Tools & actions

The assistant ships with a tool / function-calling layer so it can do things, not just describe them. Tools split into two modes:

• Read-only tools (lookup_finding, find_subprocessor, get_document, list_drift, get_recent_scan) auto-execute during a turn — there is no destructive side-effect, so we save you a click.
• Mutating tools (resolve_drift, request_review, export_audit_log) are proposed by the model and surfaced as a confirmation card. The action only runs when you click Run — and only if your role permits it.

Every tool execution writes an audit-log entry of the form assistant.tool.<name> with the exact arguments and the result, so the chain covers anything the assistant did on your behalf.

The tool catalogue

Read-only

Tool	What it does	Typical question
lookup_finding	Search scan findings by canonical key or label.	"Show me where OpenAI was detected."
find_subprocessor	Search the active subprocessor inventory.	"Do we use Sentry?"
get_document	Pull the latest version of a specific document, optionally focused on one heading.	"Quote the retention section of my Privacy Policy."
list_drift	List the most recent open drift alerts.	"What drift is currently pending?"
get_recent_scan	Summarise the most recent scan with finding counts and duration.	"When did we last scan, and was it healthy?"

Mutating (require confirmation + role)

Tool	What it does	Roles allowed
resolve_drift	Approve a drift alert (republishes the latest draft of the affected document) or dismiss it.	owner / admin / approver
request_review	Move a draft into pending_approval and reset the 48h reminder + 7d escalation timers.	owner / admin / approver / member
export_audit_log	Generate a signed CSV / JSON / PDF audit-log export and link to the download.	owner / admin

If the assistant proposes a tool you don't have a role for, the confirmation card surfaces a permission-denied error rather than silently failing.

Confirmation cards

When a mutating tool is proposed, the assistant pane renders a card showing:

• The tool name (human label, e.g. "Resolve drift alert").
• A one-sentence summary of what will happen.
• The exact JSON arguments the model produced.
• Run and Cancel buttons.

This is deliberate. The assistant cannot covertly mutate your tenant — every state change goes through a click that is recorded against your user, with the same permission gate the dashboard uses.

Tool results in the conversation

Read-only tool results render as compact cards beneath the assistant message that triggered them; mutating tool results render the same way after you confirm. Each card includes:

• A status icon (green check / amber warning / red error).
• The model-summary string the LLM saw.
• An Open link, where applicable, to the relevant dashboard page.

The model sees the tool result on the next turn, so a follow-up question like "OK, resolve drift d1234abc and then open the AI Trust Center" can chain a mutating call and a read-only call in one back-and-forth.

Programmatic access

If you need to invoke a mutating tool outside the chat pane (e.g. from a runbook or a webhook handler), use the REST API instead. Every assistant tool corresponds to an existing dashboard server action or API endpoint:

• resolve_drift → POST /api/v1/drift/{id}/resolve (REST API page)
• request_review → not exposed yet; use the dashboard
• export_audit_log → GET /api/audit/export?format=…