References
Anthropic. (2026). Redeploying Fable 5. https://www.anthropic.com/news/redeploying-fable-5
Anthropic. (January 12, 2026). Consumer health data privacy policy. https://www.anthropic.com/legal/consumer-health-data-privacy-policy
Becker’s Hospital Review. (2026, June). AI agents are going off-script: Health systems are figuring it out in real time. https://www.beckershospitalreview.com/healthcare-information-technology/ai/ai-agents-are-going-off-script-health-systems-are-figuring-it-out-in-real-time/
Bloomberg. (June 30, 2026). BOE’s Breeden says AI agents could amplify market stress. https://www.bloomberg.com/news/videos/2026-06-30/boe-s-breeden-says-ai-agents-could-amplify-market-stress-video
Cloud Security Alliance. (April 3, 2026). The AI agent governance gap: What CISOs need now. https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-agent-governance-framework-gap-20260403/
Coalition for Health AI. (2026, January). Patient survey report. With NORC at the University of Chicago and the California Health Care Foundation. https://www.chai.org/blog/chai-releases-new-patient-survey-report-on-health-ai-and-transparency
Computer Weekly. (2026, June). Gartner: Prioritize governance to beat AI hype. https://www.computerweekly.com/news/366645092/Gartner-Prioritise-governance-to-beat-AI-hype
Deloitte. (2026). State of AI in the enterprise. Deloitte Insights.
Google. (2026). Terms of service [effective July 30, 2026]. https://policies.google.com/terms/update
Harding, V. (Ed.). (2026, June). Reimagining the AI arms race. AI & Geopolitics Project (AIxGEO), Bennett School of Public Policy, University of Cambridge.
Resume Now. (2026). BYO AI report: 41% of workers say their employer has done nothing to prepare them to use AI at work. https://www.resume-now.com/job-resources/careers/byo-ai-report
Seba, F. (2026). Ungoverned: A practical guide to AI minimum viable governance. Amazon. https://us.amazon.com/Ungoverned-Practical-Minimum-Viable-Governance-ebook/dp/B0GY495GG1
TechCrunch. (June 26, 2026). OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn’t be the norm. https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/
United Nations. (2026). Independent International Scientific Panel on AI: Preliminary report. https://www.un.org/independent-international-scientific-panel-ai/en/preliminary-report
Additional sources reviewed: EU AI Act implementation timeline updates (high-risk application dates of December 2, 2027 and August 2, 2028); World Economic Forum on human-AI role design; Politico reporting on US Senate AI labeling activity; UN Global Dialogue on AI Governance; Microsoft agent deployment tracking; NIST AI Agent Standards Initiative; 2026 CISO AI Risk Report.
About Dr. Freddie Seba
Dr. Freddie Seba helps boards, trustees, and executive leadership teams build practical AI governance before AI failures make governance unavoidable.
Scholar-operator, Silicon Valley founder, and global executive — EdD · USF · MBA · Yale · MA · Stanford — he translates fast-moving AI developments into governance frameworks leaders can deploy now, across healthcare, financial services, and higher education. Non-vendor. Non-partisan. Doctoral research, not advocacy.
Ungoverned: A Practical Guide to AI Minimum Viable Governance — twelve dimensions for anchoring and analyzing the practical aspects of responsible governance — is available now on Amazon.
Booking keynotes and workshops for fall 2026 → freddieseba.com/contact
Transparency
Drafted and refined with AI-assisted tools — this issue, Anthropic’s Fable 5, the same model whose access suspension and redeployment this newsletter covered in Issues #74–77. Naming the tools we use is part of the provenance we advocate. This is disclosure, not endorsement. Final editorial control and responsibility remain with the author. Educational content only — not legal, medical, financial, or regulatory advice. For reprint or licensing inquiries: contact@freddieseba.com
Connect
freddieseba.com · LinkedIn: @freddiesebaprofile · YouTube · Spotify · Apple Podcasts: AI Governance with Dr. Freddie Seba
#AIGovernance #ResponsibleAI #BoardOversight #AgenticAI #Ungoverned #AILeadership #HealthAI #AIEthics #TrustworthyAI #HumanJudgment #AIMinimumViableGovernance #HumanAgentTeams #ShadowAI #PatientTrust #Seba12Ps
Ungoverned: AI Ethics & Governance for Leaders, Boards & Trustees By Dr. Freddie Seba © 2026 Freddie Seba. All rights reserved. The Seba 12 Ps of Responsible AI Oversight © and AI Minimum Viable Governance are author-developed governance frameworks used for educational, board-readiness, and leadership-development purposes.
Leave a Reply