Skip to content

Copilot

A copilot that changes your AI agent when you ask

Tell the copilot in your agent's dashboard what you want, in your own words: a friendlier tone, your new return policy, answers to the questions visitors keep asking, a button to book a demo. It replies with proposals that show exactly what would change, and nothing changes until you approve.

Example: a shop owner asks the copilot for changes

Visitor: Our return window is now 30 days instead of 14. Also, the agent sounds a bit stiff, please make it friendlier.

Agent: Two proposals below. Edit Q&A pair updates “What is your return policy?” to 30 days from delivery and keeps its other phrasings. Edit instructions rewrites the tone paragraph in a warmer, more personal style; the rest of your instructions stays as it is.

Visitor: Approved both. Can you also answer the top unanswered questions?

Agent: The three most asked are “Do you ship to Switzerland?”, “Can I change my order after paying?” and “Do you offer gift wrapping?”. Your instructions and Q&A pairs don’t answer them yet. Tell me the answers and I’ll propose a Q&A pair for each.

Keeping a chatbot up to date is a long list of small edits

An AI agent is only as good as its instructions and knowledge, and both drift. Your return policy changes, a new product launches, visitors keep asking something the agent can't answer, the tone feels too stiff. Each fix means finding the right tab, the right field and the right wording: rewrite the instructions without breaking the parts that work, add a Q&A pair with the phrasings people actually use, read a new page into the knowledge. The copilot turns that into a conversation. You say what should be different, it drafts the change against what the agent already has, and you review a clear before and after before anything goes live.

Why change your agent through the copilot

Say what should change, review the diff, approve.

  • Ask in plain words

    No need to know which tab holds what. Ask for “make the tone friendlier”, “add our new return policy: 30 days” or “add a button to book a demo”, or start from one of the examples in the panel. One reply can carry up to 10 proposals.

  • Nothing changes without you

    Every change arrives as a card. Long text such as the instructions or an answer is compared line by line, removed lines in red and added lines in green; shorter values show as Now and Proposed. Approve applies it, Reject leaves everything as it is.

  • Closes the gaps visitors find

    The copilot reads the 20 most asked open questions under Questions to improve and the results of your latest test run, so “answer the top unanswered questions” starts from what visitors really asked, not from guesses.

  • No invented facts

    It doesn't make up prices, policies, dates or links. It works from what you tell it in the chat and what the agent already knows from its instructions and Q&A pairs, and asks you when an answer is missing.

  • Never sees your secrets

    API keys, webhook signing secrets, an action's headers and request body, the Slack webhook URL and the identity verification secret are never read. A key that someone pasted into a Q&A answer is masked before the copilot sees it.

How to use the copilot

  1. Open the copilot

    Open your agent and click Copilot at the top, next to History. Type what you want to change and press Enter, or click one of the examples to start from it. Messages can be up to 4,000 characters long.

  2. Review the proposals

    The copilot answers in a few sentences and puts its proposals underneath. Each card names the change, such as Edit instructions or Add Q&A pair, and shows the before and after. If something is missing, such as the answer to a question, it asks you first.

  3. Approve or reject

    Click Approve to apply a card straight away, Reject to drop it, or Approve all when several are waiting. Q&A pairs, text and website pages start training right away; instructions, the welcome message and suggested questions are saved like a save in the agent form.

  4. Check the result and undo if needed

    Run your tests with simulated customers to see how the agent answers now. Changes to the instructions, welcome message and suggested questions are saved in settings history as versions marked Copilot, so you can restore the version before.

What the copilot can change, and what it reads

  • Instructions, welcome message, suggested questions

    It rewrites the instructions as a full new text compared line by line with the current one, and replaces the welcome message (up to 500 characters) or the suggested questions (up to 6, each up to 120 characters).

  • Q&A pairs

    It adds, edits and deletes Q&A pairs: a question of up to 200 characters, the answer and up to 5 other phrasings. There is one pair per question, so it proposes an edit when a pair with that question already exists.

  • Text and website pages

    It adds a text snippet of up to 20,000 characters, or reads up to 10 public pages into the knowledge with the agent's crawl settings. Reading pages uses one website crawl from this month's allowance, and a page already in the knowledge is updated.

  • Buttons and actions

    It creates and edits custom buttons: the name, when to show it, the button text (up to 40 characters) and the link. It can switch any existing action on or off; API actions and procedures themselves stay yours to edit on their tabs.

  • What it reads

    The agent's name, instructions, welcome message and suggested questions, the guardrails and reply language (read only), the list of sources with titles and statuses, up to 150 Q&A pairs, your actions and the latest test results. It never reads the text of web pages, files or other sources.

  • Data is never an instruction

    Source titles, visitor questions and test results are treated as data. A visitor who writes “ignore your instructions” can’t make the copilot do anything: only your own messages are requests, and even those wait for your approval.

  • Checked again on approval

    Approving follows the dashboard's own rules: the plan's character limit, the length limits, the checks on page addresses and links, and at most 25 actions per agent. If the item was edited after the proposal, approving fails instead of overwriting the newer version.

  • Roles and audit log

    Owners, admins and editors can send messages and approve; viewers can read the chats. A custom role only lets the copilot read what that person may see. Every approved or rejected proposal is recorded in the audit log under the name of the person who decided.

Frequently asked questions

What is an AI agent copilot?

In intoCHAT, the copilot is a chat inside your agent's dashboard that edits the agent for you. You describe a change in your own words, it drafts proposals against the agent's current setup, and you approve or reject each one. It works on how the agent is set up; it doesn't answer your visitors.

Which plan includes the copilot?

Starter, Pro and Enterprise. You can send 30 copilot messages a day on Starter, 150 on Pro and 500 on Enterprise, counted per account across all agents and team members and reset at midnight UTC. On the Free plan the panel shows Included from Starter.

Do copilot messages use my monthly messages?

No. Copilot messages have their own daily allowance and don't count toward your plan's monthly messages. Only your own messages count against it: approving or rejecting a proposal doesn't, and neither does a message the copilot couldn't answer.

Can the copilot change my agent without asking?

No. Every change is a proposal card, and nothing is applied until someone allowed to edit the agent clicks Approve. The proposal is checked again at that moment, so it can't overwrite an edit made after it was drafted.

What can't the copilot change?

Appearance and colours, the guardrails and reply language, lead forms, integrations and webhooks, procedures, API actions and client actions (it can only switch them on or off), deleting sources other than Q&A pairs, training, tests, the team and billing. Ask for one of these and it tells you where in the dashboard to do it.

How do I undo a change?

Instructions, welcome message and suggested questions are part of settings history: each approved change is a version marked Copilot, so you can restore the version before it. Q&A pairs, sources and actions are changed back on the Knowledge or Actions tab; the card keeps showing what was there before.

Does the copilot read my customers' conversations?

It reads the 20 most asked open questions to improve and the latest test results, so it can suggest answers to them. It never reads the text of your web pages, files or other sources, imported helpdesk tickets are only counted, and secrets such as API keys are never part of what it reads.

Is the copilot the agent my visitors talk to?

No. Your visitors chat with your AI agent on your website, by email or in Slack. The copilot works backstage: only your team sees it, in the dashboard, and it changes how that agent is set up.

Improve your agent by asking

Create your agent free and train it on your website, then upgrade to Starter, open Copilot and ask for your first change.

Create your agent free

Free plan, no credit card needed.