Copilot AI

Open Copilot at /app/copilot to ask questions about connected integrations. Available actions depend on the registered tools, enabled integrations, your role, customer mapping and runtime policies. Copilot cannot perform everything available in a vendor portal, and a tool available in an SOP is not automatically available in every chat or MCP connection.

Start with a read-only question

Choose Customer focus from the searchable selector, then name that customer in the prompt:

List up to five open Autotask tickets for [customer] only. Include ticket numbers, numeric IDs, titles and statuses. Explain whether the result is complete. Do not change anything.

Check the tool results, customer and any run link provided with the answer. A natural-language answer alone is not evidence that a tool succeeded. A result may be partial, refused or failed; do not assume every tool chip represents success.

For Autotask, “open” and “closed” use the account's configured terminal-status policy. Text search can match title or description and does not imply newest-first ordering. Returned counts are not necessarily the account's total. Follow the Autotask walkthrough for exact-ticket checks, pagination and saved-session verification.

Keep customer scope explicit

The selected customer supplies context to the conversation. Confirm it before each task and after reopening a saved session. When changing customer, state the new customer explicitly and check the returned records. If a name or mapping is ambiguous, resolve that ambiguity before continuing; do not silently broaden the query to every customer.

Use the session list to return to a conversation. History provides context, but an old reply is not a fresh vendor reading. Request a new read when current state matters.

Proposed changes and approvals

Writes may create a draft, require approval, execute only after an explicit apply step, or be blocked by policy. The exact route depends on the tool and current configuration. Follow the controls actually shown; do not infer that a request in chat grants unrestricted authority.

Before approving, inspect the customer, exact target, tool and arguments. After approval, verify the execution result and the intended vendor effect. An approved request is permission to attempt the action; it is not evidence that the vendor accepted it. If the response is lost or uncertain, reconcile with read-only evidence before deciding whether any retry is safe. See Approvals.

Cost and configuration

Conversation usage is shown where available; model calls consume the workspace's configured budget. Use a bounded set of prompts during evaluation rather than repeatedly polling with AI.

Admins manage model and execution settings under Settings → AI Config (/app/settings/ai-config). Changing a confidence threshold or customer autonomy mode changes behaviour; it is not a troubleshooting step for a blocked action. A confidence threshold alone does not establish safety and cannot clear mandatory human or customer-policy requirements. See How AI actions are governed.