Teams and policies

Roles

There are four, and they cover the whole surface today:

A per-team setting can restrict shared-context editing to Admins, for teams that want a tighter default. More specialized roles — context approver, billing manager, auditor — arrive with the features that need them; custom roles are an enterprise concern.

Shared context is versioned context

A team’s working knowledge — standards, product context, terminology, reusable prompts — lives in the same versioned context machinery as personal context, at team scope. Members’ conversations can draw on it, on any model.

Because context is versioned and every interaction records the versions it used, changing team context does not rewrite the past. A conversation from last week still shows the version it actually ran on. A personal project can also be moved onto a team.

A prompt library is not a separate system here: a reusable prompt is a versioned context asset with a title, tags and an optional folder path, so a team library is browsable rather than a flat list.

Policies: control at the input

Two mechanisms, both deterministic:

• Model allowlist. A team defines which models and providers may be used at all.
• Sensitivity rules. Context items and conversations can be marked sensitive, and marked content may only go to designated models — for example, only to models running on your own hardware.

A team can also set standing instructions that apply as a floor for everyone, and force a context-overflow behaviour and retention policy rather than leaving them personal.

Policies on what a model may answer — topic restrictions, answer filtering, DLP-style scanning — are the next release, and they are best-effort by nature. OneAI keeps that distinction explicit instead of presenting output filtering as an equivalent guarantee. See the overview for why that line matters.

Who did what

An activity log records team actions: conversations created or deleted, standing instructions changed, members invited, policies edited. It records structure and actor — not prompt or response bodies. It is a separate thing from conversation history, on purpose.

Auditing what was sent where is a different question, and the record already answers it: every interaction stores the model, the context versions and the policy conditions that applied. Compliance reporting — exports, retention holds — is an enterprise feature that builds on the same data.

Getting a team

Invites go by email: invite, sign in, land in the team. Team members switch between personal and team workspaces in the app. Team billing — a pooled usage budget and organization-level keys — arrives with the Team plan in the next release.