---
title: "AI chatting explained"
description: "The six parts of an automated conversation system, the unlock ratio as the single metric, and what the platforms publish about AI in conversations."
url: https://ofm-course.com/modules/chatting/ai-chatting-explained/
module: "Module 05, Chatting"
updated: 2026-09-14
lang: en
---

# AI chatting explained

## In brief

An automated conversation system assembles six parts: persona, memory, scripts, timing, guardrails and human takeover. Its only sales metric is the unlock rate, the number of paid media opened per hundred sent. On Telegram, connecting a bot to a business account has been an official feature since March 31, 2024; on OnlyFans, the wording of the help center changed in 2026.

## What you will be able to do

- Describe the six parts of an automated conversation system and the role of each one.
- Choose and track a single selling metric, correctly defined.
- Cite what the platforms publish about AI in conversations, with the documented chronology.

## The six parts

An automated conversation system is not a text generator plugged into an inbox. It is an assembly of six parts, and the industry's experience shows that a disappointing system is almost always a system that is missing one of them.

<figure class="figure">
<svg viewBox="0 0 640 300" role="img" aria-label="Diagram of the six parts of an automated conversation system">
  <title>The six parts of a conversation system</title>
  <rect width="640" height="300" fill="var(--sand-soft)"/>
  <rect x="26" y="40" width="180" height="66" fill="none" stroke="var(--campus)" stroke-width="2"/>
  <text x="40" y="70" font-family="IBM Plex Sans, sans-serif" font-size="14" font-weight="600" fill="var(--ink)">1 . PERSONA</text>
  <text x="40" y="92" font-family="IBM Plex Sans, sans-serif" font-size="12" fill="var(--muted)">who speaks, and how</text>
  <rect x="230" y="40" width="180" height="66" fill="none" stroke="var(--campus)" stroke-width="2"/>
  <text x="244" y="70" font-family="IBM Plex Sans, sans-serif" font-size="14" font-weight="600" fill="var(--ink)">2 . MEMORY</text>
  <text x="244" y="92" font-family="IBM Plex Sans, sans-serif" font-size="12" fill="var(--muted)">what has already been said</text>
  <rect x="434" y="40" width="180" height="66" fill="none" stroke="var(--campus)" stroke-width="2"/>
  <text x="448" y="70" font-family="IBM Plex Sans, sans-serif" font-size="14" font-weight="600" fill="var(--ink)">3 . SCRIPTS</text>
  <text x="448" y="92" font-family="IBM Plex Sans, sans-serif" font-size="12" fill="var(--muted)">where the conversation goes</text>
  <rect x="26" y="136" width="180" height="66" fill="none" stroke="var(--campus)" stroke-width="2"/>
  <text x="40" y="166" font-family="IBM Plex Sans, sans-serif" font-size="14" font-weight="600" fill="var(--ink)">4 . TIMING</text>
  <text x="40" y="188" font-family="IBM Plex Sans, sans-serif" font-size="12" fill="var(--muted)">when, and at what pace</text>
  <rect x="230" y="136" width="180" height="66" fill="none" stroke="var(--campus)" stroke-width="2"/>
  <text x="244" y="166" font-family="IBM Plex Sans, sans-serif" font-size="14" font-weight="600" fill="var(--ink)">5 . GUARDRAILS</text>
  <text x="244" y="188" font-family="IBM Plex Sans, sans-serif" font-size="12" fill="var(--muted)">what is never said</text>
  <rect x="434" y="136" width="180" height="66" fill="var(--ochre)" opacity="0.22" stroke="var(--ochre)" stroke-width="2"/>
  <text x="448" y="166" font-family="IBM Plex Sans, sans-serif" font-size="14" font-weight="600" fill="var(--ink)">6 . HUMAN TAKEOVER</text>
  <text x="448" y="188" font-family="IBM Plex Sans, sans-serif" font-size="12" fill="var(--muted)">who takes over, and when</text>
  <line x1="26" y1="238" x2="614" y2="238" stroke="var(--line-strong)" stroke-width="2"/>
  <text x="26" y="268" font-family="IBM Plex Sans, sans-serif" font-size="13" font-weight="600" fill="var(--ochre)">SINGLE METRIC . UNLOCK RATIO</text>
  <text x="26" y="288" font-family="IBM Plex Sans, sans-serif" font-size="12" fill="var(--muted)">paid media opened per hundred sent</text>
</svg>
<figcaption>The six parts, and the one metric that judges them all. A system with no human takeover is not an incomplete system, it is a dangerous one.</figcaption>
</figure>

**1. The persona.** This is the document from module 4, turned into parameters: identity, displayed age, register, open and closed topics, forbidden phrases. Without it, the system writes correctly and sounds like nobody.

**2. Memory.** What the system retains from one day to the next in a conversation: first name, language, what was bought, what was promised, what was refused. This is the part that solves the memory break between shifts described in [Why human chatting breaks at scale](/modules/chatting/why-human-chatting-breaks/), and it is the one a beginner underestimates most.

**3. Scripts.** The selling sequence: where the conversation goes, which tiers, in what order, with which media. This is the two script lessons of this module, in full. A system with no script produces pleasant conversations that sell nothing.

