All editions · Policymaker Edition

Your Community, Your AI — CC BY 4.0

What Community-Governed AI Actually Does Today

Why a Legislator Should Read a Factual Inventory

Policy made about AI is unusually vulnerable to being made about the wrong thing — about a demonstration that does not exist yet, a fear that outruns the technology, or a marketing claim that will not survive contact with production. The antidote is dull but useful: an inventory of what one community-governed system actually does today, stated plainly, with the still-under-development parts marked as such. This article is that inventory. Its purpose is to calibrate — so that when you hear a claim from any direction, you have a real reference point. (Any unfamiliar term in this series is defined in plain language in the glossary.)

The system described here — Village AI — has been in production since 2025. It is a young system, built by a small team and used by a small number of communities. Some parts work well; some are still being refined. It runs a focused open model rather than a frontier one. Naming what it is, and is not, is the point of the exercise.

What It Does Today

Answers questions from an organisation's own records. When a member asks "When is the next board meeting?" or "What did the council decide about the community centre?", the system searches the organisation's actual records — minutes, announcements, event descriptions, published documents — and answers from that content. It does not infer from general knowledge. If the answer is not in the records, it says so rather than filling the gap.

Helps with drafting. It can draft announcements, correspondence, and communications in the organisation's established tone, because it has learned from that organisation's previous content rather than a generic template. A human reviews and edits every draft before it goes out.

Summarises long documents. A substantial report or a run of announcements can be reduced to key points — useful for members who want to stay informed without reading everything in full.

Works across languages. The platform supports English, German, French, Dutch, and Te Reo Māori. The AI assists with translation of community content; human review is recommended for anything important.

Triages member feedback. When a member submits a question, suggestion, or report through the platform, the system classifies it, investigates where it can, and notifies the member when it has been addressed — reducing the load on the people who would otherwise sort every item by hand.

Where It Acts, Not Just Answers

Articles 1 and 2 drew the distinction a policymaker should hold onto: a chatbot answers, an agent acts. So it is fair to ask whether this system only answers, or acts on an organisation's behalf.

It acts — in carefully bounded ways. The clearest example in production today is feedback resolution. When a member marks an answer as unhelpful, the system does not merely log a complaint for someone to review later. It investigates the correct answer against the organisation's records, and where it can resolve a routine, low-stakes case on solid evidence, it does so — improving its own knowledge so the next enquiry is answered better. That is genuine agentic behaviour: a multi-step action, not a single response.

The boundaries around it are the whole point. The agent acts only within the organisation's own data. It acts only on routine, reversible matters. And the moment it detects a systemic problem — a pattern of related failures suggesting something deeper is wrong — it stops acting and escalates to a human, because that is a judgement call rather than a routine fix. The design intent is that ordinary feedback is handled automatically while anything consequential reaches a responsible person. This is the operational form of the principle in Article 3 and of the EU AI Act's human-oversight requirement: an agent under the organisation's control acts where action is safe and reversible, and steps back where it is not.

What It Does Not Do

It does not make decisions for the organisation. When a question involves values, ethics, or judgement, it stops and routes the matter to a human — the people the community has entrusted with those decisions.

It does not take consequential or irreversible actions on its own. Where it acts, it acts only on routine, reversible matters within the organisation's data. It does not issue communications in the organisation's name, commit the organisation to anything, or make changes it cannot reverse, without a responsible person in the loop. The authority — and the accountability — stays with the organisation.

It does not reach content it was not given. Restricted content stays restricted; one organisation's content stays with that organisation. The AI cannot cross those boundaries, because they are structural rather than matters of policy.

It does not operate without oversight. Every response passes through independent verification layers — checks that are structurally separate from the AI, so the watcher does not share the blind spots of the watched — before it reaches a member.

It does not present uncertainty as fact. Every response carries a confidence indicator. Where the system is drawing on solid records, it says so; where it is on thinner ground, it says that too, and a member can trace a claim back to the specific record that supports it.

How Bias Is Addressed: The Vocabulary System

One of the subtlest forms of bias in AI is linguistic. A system trained on corporate data calls constituents "users" and minutes "posts" — importing a worldview in which communities are consumer platforms and governance is content management. The system addresses this through a vocabulary layer that adapts the whole platform to the organisation type: members or constituents rather than "users," announcements and reports rather than "posts," governance rather than "admin settings," community records rather than "content." This is not cosmetic. The vocabulary shapes how the AI frames a question — it recognises that "how do we communicate this decision to affected residents?" is a different question from "how do we update our user base?", where a generic system would treat them alike.

What Is Still Under Development

Stated plainly, because policy should be grounded in the real state of a thing:

• A higher-capability tier. The system runs a focused open model, fine-tuned per organisation type, so a council, a trust, and a club each receive an AI tuned to their context. The architecture reserves a more capable tier for harder questions; it is defined but not yet in service, so today every question is answered by the organisation-tuned model.
• Individual personalisation — where the AI learns individual members' preferences — is planned but not yet built. For now the AI knows the organisation as a community, not its members as individuals unless they interact with it directly.
• A moderator accreditation path — structured training for members who take on the moderator role — is designed and being rolled out progressively.

What This Means for Policy

The reason to read an inventory like this is to keep policy anchored to reality. What the example shows is that a community-governed AI can, today, answer from an institution's own records rather than the internet's approximation of them; verify its responses against those records through layers that operate independently of the AI; keep its data within the organisation's boundary and out of external training; act only where action is routine and reversible; and stop to ask a human when a matter needs judgement. It also shows the limits: this is a young, narrow system, transparent about what it does not yet do.

Neither hype nor alarm is the right register for a policymaker. The useful register is calibration — knowing what the technology actually does, so that whatever framework you support is built on the thing itself rather than on a claim about it. Article 5 sets out the levers.

Want to use AI tools like these well, and safely? Our free courses — Working with Claude and Agents at Work — teach the practical skills, from getting trustworthy answers to deciding what to hand an agent. For the full technical architecture behind Village AI, see Village AI — Agentic Governance.