What the AI can and cannot do
Last updated 2026-09-01 · Governing document: the Avieros Constitution, v1.3
Avieros runs an AI workforce inside your business. That deserves a straight answer to four questions, in this order: what may it read, what may it do on its own, what must it ask you first, and what is it forbidden from ever doing. Then: who is accountable when it gets something wrong, and how you appeal.
1. Read, act, ask, never
| Class | What it covers | What that means in practice |
|---|---|---|
| May read | The data of the one business it is working for — its records, its documents, its connected accounts — scoped by the same database rules that apply to a person in that business. | An agent has no wider view than the member it acts for. It cannot read another business, because the database refuses the query, not because the agent was told not to try. |
| May act alone | Reversible work inside that business: drafting, organising, summarising, preparing, analysing, and writing to Avieros’s own records. | If undoing it costs you nothing but a click, the workforce may do it and tell you afterwards. |
| Must ask first | Anything that leaves the building or cannot be taken back: sending a message to a third party, moving money, changing a connected account, altering infrastructure or DNS, deleting anything. | The action is prepared and then stops, waiting for a human. Approval is per action — not a mode you switch on once and forget. |
| Never, at all | The three below, plus operating power, water, emergency services or weapons systems — a refusal the Constitution requires to be architectural rather than a policy someone could relax. | Not gated. Absent. |
2. The three it may never do
No agent initiates a payment, transfer, payout, refund or purchase without a human approving that specific movement. Autonomy at Avieros never extends to your money. Payment processing is disabled entirely at present — see Subprocessors for what is switched on today.
The workforce drafts; a human sends. An email, message or filing that will arrive under your name, your domain or your signature waits for a person to approve that exact text. There is no “send on my behalf” setting that survives a restart, because there is no such setting.
Deletion of your records is a human action. Where the workforce would otherwise remove something, it marks it instead and asks. Your own right to delete your data is separate, always available, and exercised by you — see Privacy & Data Rights.
3. Who is accountable when it errs
Avieros is accountable. Not the model, not the vendor, not “the algorithm”.
- An AI action is an Avieros action. If the workforce does something wrong inside your business, that is Avieros’s error to answer for and to fix. “The AI did it” is not a defence a platform gets to use about its own product.
- A named human owns every consequential action, because a human approved it. The approval is the accountability: nothing consequential happens that a person did not authorise.
- The model vendor is a subprocessor, not a shield. Avieros runs on external model providers and, like every AI system, they can be wrong. The rules on this page govern what the platform does with model output. They are guardrails on action and publication — not a claim that any model is correct.
- Failures are reported, not smoothed. If an AI-driven action cannot complete, the system says so with a reason. It does not report success it did not achieve, and an automated gate in the release pipeline fails the build for code that would.
4. How you appeal
An appeal path that exists only as a promise is not an appeal path. Here is the real one, with the real limits.
| If | Do this | What happens, honestly |
|---|---|---|
| The workforce did something wrong in your business | legal@avieros.com with what happened and roughly when. | It reaches a person, not a queue. Avieros is a small team, so the response window is a person’s working hours — not a support desk. That is stated plainly rather than dressed as a service level nobody staffs. |
| Something Avieros published about you is wrong | corrections@avieros.com | Handled as a factual dispute against the primary record, not as a complaint to be managed. If the record is wrong, the page changes. |
| You want the workforce stopped in your business | Halt it from your own controls. | A human can always countermand or halt AI activity. The platform-wide emergency stop is a separate control and its live activation is not yet verified — see honest status below. Your per-business controls do not depend on it. |
| You want it all to stop and to leave | Export and delete from avieros.com/data-rights. | Self-serve and working today. Leaving with your data is a guaranteed right, not a retention conversation. |
5. What makes these more than words
| Rule | Enforced by |
|---|---|
| Human authority over consequential actions | Written permission boundaries; payments, sends, data mutations and infrastructure changes stop pending human approval. |
| Never reports success it did not achieve | An automated honesty gate in CI fails the build for code reporting connected, sent or done for things that did not happen. Deploy claims require proof against the real production hostname, never a preview. |
| Cannot read another business | Row-level security on 545 of 545 production tables, with cross-tenant checks in CI that are themselves proven able to fail. See Security, including the one open exception. |
| Party neutrality in civic work | One scoring path for every official, party-balanced test data, and a published methodology kept in sync by a build-failing test. See Methodology. |
| No claim without a primary source | Enforced in the publishing pipeline: no source, no render. |
| No silent rule changes | A methodology or rule change must ship with a public changelog entry in the same change, or the build fails. See Transparency. |
Control & authority The operator surface where these boundaries are set and AI activity is halted. Signing in is required — it is your business’s controls, not a public page.
6. Honest status
- The Constitution’s cryptographic ratification is not complete. A signed copy with its SHA-256 hash in a public immutable log is defined and not yet done. The commitments are in force as operating law today; the ratification record is pending, and this page will not say otherwise until it exists.
- The platform-wide AI kill switch is built and its live-production activation is not verified. Tracked as
KILL-SWITCH-DOES-NOT-STOP-WORKERS. An unverified emergency control is not an emergency control, so it is not claimed as one. Per-business human override does not depend on it. - AI-generated summaries are not shipped yet. The labelling rule — a persistent AI-generated label and a link to the official source — is locked before the feature rather than after, which is the only order in which such a rule means anything.
Back This site runs no script, so it cannot know where you came from and will not pretend to. Your browser’s back button does. The Trust Center overview is here otherwise.