# Knowledge, Copilot, Autopilot, and Human in the Loop | Mihu AI Launch Week · Day 4 · LOOP by mihu # Knowledge, Copilot, Autopilot, and Human in the Loop Where the answer comes from, who gets to send it, and how the human-in-the-loop process moves a conversation between an AI agent and a team member — in both directions, in one click, with the reason written down. Mihu AI Team September 3, 2026 10 min read v1.0.0 Launch Week · Day 4 Knowledge · Retrieval · Draft · Handover Your knowledge The draft Who sends it The same retrieval every time. What changes is whether the AI agent sends it, or your team member does. In this article 1 The human-in-the-loop process (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#approval) 2 Where the answer comes from (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#knowledge) 3 Retrieval you can see in the thread (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#retrieval) 4 Autopilot: the AI agent sends (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#autopilot) 5 Copilot: it drafts, your team sends (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#copilot) 6 Say next: the line in front of you on a live call (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#saynext) 7 Take over, hand back (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#handover) 8 When it does not know (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#gaps) 9 Tuning the match (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#tuning) 10 Frequently asked questions (https://mihu.ai/nl/blog/launch-week-day-4-03-09-2026#faq) LOOP is the Mihu Contact Center: a cloud call center and an omnichannel inbox where your team and your AI agents answer the same calls and the same messages. This post is about what happens inside a single reply. The AI agent searches your Knowledge Library , writes what it found into the thread, and then either sends the answer itself or hands it to a team member to send. On a live call it says nothing and puts the next line on your screen instead. And either side can take the conversation from the other, in one click, with the reason written down. ## The Human-in-the-Loop Process Nothing you care about happens without a person saying yes. Human-in-the-loop is not a mode you switch on. It is a gate the AI agent passes through on every reply and every action, and you decide where that gate sits. The whole process is five steps, and each one of them leaves a line in the conversation. It retrieves, then it proposes. The AI agent searches your Knowledge Library and writes the reply. On a Copilot queue that reply arrives on a team member’s screen as an AI draft , not in the customer’s inbox. Actions come with a confidence score. Beyond the reply, it proposes what to do — book this slot, update this field, open this ticket — each with a number next to it and an Approve or Reject . You draw the line, field by field. A corrected address can write itself. Anything touching money, a commitment or a field you marked sensitive waits for the tap. This is a setting in your workspace, not a promise in a brochure. Below the threshold, it escalates instead of guessing. A knowledge gap, an SLA about to break, sentiment falling, the customer asking for a person, a message outside working hours, the queue busy — any of these routes the conversation to a named team member, with the reason attached. A person can step in at any point, and step back out. Take over conversation silences the AI agent mid-thread or mid-sentence; handing it back resumes it with everything said in between. And the loop runs the other way too: a team member can hand work to an AI agent and approve what comes back. What is about to happen Who decides A reply on an Autopilot queue The AI agent sends it A reply on a Copilot queue A team member presses Send An action on a field you marked sensitive A team member taps Approve An action on a field you did not It goes through, and is logged Confidence below your threshold It escalates to a person Every decision in that table is written into the thread with who made it and when — the approval, the rejection, the takeover, the hand-back. So the question a regulator, an auditor or your own head of support asks — who let this go out? — has an answer that nobody has to reconstruct from memory. ## Where the Answer Comes From An AI agent is only as good as what you let it read. A model out of the box knows the internet. It does not know your stock, your prices, your return window, or which of your three branches does bodywork. That gap is the whole distance between a demo that impresses you and a Tuesday that does not embarrass you. The Knowledge Library is where you close it. It sits in the inbox sidebar, under your queues, and it is the same library your AI agents read from — not a copy, not an