On 9 July at the AI for Good Summit in Geneva, the ITU — the United Nations' digital technology agency — announced a new global Focus Group with one mission: make AI agents identifiable, trustworthy, and subject to meaningful human control. The concern is real. Agents that act without direct human input can impersonate people, make unauthorised decisions, and touch sensitive domains like finances and family data. The first working meeting is in Paris this November.

We read that announcement with some satisfaction — because at allMe.life, those three pillars aren't a framework we're waiting for. They're mechanisms already running in our codebase. Here's what each one looks like when you actually build it.

Pillar 1: Identifiable agents

Every AI agent in allMe.life — Jerry (your personal assistant), Sage, and Oracle — carries a named identity on every action it takes. When an agent needs information from the outside world, the request goes through our Sovereign External-Info Gateway with the agent's identity attached, and the gateway checks a per-agent policy: is this agent allowed this capability, for this user, at this account level? No policy row, no call.

Every LLM interaction is also audited: which agent, which model, which user, how long, how many tokens, tied together with a correlation ID. And here's the part we're proudest of: the audit log is zero-knowledge by default. We record the shape of every AI conversation — roles, sizes, timing — but never the content. Your decrypted email body doesn't sit in an audit table. We can prove what our agents did without ever storing what you said.

Pillar 2: Trustworthy behaviour

Trust isn't a promise; it's an architecture. Three layers of it:

A decision spine that rejects its own bad answers. Every agent decision flows through KEEL, our decision engine. It generates multiple options, runs an AI "pre-mortem" (how could each option fail?) on a local model, scores options across six axes — impact, cost, urgency, reversibility, confidence, cross-domain effects — and then applies output discipline: if the winning recommendation is vague ("you might want to consider...") or bloated, KEEL demotes it to no action. Every output field must declare its provenance — which rule, which model, which agent produced it. And every decision records why the chosen option won and why the others lost.

Principles a human approved. Our agents draw on a curated library of thousands of guidance principles (ATHENA). Every principle is extracted, staged — and invisible to the AI until a human reviewer approves it. Un-vetted guidance is dark to runtime by construction.

Your AI runs on our hardware, not someone else's cloud. Agent conversations are served by a sovereign LLM — an open model running on our own GPU servers. Your data isn't training someone else's model, and no third-party AI provider ever holds a conversation with your personal context.

Frontier models see only what a local guard lets through. When a task genuinely benefits from a frontier model (deep research, for instance), the request doesn't go there directly. It passes through the same Sovereign Gateway as everything else: a local safety classifier and a local context-stripping rewriter turn your personal question into an anonymous, self-contained public one before any external model sees it. The frontier LLM answers a stranger's question — it never meets you.

A single, guarded door to the internet. The only way a question leaves allMe.life is through the Sovereign Gateway, which runs — in order — a policy check, a safety classifier (a local model that fails closed: if it can't classify, the request is blocked; minors get stricter thresholds; a query signalling a health emergency returns crisis resources instead of going online), and a context-stripping rewriter that removes names, relationships, and identifiers before anything external sees the query — backed by a regex scrubber that hard-rejects anything resembling an email, phone number, IBAN or card number. One redacted audit row per request, purged after 90 days. No raw query, no token, no PII ever persisted.

Pillar 3: Meaningful human control

Any agent action that affects your data carries a confirmation gate: the action sits in a pending state until you confirm it, and the system records who confirmed and when. Sharing features use dual opt-in — both parties must agree, either can revoke. Batch AI processing (our Eco Plan) only happens under an explicit, time-limited delegation you grant.

The pillar the UN hasn't named yet: don't trust the AI with your data at all

Here's where we go further than the framework. Most companies bolt trust controls onto agents that have standing access to user data. We inverted that: our data layer doesn't trust our AI.

Your content is stored zero-knowledge — encrypted with keys derived from your login that our servers never hold. When AI processing is needed, our patent-pending Temporal Process Isolation (TPI) opens a short-lived, memory-locked session — measured in seconds — decrypts just what's needed, processes it with local AI, re-encrypts the result, and cryptographically shreds the cleartext. The AI gets a time-boxed window, not a master key. Sessions can't even be correlated across time boundaries.

We store your data. We just can't read it. And neither can our agents — except for the seconds you allow, under controls you can audit.

Why this matters

The ITU Focus Group brings together technical, policy, and legal experts to define what trustworthy AI agents should look like. We think the strongest contribution to that conversation is running code. Identifiable agents, self-auditing decisions, fail-closed safety, human confirmation gates, and a data layer that treats the AI itself as untrusted — that's our answer, shipped.

The world is starting to ask the right questions about AI agents. We've been building the answers.

allMe.life is a sovereign personal data platform for individuals and families. Jerry, our AI companion, helps you organise your digital life — without ever holding the keys to it.

Sources: ITU announcement, AI for Good Summit, Geneva, 9 July 2026 (Reuters; The Next Web).