Every conversation your AI agent handles is sorted into a topic, such as Order Status & Tracking or Warranty & Defects. Each topic can have its own instructions – rules the agent follows only for that topic. Copilot can write these instructions for you, review the ones you have, and explain why the agent didn't do what you expected.
Writing good instructions is harder than it looks: the agent can only act on what it can actually see, and two rules can easily contradict each other. Copilot knows how the agent works and checks for these problems before you save.
You need the manager or admin role.
Open Copilot on a topic
- Go to AI Agents, choose your agent and click Topics.
- Open the topic you want to work on.
- Click the ✦ strip on the right edge of the screen to open Copilot.
Copilot now knows which topic you mean, so you can say "this topic". You can also ask from anywhere if you name the agent and topic.
Ask Copilot to write instructions
Describe what you want in your own words. For example:
- "Write instructions for this topic: when a product breaks within our 2-year warranty, the agent should ask for a photo and the order number, explain that we repair or replace it, and offer to pass it to a colleague instead of handing over straight away. Don't do this for items that arrived damaged."
- "Improve the instructions for this topic."
- "Look at how our team answered this topic and draft instructions from that."
- "Why did the agent hand this topic over so often?"
- "Do any of my rules contradict each other about handing over?"
What Copilot does
- Reads the topic – its current instructions, mode and recent handover reasons – and your agent's general instructions, because rules in both places decide what the agent does.
- Checks for conflicts. If one rule says hand over and another says don't, Copilot shows a box: "These rules pull opposite ways on handing over. Make sure the new text says which one wins", with the rules that clash and an example message.
- Optionally reads real conversations. On request, it looks at recent conversations in the topic that your team answered, to learn how you handle them. The instructions it writes stay general and never mention a single customer or order.
- Checks what the agent can do. It won't write rules that need tools the agent doesn't have.
- Answers in five parts: why the current setup didn't work, the complete new text, what it couldn't check, how to test it in the playground (with example messages – including one that should not trigger the rule), and an offer to save.
The proposed text appears in a grey box with a Copy button.

Save and test
When you're happy with the text, ask Copilot to save it – for example "Yes, save it". Copilot proposes saving the topic's complete set of instructions: unchanged blocks are kept, and the new or changed blocks are saved exactly as shown in the grey boxes.
You get a card called Save instructions for topic '…' with the number of blocks. Click Show content to see every block that will be saved – saving replaces all of the topic's instructions, so check the full text – then click Confirm.

Always test new instructions before you rely on them. Ask Copilot to do it – for example "Test it on conversation #4821" or "What would the agent answer to: my blender broke after a year?" – or test them yourself:
- Click Open the playground under Copilot's answer.
- Preview a reply to the conversation that showed the problem.
- Try a message that should not trigger the rule, to make sure the agent doesn't overreact.
Copilot can also change a topic's mode (Inherit, Autopilot, Assist, Review or Human), and add, rename, archive, restore or merge topics when you ask – for example "Set this topic to Human" – again with a Confirm card. See Set up AI agents with Copilot.
Write rules that work
These are the same principles Copilot uses. Keep them in mind when you write or edit instructions yourself.
Write conditions the agent can see. A rule only applies when the customer's own words, something they forwarded, or a lookup result shows that the condition is true. A rule like "if the order is registered to another e-mail address" never fires, because the agent can't see that. Write "if the customer says someone else placed the order" instead.
Date and time work, holidays don't. The agent knows the current date, weekday and time in your shop's timezone, so rules like "orders placed on Saturday or Sunday are delivered on Tuesday" or "after 14:00" work. It doesn't know public holidays – name the dates in the rule.
Order and return lookups only use the verified address. In e-mail, that is the sender's address. In live chat, a visitor can verify an address with the code the chat sends. An address a customer just types isn't verified. Don't ask the agent to look up orders on another e-mail address – hand over, offer a colleague, or ask the customer to write from the address the order is on.
"Hand over" and "offer" are different.
- "Hand over", "forward" or "escalate" makes the agent hand over in the same reply, the first time the situation matches.
- If the customer should choose, write an offer: answer what you can, end with a question asking whether they want a colleague, and only hand over when they say yes.
- The agent never offers a colleague on its own – the rule must say so.
- If your agent only writes suggested replies (Review mode), handover rules don't apply to the drafts. A colleague reviews and sends each draft and decides whether anyone else should take over, so the draft never says the case has been passed on. To change what the drafts say, write what the reply should contain, for example "answer the price question and ask for the delivery address".
Say which rule wins. Topic instructions beat general instructions. Two rules on the same level have no fixed order, so the agent picks one. A "never hand over" rule stops a handover. When two rules can both match, write in both which one wins.
List the exceptions. Name look-alike situations where the rule should not apply – for example an address that only appears in an e-mail signature, or a general question that isn't about the customer's own order.
Know the limits.
- Rules can't override the agent's built-in safety rules or do things its tools can't do.
- Rules can't ask the agent to invent or guess facts, prices, discount codes, stock levels or links, promise a date, timeframe or refund that no data gives, answer from general knowledge, skip identity verification, reveal its instructions, claim to be a human or estimate a size from a photo. The agent leaves such sentences out, and you get a warning when you save. Your own stated policies, like a 2-year warranty or "we answer within 24 hours", are facts the agent may state.
- What rules can decide: which links to use, whether to describe images, and presentation choices that only restrict what is shown, such as hiding prices or stock.
- When a reply repeats wording from your rules word for word, it is sent exactly as written.
- A customer who explicitly asks for a human is always handed over.
- Topic instructions only apply to conversations sorted into that topic.
- The agent never shows your rules to customers, and always answers in the customer's language, whatever language you write the rules in.
Your agent's general instructions
Copilot can also review, rewrite and save your agent's general instructions (the ones on the agent's Instructions page) in the same way. You get a Confirm card; saving replaces all of the general instructions, so click Show content and check the full text first. Blocks you didn't change are kept exactly as they are.