export. Two kinds of thing live in it. Documents — FAQs, product documentation, policies, procedures. PDF, DOCX, CSV or TXT, up to 50 MB a file. You drop them in, they are processed once, and after that they are simply known. Tables — your structured data, as rows. Stock, prices, availability, specifications, opening hours. Open All tables , pick one, and you get its record count, when it last changed, and a search box across every row in it. Knowledge Library · Company knowledge base ‹ All tables Stock Data 16,851 records Updated 2 Sep 🔍 estate 2.0 diesel Match Brand Model Updated 0.86 Volvo V60 D4 2026-09-01 0.81 Volvo V90 D5 2026-09-01 0.64 Volvo XC60 B4 2026-08-28 0.31 Volvo XC90 Recharge 2026-08-28 Sixteen thousand rows, and a score on every one of them. That score is the whole product. Because the search box is not a text search. Type a phrase and every row comes back scored — 0.86, 0.64, 0.31 — and that number is the same match your AI agent computes in the middle of a conversation. What you see in the library is what the agent sees on the call. So a question you know your customers ask, typed into that box, tells you today what the answer will be tomorrow. If it comes back at 0.31, you have just found next week’s bad answer before a customer did. ## Retrieval You Can See in the Thread Every answer says where it came from, in the conversation itself. In most tools, an AI answer grounded in your documents is a black box with a nice sentence coming out of it. You find out it was wrong when a customer tells you, and then you get to guess why. In LOOP the retrieval is a line in the thread, sitting right above the message it produced: Inbox · Support queue · WhatsApp 📖 Knowledge base: 5 records retrieved (top match 73%) 19:58 ✦ AI draft Yes — that model is in stock in the estate body with the 2.0 diesel, and it is at the Austin branch. Would you like me to hold one for a viewing this week? ↻ Regen ✎ Edit ➤ Send Five records, top match 73%, and the reply they produced. Nobody has to take the answer on trust. Anyone who reads that conversation afterwards — a supervisor, a QA reviewer, the person who takes it over — can see whether the agent had something to answer with or was working from thin air. When a reply was wrong, that one line tells you which of two very different problems you have: the knowledge was missing , or the knowledge was there and the reply was poor . Most teams cannot tell those apart, and so they fix the wrong one for months. You can also ask it yourself, mid-conversation. Ask the knowledge base is a question box over the same library, and the answer comes back with the source attached — the document, the article, the row. A new starter on their second day has the same recall as the person who wrote the policy. ## Autopilot: The AI Agent Sends It retrieves, it writes, it sends — and all three are on the record. You met the three modes on Day 3: Autopilot , Copilot and Human First , set on a queue, on a channel, or on one conversation. Today is what the first two actually do in the seconds between your knowledge and your customer. In Autopilot the AI agent reads the thread, retrieves, decides and sends. Nobody presses anything, and three things are true every time: It retrieves before it writes. The reply is grounded in your library, and the retrieval line goes into the thread whether the answer turns out well or badly. It acts inside the limits you set. Booking against real availability, updating the contact record, firing the workflow — with anything sensitive still waiting for a person’s tap. It hands over the moment your rules say so , and the reason is written into the thread rather than inferred from it later. Autopilot is not “the AI handles support now”. It is the questions that repeat, answered at three in the morning, from a document you approved. ## Copilot: It Drafts, Your Team Sends The keystrokes disappear. The judgement stays exactly where it was. Copilot is the same machine, one step back. The AI agent still reads the thread, still retrieves the same records at the same scores, still writes the whole reply. It just stops at your team member’s screen, marked AI draft , with three buttons under it. Button What it does What it tells you Send Goes out as written The common case. Count these — it is the work your team no longer types. Edit Change a line, then send What they changed is what your library was missing. Read the edits, not the drafts. Regen Another take, same facts The tone was off, not the answer. A different problem, and a smaller one. Nothing leaves in Copilot without a person pressing something. That is the whole promise, and it is a setting rather than a claim: anything touching money, a commitment or a sensitive field waits for the tap, and every decision is logged with who made it. Beyond replies, the AI agent