To correct an AI that gave a customer a wrong answer, fix what it rereads before writing, not just the message. With a Muzau agent, from your first edit to a draft on tone, length or wording, a writing rule is created for all your agents. A corrected price or date becomes knowledge suggested to the Brain, which waits for your click. An answer that has already gone out is fixed on the Brain page, then with the customer.
A wrong answer almost always has a cause: missing information, an old price, a way of writing that doesn't sound like you, or a request that went beyond what the AI knew. Fixing the message repairs the exchange with that customer; the next question of the same kind will get the same answer, as long as the cause is still there.
For an AI not to repeat itself, the correction has to reach what it rereads before writing: its writing rules and what it knows about your business. If the answer has already gone out, also correct it with the customer by writing to them from your connected inbox: the follow-up in progress stops (see how to take back control of a conversation).
In "To approve", you correct the agent's text directly, then send it. Muzau compares what it had written with what went out, and works out the direction you took: shorter, a different tone, a different wording. From the very first time, that direction becomes a writing rule for all your agents. Edits are sorted at the platform's regular pass, not at the moment of sending.
If your edit corrects a piece of information, a price, a time or a link, it becomes knowledge suggested to the Brain, the memory where your agents read your offers, your prices and your lead times. That knowledge only goes in after your "Keep" click. A typo teaches nothing, and a sentence specific to one customer does not become a rule.
You can also have it rewrite the text of an email with an instruction: "shorter, and offer the meeting link". The new version never goes out on its own, you reread it before approving. By ticking "Remember this instruction for its future answers", you keep it for the future, after your click. For a scheduled follow-up, the instruction given with "Redo it" applies right away to the follow-ups of all your agents.
If the agent answered on its own, fix what it read. On the Brain page, you add the right knowledge, which goes in directly since you are the one writing it, and you remove the one that was wrong, which is then archived, never deleted. For the site chat, the missing answer is added from the card of the request passed on to the team. If it is open to everyone, the chat can use it for the next question of that kind. For what must never be said, write a red line ("never a discount above 10%"): it goes at the top of each agent's instructions. While you check, put the agent in trial mode: everything it prepares then lands in "To approve" (see how much customer follow-up to hand to an AI).
With Muzau or another tool, the right correction depends on the cause.
Write the correction for the next question, not just this one. A piece of information fits in one sentence, which also says what makes it change: "The lead time is [usual lead time], longer for [case that extends it]." A tone instruction says what to do, not what to avoid: "Answer in three sentences at most" works better than "Don't be too long". Then watch the next request of the same kind: if the error comes back, the cause was elsewhere. To see what holds back an answer before it goes out, see whether an AI that writes to your customers can get it wrong.
Autonomous does not mean unchecked. You set the level of autonomy of every agent, its limits cannot be crossed, everything it does is recorded in a plain-language log, and one gesture stops it all.
For the tasks you don't have time for. You give the address of your website, then describe the work in your own words. Nothing to install.