Policies

If Thresholds decide what counts as a signal and AI Agents decide who is watching, Policies decide what those assistants are allowed to do.

Permission required: Manage organisation. The screen has three zones: hard ceilings, organisation defaults, and the override grid.

The four rule types

For any combination of workflow and assistant, the rule is one of:

Rule	Meaning	UI label
Allow	Action runs without approval.	Allowed
Deny	Action is blocked.	Blocked
Ask first	Assistant proposes, a human approves before it runs.	Asks first
Limit	Action runs but is rate-limited (for example, max 3 per hour).	Limited

The six workflows

These are the only workflows assistants can be granted. Anything not listed is denied by default.

Workflow	What it covers
Apply scenario	Apply a scenario as the new active plan.
Send communication	Dispatch a report, summary, or notification.
Open decision case	Create a new decision case.
Auto-resolve signal	Close a low-confidence or duplicate signal.
Record data	Record actuals (hours, finance, KPIs) on a user’s behalf.
Run assistant	Master gate: without this, nothing else applies.

If Run assistant is denied, no other policy matters: the assistant simply does not run. This is the per-assistant kill switch.

The three zones on the screen

Zone 1: hard ceilings

The maximum any assistant can ever do, regardless of grid rules. Even with Allow set further down, an assistant cannot exceed the ceiling. Treat this as the constitutional limit.

Typical ceilings:

Apply scenario: never auto-allow without human review.
Send communication: never email external recipients without review.

Zone 2: organisation defaults

The fleet-wide default for each workflow. Applied to every assistant unless overridden in Zone 3. Most organisations set these to Asks first and override per assistant below.

Zone 3: override grid

Every assistant by every workflow as a grid. Each cell shows the effective rule:

Inherit (greyed) uses the organisation default from Zone 2.
An explicit rule (Allow, Deny, Asks first, Limit) overrides the default.

Click any cell to change it. Inheritance keeps the grid readable: only deviations show colour.

Pause-everything switch

A single toggle at the top: Pause all automation.

When on:

All assistants stop. Scheduled and event-driven runs do not fire.
A banner shows in Pulse and Activity.
API calls that would trigger an assistant return a clear paused response.

Use this if you suspect bad behaviour and need to investigate before any more automated actions happen. The switch is logged with user and timestamp.

Versioning and activity log

Every policy change creates a new immutable version, with a link from the old to the new. You can view the full history of who changed what and when. This is the trail clients and regulators ask for.

When your organisation is set up, policies start with a safe Observe default: assistants can detect everything but cannot act without approval.

Mapping to autonomy tiers

AI Agents lets you assign a tier per assistant. Tiers are starting points that pre-fill the grid:

Tier	Pre-fills with
Observe	All workflows except Run assistant set to Deny.
Assist	All action workflows set to Asks first.
Act	Workflows set to Allow with limits on apply and send.
Custom	Whatever you have manually set.

You can override any cell after picking a tier.