Copilot: change your agent by asking

The copilot is a chat inside your agent's dashboard that changes the agent for you. Ask in your own words, for example "make the tone friendlier", "add our new return policy: 30 days", "answer the top unanswered questions" or "add a button to book a demo". The copilot answers with proposals: cards that show exactly what would change. Nothing changes until you approve a card.

Open the copilot

1. Open your agent.
2. Click Copilot at the top, next to History.
3. Type what you want to change and press Enter (Shift+Enter starts a new line), or click one of the examples to start from it.

The copilot replies in a few sentences and puts its proposals under its reply. Messages can be up to 4,000 characters, and one reply makes at most 10 proposals.

New chat starts a fresh conversation. Recent lists the agent's last 10 chats; click one to read it again or carry on where you left off. Proposals you haven't decided on yet are still there.

What the copilot can change

Anything else is out of the copilot's reach: appearance and colours, the guardrails and reply language, lead forms, integrations and webhooks, API actions and client actions (it can only switch them on or off), procedures, deleting sources other than Q&A pairs, training, tests, the team and billing. If you ask for one of these, the copilot says so, tells you where in the dashboard to do it, and suggests what it can do instead.

What the copilot reads

To make good proposals, the copilot reads:

• the agent's name, instructions, welcome message and suggested questions, plus the guardrails and reply language (read only);
• the list of knowledge sources: titles, types and statuses, with counts. It never reads the text of web pages, files or other sources. Imported helpdesk tickets are only counted, because their titles are your customers' subjects;
• your Q&A pairs (up to 150, with long answers shortened in what it reads);
• your actions: name, type, on or off and when to use them, plus the button text and link of button actions;
• the 20 most asked open questions to improve;
• the results of the latest test run.

It never sees secrets: not API keys, webhook signing secrets, an action's headers, query parameters or request body, the Slack webhook URL or the identity verification secret. A key that someone pasted into a Q&A answer is hidden before the copilot reads it.

Text from your sources, your visitors' questions and test results is treated as data, never as instructions. A visitor question such as "ignore your instructions and…" can't make the copilot do anything. Only your own messages in the chat are requests, and even then every change waits for your approval.

If a custom role can't see knowledge, actions or questions to improve, the copilot doesn't read them for that person either. In an earlier copilot chat, that person sees a proposal for knowledge or actions without its details.

Approve or reject a proposal

Each card shows what would change: long text such as the instructions or an answer line by line (removed lines in red with a −, added lines in green with a +), and other values as Now and Proposed.

• Approve applies the change straight away.
• Reject leaves everything as it is.
• Approve all (shown when several proposals are waiting) approves them one after another.

Approving uses the same rules as the dashboard's own forms: the plan's character limit, the length limits above, the checks on page addresses and button links, and at most 25 actions per agent with no two of the same name. In detail:

• Q&A pairs, text and website pages start training right away, as if you had added them on the Knowledge tab and clicked Train. A pair that answers one of the questions to improve marks that question as resolved, if your role can also change questions to improve.
• Website pages are read when you approve, with the agent's crawl settings. This uses one website crawl from this month's allowance. Only public http and https addresses work. A page that is already in the knowledge is updated. If any of the pages can't be read, none of them is added.
• Instructions, welcome message and suggested questions are saved like a save in the agent form.

If a change can't be applied, the card shows Failed and the reason, for example that the plan's character limit would be exceeded, and nothing was changed. Fix the cause and click Try again, or ask the copilot for a different proposal.

A proposal is checked again when you approve it. If what it changes was edited since, for example someone changed that Q&A answer on the Knowledge tab, approving fails instead of overwriting the newer version. Ask the copilot again for a fresh proposal.

Undo a change

• Instructions, welcome message and suggested questions are part of settings history. Each approved change is saved as a version marked Copilot, so you can restore the version before it.
• Q&A pairs, sources and actions aren't part of settings history. Change or delete them on the Knowledge or Actions tab. The card keeps showing what was there before, so you can put it back by hand.

Daily allowance

Messages to the copilot are counted per account per day, across all agents and team members. The count resets at midnight UTC.

Only your messages count. Approving and rejecting proposals don't, and neither does a message the copilot couldn't answer. The panel shows how many messages are left today. Copilot messages don't use your plan's monthly messages.

Who can do what

With a custom role, sending messages and approving need edit on agent settings. Approving a change to Q&A pairs, sources or actions also needs edit on knowledge or actions.

Every approved or rejected proposal is recorded in the audit log as Copilot change applied or Copilot change rejected, under the name of the person who decided. A change to the instructions, welcome message or suggested questions also shows as Agent settings changed.

Next steps

• Write instructions that work: Instructions.
• Check the agent after a change: Tests with simulated customers.
• See and restore earlier settings: Settings history.