proposes actions the same way — book this slot, update this field — each with a confidence score and an Approve or Reject beside it. And in Human First , where the AI agent stays silent, the same drafting is there the moment you want it. Draft with Mihu AI sits above the composer, reads the conversation and writes the reply, with a tone to choose — Empathetic, Professional, Friendly, or the agent’s own suggestion — then Insert , Regenerate , Dismiss or Approve & send . It needs no configured AI agent at all, so a team that has not built one yet still gets this on their first day. Try this: put one queue on Copilot for a week, then pull the drafts your team edited and read only the edits. That list is your next week of knowledge base work, written for you by the people who do the job. ## Say Next: The Line in Front of You on a Live Call The AI agent stops talking and starts prompting. Copilot on a live call is the part people do not expect. The AI agent is on the line. It hears both sides. It never speaks. What it does instead is turn the call into a transcript as it happens and run the same retrieval it would run if it were answering — then put the result on your screen while the customer is still talking. Somebody asks where their order is, and the order is in front of you before they finish the sentence. They ask what the refund policy actually says, and the clause is there with the article it came from. It arrives as a Say next card. ### No hold “Let me just check that for you” is a sentence your team stops saying, along with the forty seconds after it. ### No second screen The answer arrives where they are already looking. No tab, no search box, no CTRL-F through a PDF. ### New starters sound experienced The recall belongs to the library. The judgement, the tone and the decision stay theirs. ### Same engine on text On WhatsApp, email or chat the same retrieval lands in the draft above the composer instead. One mechanism, two surfaces. The customer is talking to a person the entire time. That person simply already has the answer. ## Take Over, Hand Back One button, both directions, and the reason written into the thread. The handover is not a setting you go and change. It is a button on the conversation, and it works both ways. Take over conversation sits under the thread. Press it and the AI agent goes quiet — mid-sentence on a call, mid-thread on messaging — and the conversation is yours. Nothing is re-sent and nothing restarts. On messaging the thread simply continues; on voice you are on the same live call, so the customer hears one conversation, not a hold tone. Give it back the same way. On messaging the mode returns to Autopilot; on a live call the button reads Hand back to AI agent , and the agent resumes carrying everything that was said while you had it. And it is not only the AI agent that gets relieved of work. A team member can hand a task to an AI agent — call this customer back, find out what they actually need, chase the document that never arrived — with a time on it if it is not for now. The desk has a Tasks switch beside Conversations for exactly this. The agent makes the call, and what it learns is on the record before anyone opens the thread. Same gate on the way back: what it did is a line in the conversation, and anything you marked sensitive still waits for your tap. One thread · the whole audit trail ⇆ AI → Human — Rachel took over ⇆ AI → Human · escalated to Rachel ⇆ AI → Co-pilot — AI drafts, Rachel sends ⇆ Human → AI — autopilot resumed by Rachel Four lines in the conversation. Nobody has to reconstruct how it got where it got. What you see in the thread What happened Started by AI → Human — took over A team member pressed Take over conversation Team AI → Human · escalated A rule fired and routed the thread to a named person Your rules AI → Co-pilot Same conversation, one step back: the AI drafts, a person sends Team Human → AI — autopilot resumed Handed back to the AI agent, which picks up with everything said since Team The ones that say escalated did not come from a button at all. They came from rules you wrote: a knowledge gap, an SLA about to break, sentiment falling, the customer asking for a person, a message outside working hours, the queue busy. The trigger is in the line, so the argument about whether the AI agent should have handled it is an argument with evidence in it. ## When It Does Not Know The useful failure is the one that gets logged. An AI agent that answers everything confidently is worse than one that stops. When the match falls below your threshold, it does not improvise. It flags a knowledge gap and hands the conversation to a person, with the question attached. Those gaps do not go into a log nobody opens. They are ranked by how often they happen , so the top of the list is the thing your customers ask