--- title: "When to add humans back, and what you never automate" description: "The reverse question of module 5: the four situations that require a person, supervision as a job, and what you never automate." url: https://ofm-course.com/modules/scale/when-to-add-humans-back/ module: "Module 07, Scaling up" updated: 2026-09-14 lang: en --- # When to add humans back, and what you never automate ## In brief Four situations require a person to take over the conversation immediately: distress, a money dispute, a request outside the limits, and a fan who accounts for a large share of revenue. Supervision is a role, with a daily sample of five to ten conversations read in full. Never automated: age and consent, pricing decisions, the relationship with the creator, and the reply to a person in difficulty. ## What you will be able to do - Write your human takeover rules and make them applicable in one action. - Organize sample-based supervision that fits in a daily routine. - Know the short list of what is not automated, and why. ## The reverse question Module 5 explained why human conversation does not hold at scale. This lesson asks the reverse question, and it is just as important: **where is a human indispensable**. This is not a concession. An operation that automates without knowing how to answer that question is not saving money, it is transferring a risk it no longer sees. ## The four takeover triggers Four situations require a person to take over immediately. They are written in advance, they are few so that they get applied, and they must be doable in one or two actions. **1. Distress.** Someone who talks about severe loneliness, mental health, suicide, or whose messages change in nature. This is not a selling situation. A person takes over, stops the commercial sequence, and replies as a human being. It is the first rule on the list and it is not negotiable. **2. The money dispute.** A contested payment, content judged not to match what was announced, a refund request. A system can mishandle a dispute and turn a complaint into a public incident. A person decides, and documents it. **3. The out-of-bounds request.** Anything outside the limits written in the module 4 persona sheet: refused content, meeting requests, requests involving third parties, any request raising a doubt about anyone's age or consent. The takeover is immediate and the conversation stops if necessary. **4. The important person.** The fan who represents a significant share of a creator's revenue. This is not a question of risk but of return: these conversations deserve human time, and they are worth the cost. Two properties make these rules useful. They are **few**: four, not fifteen, otherwise nobody remembers them. And the takeover must be **immediate in practice**: if taking over requires five manipulations, the rule will not be applied when it matters. ## Supervision is a role The lesson [Roles, tool stack, bans and recovery](/modules/scale/roles-tools-and-bans/) listed four roles. In an automated operation, the conversation role becomes a supervision role, and it must be described precisely, otherwise it disappears. **The daily sample.** Five to ten conversations read in full, picked at random, including conversations that bought nothing. It is the only device that detects a drift before it shows in the numbers. The industry's job listings show, moreover, that this work exists as a paid position in its own right, in charge of auditing conversation logs. **Rereading new scripts.** A new script is reread in its first twenty real runs, not in its file. **Reviewing refusals.** The conversations where a tier was not opened teach more than those where it was. Nobody reads them, and that is a mistake. **Checking the persona.** Once a week, compare ten recent messages to the persona sheet. The register always slips, whoever is executing. Without these four actions, an automated operation works perfectly for several weeks, then drifts without any indicator moving fast enough to flag it. That is the risk specific to scale: **the gap between what you believe is happening and what is happening grows in silence**. ## What you never automate The list is short. It has to stay short to be respected. **Age and consent verification.** No system, no automatic procedure. A person verifies, a person keeps the documents, a person takes responsibility. It is a point of law before it is a point of method. **Pricing decisions.** A system executes a price ladder. It does not decide it, it does not adjust it on its own initiative, and it does not grant an unplanned discount. The lesson The price ladder, in module 5, explained what a price ladder that starts moving on its own costs. **The relationship with the creator.** The weekly check-ins, the calls on limits, the discussions about the contract. Remember that the platform's terms of service are explicit: the contractual relationship is with her, and the fact that a third party assists her "does not affect your legal responsibility". That relationship is the only thing holding the whole together. **The reply to a person in difficulty.** First trigger on the previous list. It is not a special case, it is a limit. ## The last word This course has a thesis, and it has been defended with facts rather than slogans. Handling sales conversations with human teams is expensive, degrades at night, copes poorly with languages and depends on work whose conditions the press has documented: less than two dollars an hour, eight-hour shifts, five days a week, long training, a real load. The operations that scale have moved the conversation to supervised systems. That move is documented on the Telegram side by official, dated features, and on the content platform side by a rule wording that has changed and now concerns direct integration, in the absence of a public interface. But the same thesis has a second half, and that is what separates a serious operation from one that thinks it is modern. A system executes; it does not decide. It sells; it does not judge. It keeps a register; it does not take responsibility. The four things on the list above remain human things, and they will stay that way. The method track is finished; what remains is [module 8](/modules/best-tools/), which reviews the tools category by category. Then go back to the module you missed the most, and reread the Sources and method page: that is where this course says what it knows, what it does not know, and on what date it checked it. ## Key takeaways - Four takeover triggers: distress, a money dispute, an out-of-bounds request, and the person who represents a large share of the revenue. - Supervision is a role, not a good intention. Without a daily sample, a drift takes weeks to show. - Never automated: age and consent verification, pricing decisions, the relationship with the creator, and the reply to a person in difficulty. - The more automated an operation is, the more the gap between what you believe and what is happening can grow without a signal. The sample is the only remedy. ## Do this now Write your takeover rules and test them. Thirty minutes, and it is the last exercise of the course. 1. Write your four human takeover triggers, one sentence each, with a concrete example of a message that would trigger them. 2. Check in your setup how many actions it takes to take over a conversation. If it is more than two, fix it today. 3. Set up the sample: five conversations per day, read in full, noted in one line in the journal. 4. Write your list of what is never automated, post it where you work, and reread it in three months. ## Sources - OnlyFans, Terms of Service (creator responsibility, relationship with her and not with a third party), https://onlyfans.com/terms, checked on 2026-09-14 - OnlyFans, help center, Can I use an AI chatbot on OnlyFans, https://onlyfans.com/help/196/205/213, checked on 2026-09-13 - BBC News, Icky and heartbreaking: the $2 per hour worker behind the OnlyFans boom, https://www.bbc.com/news/articles/cq571g9gd4lo, checked on 2026-09-14 - BBC News, OnlyFans agents control and threaten creators while taking half their earnings, https://www.bbc.com/news/articles/c932gpk2l1do, checked on 2026-09-14 - OnlyChat, help center, The AI section (customization, hours, pace, behavior limits), https://docs.only-chat.ai/ai, checked on 2026-09-14