On June 3, 2026, a room at UN Headquarters spent a morning on one question: what does AI transformation actually look like for governments and public institutions?
The answer the room landed on was multi-agent AI systems. And OORT AION was the only company in the room with a live, running proof.
The half-day workshop "AI Transformation and the Future of Public Institutions: From Theory to Implementation" was co-hosted by the Permanent Missions of Moldova and Tajikistan to the United Nations, UNITAR, and the Blockchain Alliance International (BAI). Paraguay and Kenya co-sponsored. Columbia University served as academic partner. OORT, alongside Lightening AI and Dragon Global, was a named technology partner.
Its outputs are formally designated for reporting to the second session of the UN Global Dialogue on AI Governance in May 2027, which means the morning's discussions don't end with the event. They become an input to a formal intergovernmental process.
The room included diplomats, ambassadors, intergovernmental officials, academics from NJIT and NYU Stern, and UNDP's Chief Digital Officer. This was not a conference panel. It was a structured policy session with a defined institutional pipeline.
OORT appeared in three of the workshop's four substantive segments.
Dr. Max Li, OORT's founder and CEO and adjunct professor of electrical engineering at Columbia University, delivered the keynote on multi-agent AI architecture and its implications for governments.
Dr. Sean Yang followed with a twenty-minute live demonstration of AION - OORT's own deployed multi-agent system, running on OORT's own operations - and later moderated the academic panel.
Michael A. Robinson, Chairman of the OORT Foundation, joined UNDP's Chief Digital Officer and the Ambassadors of Paraguay and Kenya at the closing policy roundtable.
Dr. Li's keynote opened by retiring the prompt-and-response mental model most people still carry. An agent is not a chatbot. It runs a continuous loop, perceiving its environment, reasoning, and acting, with memory and tools far beyond text generation. The shift is from a single generalist model to a coordinated team of purpose-built agents: one classifies correspondence, a second synthesizes research, a third drafts responses, a fourth checks compliance, a fifth routes edge cases to human review.
The intelligence of the system emerges from orchestration, the layer that sequences agents, passes context, and manages the pipeline, not from any single component.
Human-in-the-loop (HITL) oversight was treated as an architectural choice, not rhetorical reassurance. At defined points, such as. high-consequence decisions, novel situations, irreversible actions, the system pauses and routes to a human reviewer. The argument was precise: multi-agent AI does not remove institutional accountability. It relocates and concentrates it at the orchestration layer, making that design work more consequential than ever.
For governments, the implication is direct. A ministry that never built a centralized legacy case-management system is better positioned to deploy multi-agent workflows than one migrating from decades of technical debt. Digital leapfrogging, applied to AI, is a strategic position.
When Dr. Li stepped back from the podium, Dr. Sean Yang walked forward with one opening slide:
"Not a mockup. The actual product, running on ourselves."
OORT didn't build AION and test it on clients first. It built AION for itself, putting its own communications operations on the line. The twenty minutes that followed were two live walkthroughs — each built around a single agent, each translated into the institutional context the audience came to evaluate.
The first agent monitors approximately 80 influential voices in OORT's space on X: technical commentators, community voices, journalists, early-adopter advocates. At that volume, raw monitoring produces noise. So the agent's first job is filtration: read each post against criteria, discard irrelevant ones silently. For relevant posts, it does three things in sequence.
It flags the post. It drafts a suggested reply in OORT's brand voice, trained on OORT's own communications corpus, not just any generic corporate language. And it explains, in plain language, why: what it saw, why the moment warranted a response, what opportunity or risk it read.
That third step -the explanation - is not secondary. A system that flags and drafts without reasoning creates delegation, not collaboration. The reviewer accepts or rejects a recommendation without evaluating the judgment behind it. When the agent explains its reasoning, the human reads an argument, agrees or disagrees, and decides. The locus of evaluation stays with the person.
Nothing is ever posted automatically. The workflow ends with a human reading the flagged post, the suggested reply, and the explanation, then choosing whether to post, adapt, rewrite, or do nothing. The agent proposes. A person speaks.
For governments: replace "80 influential voices in OORT's sector" with citizens, journalists, civil society organizations, and local media. The function is identical; a continuous, filtered, prioritized pulse on what a constituency is saying, paired with a draft a communications officer can evaluate and post.
The second agent processes inbound questions, the kind that arrive continuously across platforms and Telegram, at all hours, in multiple languages, at volumes human teams cannot cover off-hours.
The design decision that defines its character is not capability. It's constraint.
