> ## Documentation Index > Fetch the complete documentation index at: https://docs.kaisupport.com/llms.txt > Use this file to discover all available pages before exploring further. # Best practices > Ten rules for the first weeks with Kai, and the reason behind each one. This page is opinionated on purpose. Each rule is short, and the paragraph under it says why. Read it in your first week, then again after your first retrain. ## Start in copilot, and stay there Run every channel in copilot until draft acceptance is steady. A draft costs you a read. A wrong autopilot reply costs you a customer, and you find out about it later. Copilot also produces the only evidence that autopilot is safe. Every draft your team sends unchanged is a vote that the answer behind it works. Watch draft acceptance on the Overview page for two or three weeks. When it holds steady and high, you have a number, not a feeling. ## Review the blockers first, not the longest list The review queue marks some items **Blocks publishing**. Start there. The answer library is usually the longest category, and it is rarely the one stopping you. Publishing is blocked only by what could reach a customer with nobody reading it: your brand voice, every human handoff rule, every conversation flow that autopilot may use, and the answers those flows use. Everything else is allowed into a published Brain, and it drafts instead. Reviewing the tail from real traffic is faster than reviewing it from a list, and it is better ordered, because the questions customers actually ask come up first. A review queue row badged Blocks publishing ## Write answers that hold no numbers Keep a price, a fee, a date and a discount out of the answer library. Put them in current information, and let the answer point at them. Kai checks every number in a reply against current information, a non-volatile answer, or the conversation itself. A number it cannot verify downgrades the reply to a draft. An answer with a price baked into its text therefore drafts forever, and worse, it goes stale the day the price changes and nobody remembers which entry held it. Current information is not part of a Brain version: change a price there and the next reply uses it, with no review and no publish. ## Keep conversation flows short Three or four steps is a flow. Twelve steps is a script, and a customer will leave it at step two. A flow tells Kai what to do in a situation, not what to say word by word. The brand voice supplies the wording, and the answer library supplies the facts. A long flow mostly repeats those two, and every repeated sentence is one more thing to keep in agreement when a price or a policy changes. ## Write handoff rules for the expensive mistakes List the cases where being wrong costs you money, trust or a legal problem. Refunds, complaints, anything a lawyer would read, anything a parent is angry about. Those are human handoff rules. Kai stops on a handoff rule whatever else the policy says, and whatever its confidence is. That makes the rule the cheapest control you own: one sentence that a hundred edge cases cannot get past. Write the rules broadly. A handoff that was not needed costs one teammate one minute. ## Rate the drafts Rate what Kai already said, in the console's Review mode. Mark a reply good, or write the correction you would have sent. A rating is the only thing that lets a live reply teach the next Brain. Kai never learns from its own unreviewed output: an AI reply enters the corpus only when a reviewer rated it good or wrote a correction, and the correction replaces the words Kai used. An unrated reply is discarded at retrain time. If your team rates nothing, a retrain learns nothing from the months you ran. ## Move one source to autopilot, not a whole channel When you do switch on autopilot, switch on one source. One Intercom workspace carries the Messenger widget, WhatsApp, email and SMS on one connection, and those are not one audience. A widget visitor is often anonymous and asks anything. A WhatsApp thread is a phone number that already chose to write to you, usually about one thing. Start with the narrower audience, watch the activity feed for a week, then widen. A per-source mode is one control on the channel card, so this costs you nothing. ## Retrain on a schedule, not after a complaint Put a retrain in the calendar: monthly, or after a season ends. Do not run one because a reply went wrong yesterday. One bad reply is a review job, not a retrain. Edit the answer, approve it, publish. A retrain rebuilds the whole Brain, and it needs the reviews to be in before it runs, or it carries the same gaps forward. Scheduled retrains also cost less: labels are cached per conversation, so a retrain over a mostly unchanged history pays only for what is new. ## Never approve an answer you would not send Approve means "a customer may read this". If you would rewrite it before sending, edit it now, or reject it. Approving is what lets an answer reach a customer with nobody watching, under the **Anything a reviewer approved** scope. An approval given to clear a queue is not a review, and it removes the one gate that stands between a draft answer and a live one. Rejecting is cheap. The topic stays known, and Kai escalates it instead of guessing. ## Give the account owner to the person who knows the customer Set the account owner to the teammate who has actually spoken to that customer, not to whoever is free. The owner is how the next conversation from that customer reaches the same person, so the customer does not explain everything again. Kai assigns a new conversation to the owner only while it is unassigned, so ownership never takes a thread away from the teammate a manager handed it to. ## Next The modes, the two autopilot scopes, and per-source settings. Every term on one page.