for most and your library does not have. Write it once and the whole floor stops being asked it — by the AI agent and by your team, because they are reading the same library. Every conversation also carries an AI Analysis : sentiment, emotion, intent, knowledge gap, whether a team member was needed, satisfaction, resolution outcome, duration and language. The same fields for AI conversations and human ones — which is the only way to compare them without flattering either. ## Tuning the Match Three settings and a test that tells you the truth. Strictness — Relaxed , Balanced or Strict . Relaxed answers more and risks a loose match; Strict says “let me get someone” sooner. Set it by what the queue is for. A question about a specification and a question about a warranty do not deserve the same threshold. Test this agent — type a question and it runs the exact retrieval it would run with a customer. Not an approximation of it: the same records, the same scores, the same answer. The library is live — fix a document or a row and the next conversation uses it. No retraining, no redeploy, no window where the old answer is still going out. Try this: take last week’s five worst conversations and look only at the retrieval line on each. Missing knowledge, or a weak reply from good knowledge? They will not all be the same, and only one of the two is fixed by writing something. ## Frequently Asked Questions Where does my AI agent get its answers? From your Knowledge Library — the documents you upload (FAQs, product documentation, policies, procedures) and the tables of structured data you keep there. It retrieves from that library before it writes, and the retrieval is recorded in the conversation. Can I see what it retrieved for a particular reply? Yes. A line in the thread says how many records came back and what the top match scored — for example, five records retrieved, top match 73%. It sits above the message it produced, so a supervisor reading afterwards can tell whether the agent had something to work with. What is the difference between Autopilot and Copilot? Who presses send. In Autopilot the AI agent writes the reply and sends it itself. In Copilot it writes exactly the same reply, and your team member sends it, edits it first, or regenerates it. The retrieval underneath is identical. Can a team member take a conversation from the AI agent, and give it back? Yes, in both directions and in one click. Take over conversation silences the AI agent mid-thread or mid-sentence; on a live call, Hand back to AI agent returns it with everything said since. Each change is written into the thread with who did it. What happens when the AI agent does not know the answer? It stops rather than improvises. Below your confidence threshold it flags a knowledge gap and hands the conversation to a person with the question attached. The gaps are then ranked by how often they occur, so the most expensive missing answer is at the top of the list. Does the AI agent speak while my team member is on the call? No. In Copilot on a live call it never speaks. It listens to both sides, retrieves, and puts a Say next card on the team member’s screen while the customer is still talking. What can I put in the knowledge base, and in what format? FAQs, product documentation, policies and procedures as PDF, DOCX, CSV or TXT files up to 50 MB each, plus tables of structured data such as stock, prices and availability that you can browse and search row by row. Do I have to retrain the agent after I change a document? No. The library is live: fix a document or a row and the next conversation uses it. There is no retraining step and no window in which the old answer is still going out. Does Copilot require a configured AI agent? For the automatic drafts on a Copilot queue, yes. But Draft with Mihu AI above the composer works without one, so a team that has not built an agent yet still gets a written reply, a tone to pick, and Insert, Regenerate, Dismiss or Approve & send. ## Put one queue on Copilot today Register on the free plan, drop your FAQ and your price list into the Knowledge Library, and let the AI agent draft while your team keeps the send button. Seats are free, and you decide what it is allowed to answer on its own. Register free → (https://app.mihu.ai/register) Launch Week LOOP Knowledge Base Copilot Autopilot Handover Human in the Loop ## Keep exploring Launch Week More ways to run your contact center with Mihu AI 💬 Launch Week · Day 3 #### LOOP by mihu: The Contact Center Where AI Agents Are Colleagues Sep 2026 (https://mihu.ai/nl/blog/launch-week-day-3-02-09-2026) 👥 Launch Week · Day 2 #### Team, Roles, Logs, Queues, and Smart Routing Sep 2026 (https://mihu.ai/nl/blog/launch-week-day-2-01-09-2026) 🤖 Product #### Mihu Contact Center Product page (https://mihu.ai/nl/mihu-contact-center)