Office

Office is an optional product for a paid organization. It adds an operator that reviews the business records available to that organization, prepares useful work, and brings decisions back to a person.

An owner can customize what Office watches and what it does with it by asking for skills — standing instructions such as a morning check of a customer sheet — and by connecting Google Sheets as a read-only source.

Who can use it

An organization owner enables Office explicitly. Enabling is a single action, not a questionnaire or an onboarding wizard. Other members can view an enabled Office, but only an owner can enable it or approve the actions that need a human.

Eligibility comes from the organization’s plan. An owner’s personal plan does not grant Office to an organization they belong to, and a member cannot enable it on the organization’s behalf.

What it may do on its own

Office may read the organization’s available records, prepare work, write files in a sandbox, and build. It may not deploy, contact a customer, or delete or merge records without an owner’s approval. An approval is for one recorded operation and one set of arguments; a later code change is a new decision, not a standing permission.

Skills guide what Office looks at and when. They do not grant capabilities beyond this policy: a skill is instructions, not a new tool or a new permission.

Customizing Office with skills

A skill is a named piece of standing guidance: what to look at, when, and what to bring back. An owner creates one by asking in plain language, for example “Check the Customers sheet every morning and tell me who is missing a follow-up.” Office drafts the skill and shows it for review before it governs any run. If an existing skill clearly matches the request, Office proposes an update to it instead of a duplicate; if two could match, it asks rather than guessing.

• Review and approval. Office runs a read-only test of a new or changed skill against real organization data, in a conversation other members cannot see. A skill cannot be approved until that test passes; a failed or blocked test — for example, no Google Sheets connection — leaves the previous version in charge. Only approval makes a version current; until then the previous version keeps running.
• Versions. Every change is a new version. The approved version is what scheduled runs use; earlier versions are kept, and an owner can restore one. Up to 20 versions are retained per skill, and an organization can have up to 50 skills.
• Schedules. A skill carries a schedule — daily, weekdays, weekly, monthly, hourly, or at intervals of ten minutes or more — or it can be run on demand. Office checks schedules on the existing operator loop, so anything finer than every ten minutes is rejected; time spent paused or disabled is not caught up afterwards.
• Runs and history. Each run is recorded with the skill and version that produced it and appears in Office history; test runs are visible to owners only. “Run now” queues a manual run that uses the same policy as a scheduled one and can still ask for approval. A due run is recorded as skipped when an approval is pending, the Office is paused, or credits are unavailable.
• What a skill can reach. Skill scripts run in a sandbox with no network and no credentials. They can compute over the data the operator supplies; anything from outside arrives through a read the operator performs itself, never through the script. Skills carry no tool allowlist of their own — they cannot lift an approval or add a capability.
• Who can change what. Only an owner can create, edit, approve, enable, disable, restore, or run a skill. Members can view approved skills and non-test history.

Google Sheets

Office can read a Google Sheet that an owner connects for the organization. Open Office → Connections, choose Google Sheets, and complete the provider’s consent screen. The connection belongs to the organization, not to the person who connected it.

• Reads are read-only: Office can read values, not change the sheet.
• If the connection is revoked or expires, skills that depend on it fail closed with a connection error until an owner reconnects it, and their tests cannot be approved.
• Disconnecting in Office removes the organization’s connection. The provider-side grant can be removed from the provider’s account settings.

Billing

Model usage consumes the organization’s AI credits. Office does not have a separate token meter or a separate compute invoice, and it does not invent a price for work. If credits or paid access stop, new work stops.

When work stops

A downgrade, a suspended account, or exhausted credits stops new work without discarding what already exists. Records, conversations and pending decisions remain readable, and an owner can still decline a pending decision. Consequential approvals resume when the organization can start work again.

The same applies to a broken credential: new work is unavailable until an owner repairs setup, while the records stay where they are.

If a skill run is skipped or blocked

• Connection required — reconnect Google Sheets in Office → Connections, then ask Office to re-test the skill.
• Test has not passed — approval stays disabled until a new test passes.
• Run in progress — wait for the queued or running occurrence to finish.
• Approval pending — a due run is skipped while an earlier version of a skill is waiting for approval; approve or decline it and the next occurrence runs normally.
• Interrupted run — if a turn stops mid-run, the run is closed automatically and the next scheduled occurrence is the retry; nothing needs fixing by hand.
• Paused or out of credits — the occurrence is recorded as skipped and is not caught up when work resumes.
• Sandbox unavailable — the deployment has no isolated runner; contact support.