**4. Timing.** The delays between messages, the persona's sleeping hours, the pauses. It serves two purposes: persona consistency, and the fact that an instant reply at any hour is not credible.

**5. Guardrails.** What must never be written: forbidden topics, impossible promises, unauthorized discounts, claims about real people. This part is a list, not an intention.

**6. Human takeover.** The moment a person takes over. It must be possible in one action, and the situations that trigger it must be written in advance. A system with no human takeover is not an autonomous system, it is a system with no brakes.

## A single metric

It is tempting to judge a conversation system on its style. That is a mistake: style is judged over ten messages, selling is judged over hundreds.

The metric is the **unlock ratio**: of a hundred paid media sent, how many are opened. Its definition matters as much as its value, and the two must be published together. One tool vendor, for example, publishes a ratio of 43% with its explicit definition, "of every 100 paid photos or videos the AI sends, 43 are unlocked", and the clarification that it is a "cumulative platform statistic since launch, not a single account and not a comparison with human chatters". That is the correct form of a performance claim: a number, a definition, a scope.

Three traps to know.

**An unlock ratio without a scope means nothing.** On which conversations, over what period, at what average price.

**A high ratio with a low price is not a performance.** Lowering prices pushes the ratio up and the revenue down. Always track the two together.

**An average hides variance.** True for human teams, true for systems. Look per creator, per language and per tier.

## What the platforms publish

This is the point where the industry tells the most false stories, in both directions. Here are the published facts.

**On OnlyFans.** The question is handled in a help center article, not in the terms of service. An investigation published by an industry vendor, which documents its method and its archive captures, establishes the following chronology: in January 2026 the help center article displayed the wording "No. You cannot use an AI chatbot to write chats or direct messages on OnlyFans", and it has since displayed the wording "OnlyFans does not have a public facing API and Creators may not directly integrate AI chat bots into the OnlyFans platform", recorded on September 13, 2026. The same investigation says it checked the terms of service, the acceptable use policy and the fan and creator agreement without finding any clause on chatbots, and recalls that a news agency dispatch from July 2024 quoted the prohibition while attributing it to the terms, when it actually appeared in a help center answer.

This course therefore concludes neither "it is allowed" nor "it is forbidden". It notes that the published wording has changed, that it is narrower than before, and that it concerns direct integration in the absence of a public interface. What you do with that is up to you, and the responsibility remains the creator's, since the terms of service state that the platform's relationship is with her and not with a third party.

**On Telegram.** There is no ambiguity: connecting a bot to a business account so that it can "process and answer messages on their behalf" is a feature officially announced on March 31, 2024 and documented in the API since version 7.2 of the same day. Sending paid media inside the conversation has been documented since version 7.6 of July 1, 2024.

This asymmetry alone explains why operations that automate almost always start with Telegram.

## What a system does badly

To finish, honesty about the limits, because no lesson in this course sells a tool.

A system **does not fix a bad script**. It runs it faster.

A system **does not replace a vague persona**. It exposes the vagueness.

A system **does not judge a sensitive situation**. That is precisely the role of part number 6.

A system **does not produce catalog**. If the tiers of the price ladder are empty, it has nothing to sell.

The next lesson moves to practice: how you set this up on a creator, step by step, and what the tools on the market cost. For a broader comparison of conversation AIs, the site [chatting-ofm.com](https://chatting-ofm.com) is entirely devoted to chatting.

## Key takeaways

- Six parts: persona, memory, scripts, timing, guardrails, human takeover. Systems that disappoint are always missing one of them.
- The selling metric is the unlock ratio: the number of paid media opened per hundred sent. Any other reading is comfort.
- The OnlyFans help center published a blanket prohibition in January 2026, and has since published a narrower wording about direct integration, given the absence of a public API.
- On Telegram, connecting a bot to a business account is an official feature, announced on March 31, 2024 and documented in the API.

## Do this now

Audit your current conversation against the six parts, even if it is entirely human. Thirty minutes.

1. Write the six parts in a column. Next to each one, write who or what fills it today in your operation, or nobody.
2. For memory, open three conversations older than three days and check whether the context from day one is still being used. Note the result.
3. For human takeover, write the three situations in which a person must take over immediately. Three, not ten.
4. Open the help center page cited in the sources, read the wording in force on the day you do this exercise, and copy it with the date into your file.

## Sources

- OnlyFans, help center, Can I use an AI chatbot on OnlyFans, https://onlyfans.com/help/196/205/213, checked on 2026-09-13
- OnlyChat, investigation into the OnlyFans rule on chatbots and the terms of service (chronology and archive captures), https://only-chat.ai/blog/onlyfans-ai-chatbot-rule-terms-of-service, checked on 2026-09-14
- Telegram, Telegram Business (connecting bots to a business account, March 31, 2024), https://telegram.org/blog/telegram-business, checked on 2026-09-14
- Telegram, Bot API changelog (version 7.2 of March 31, 2024, version 7.6 of July 1, 2024), https://core.telegram.org/bots/api-changelog, checked on 2026-09-14
- OnlyChat, About page (definition of the unlock ratio), https://only-chat.ai/about, checked on 2026-09-14
- OnlyFans, Terms of Service, https://onlyfans.com/terms, checked on 2026-09-14