The agent answers only from a curated knowledge base built exclusively from OORT's own verified sources: official news posts, product documentation, public announcements. It does not access the open internet. It does not synthesize from general training data. It does not extrapolate beyond what it has been given.
Every answer is dated: the agent states when its information was last updated. When that exceeds a defined threshold (30 days for time-sensitive topics, 180 days for stable content) it flags the staleness explicitly and recommends verification against the live source.
And when it doesn't know something? It says so.
The demo's standout moment, what Sean called "the money shot", came when the agent was asked about a recent development not yet in the knowledge base. It did not attempt an answer. It stated directly that it lacked current information, named the date of its most recent entry, flagged potential staleness, and pointed to a specific live source.
Sean let it sit before speaking: "That refusal to make things up is the feature."
He was not being ironic. An AI system in a public-facing role that fabricates when it doesn't know is not a communications tool. It's a liability. In a consumer context, that's a product problem. In a government context, where a citizen may decide about eligibility, compliance, or legal rights based on the answer, it's a governance failure.
The designed preference for honest incompleteness over confident fabrication is what makes the system trustworthy in high-stakes contexts.
For governments: citizens ask about benefit eligibility, business registration, permit requirements, regulatory timelines. The gap between what they need and what's accessible is currently filled by expensive call centers, link-returning search interfaces, or silence. A Q&A agent grounded in official policy documents, dated and updated as policy evolves, configured to admit its limits, is a fourth option, one that operates 24 hours a day, in the citizen's language, from authorized sources, without making things up.
Dr. Sean Yang moderated a fifty-minute academic panel with Prof. Grace Wang (NJIT; IEEE Fellow; sole academic on the New Jersey Supreme Court's committee on AI and the Courts; AI Subject-Matter Expert to DHS's SAGE program) and Prof. Xi Chen (NYU Stern; former Principal Research Scientist at Amazon Ads; research spanning preference alignment, fairness, differential privacy, and mechanism design).
The panel's synthesis was direct: the research frontier is ahead of deployment. Deployment is ahead of governance. Governance at every level is playing catch-up to a technology that is not waiting.
The frameworks that dominate current AI policy discourse are organized around pre-deployment approval — and are far weaker on the obligations that matter most: regular third-party auditing, mandatory disaggregated error-rate disclosure, and clear contestation processes. The gap will not close by waiting. It closes by building systems with accountability designed in from the start.
The closing roundtable was moderated by Robert Opp, UNDP's Chief Digital Officer, and brought together Michael A. Robinson of the OORT Foundation alongside the Ambassadors of Paraguay and Kenya.
The question threading through the discussion: how does a developing nation, with constrained budgets, legacy infrastructure, and a civil service not yet fluent in machine learning, make AI decisions that are sovereign, ethical, and durable?
Robinson framed digital sovereignty as a procurement decision facing finance ministries today. The conventional path to AI capability runs through hyperscale cloud providers: infrastructure abroad, under foreign jurisdiction, dependent on commercial terms that can shift. Decentralized infrastructure offers an alternative: computation distributed across nodes that can include locally controlled infrastructure, data processed without leaving national boundaries, governance embedded in the architecture rather than in a vendor's terms of service.
For public institutions where accountability stakes are high and opaque outputs carry severe consequences, verifiability is foundational, not a luxury.
The workshop's most consequential attribute is not what was said in the room. It's where the outputs are going.
The June 3 event is formally designated for reporting at the second session of the UN Global Dialogue on AI Governance in May 2027. The Global Dialogue is the process through which member states and stakeholders develop shared positions that inform normative instruments, capacity-building priorities, and UN-agency coordination. When a workshop's outcomes are designated for reporting into it, the June 3 discussions become an input to a formal governance track.
For OORT, that pathway carries the structural-fit argument, that the architecture exists, runs today, and addresses problems public institutions call urgent, from a single demonstration into a recurring series and, eventually, the documented record of a formal UN process.
A UN-convened room independently identified multi-agent AI systems as the leading model for institutional AI transformation. OORT AION was the working proof it already runs.
The convenors validated the concept. AION was the evidence.
Learn more about OORT AION at aion.oortech.com
A UN workshop on AI transformation for public institutions. Multi-agent AI, governance gaps, and a live AION demo. Full recap from OORT.
%20(1).png)
The OORT Foundation completes its final quarterly buyback and burn: 3,000,000 $OORT removed from circulation; and activates Equilibrium Burn, a community-driven 1:1 matched framework backed by a 30,000,000 $OORT allocation.
