How the direct message became a sales floor, and what the person on the other end deserves to know
Somewhere tonight, with odds approaching certainty, a woman is replying to an Instagram story. It is late. The creator she follows has posted something that landed close to the bone, and under it, the little text field is open: an invitation. She types something a little braver than what she posts publicly. She hits send.
Four seconds later, the reply arrives. It is warm. It refers to the exact story she responded to. It asks her a question about herself, an easy one, and when she answers, the next message acknowledges what she said before offering her something useful: a guide, a checklist, a link to a free class this Thursday. The conversation feels like the thing she quietly hoped for when she hit send. It feels like being seen by someone she admires.
The messages were written in March.
Not by a scammer. Not by an overseas click farm running a stolen account. The creator wrote them herself, or paid someone to write them in her voice, one time, for everyone. A piece of software watched for the reply, matched it to a trigger, waited four engineered seconds so the response would not feel robotic, and sent message one of a sequence that has run thousands of times. The woman at her kitchen table is inside a funnel. She was inside it before she finished typing.
Nobody in this exchange is a villain. That is what makes it worth understanding. The creator cannot answer four thousand messages a month and never could. The follower wanted connection and received something engineered to feel like it. Between them sits an industry that has grown enormous while somehow staying invisible, and a question that neither the industry nor the platforms have wanted to answer out loud: what does the person on the other end deserve to know?
The machinery, plainly
Here is how the machine works, stripped of its jargon.
A creator posts a video and says: comment the word READY and I will send it to you. Thousands comment. Software watches the comment stream, catches each one, and opens a private conversation with the commenter. Message one delivers the promised thing. Message two asks a question, usually multiple choice, and the answer quietly sorts the person into a category: beginner or advanced, curious or ready to buy. Message three responds to whichever button they pressed, often with a short testimonial matched to their answer. Somewhere around message four or five, having asked permission first, the sequence offers a link. A webinar registration. A sales page. A calendar.
Practitioners call this a DM funnel, and the reason you have never read a serious examination of one is that nearly everything written about them is written by people selling them. Search the topic and you will find vendor blogs, affiliate roundups of "the 12 best automation tools," and course creators teaching other course creators. The trade coverage is a hall of mirrors. The mainstream coverage barely exists.
The money is not imaginary, though. ManyChat, the dominant infrastructure company in the space, raised 140 million dollars in a round led by Summit Partners in April 2025, bringing its total funding above 160 million since its 2015 founding. The company reports roughly 1.5 million customers across 170 countries and claims billions of automated messages a year; its client list, by its own telling, runs from solo creators to Nike and The New York Times. Those scale figures come from the company and its investors, so hold them loosely. The funding round is a matter of public record, and so is the direction of travel: the company shipped an AI layer in 2024, moving the field's ambition from scripted button-trees toward software that composes its own replies.
One more mechanical fact completes the picture. Every answer a follower gives inside one of these conversations is stored as a tag on her contact record. The chat is also a database write. The woman who taps "just getting started" at midnight has not only received a friendlier message; she has been filed, durably, for every campaign that comes after this one.
So is what happened to the price of entry. In March 2026, ManyChat restructured its plans and cut its free tier from 1,000 contacts to 25, a change the industry press read as the end of the hobbyist era. Automated conversation is priced like infrastructure now, metered per human being on the other end. That detail matters more than it first appears, and we will come back to it.
Intimacy at a distance
The strange thing is that a pair of researchers described the emotional core of this machine in 1956, before anyone it would ever run on was born.
That year, in the journal Psychiatry, Donald Horton and R. Richard Wohl published a paper called "Mass Communication and Para-Social Interaction: Observations on Intimacy at a Distance." Watching the new mass media of their day, they observed that radio and television gave audiences "the illusion of face-to-face relationship with the performer." Listeners came to feel they knew the host who spoke to them each night, felt entitled to him, grieved him when he left the air. "The most remote and illustrious men," the authors wrote, "are met as if they were in the circle of one's peers." The relationship ran in one direction only, but it did not feel that way from the couch. Horton and Wohl named it para-social interaction, and the term outlived them into every corner of internet culture.
They also noticed something their successors keep rediscovering: the audience deserved more sympathy than scorn. The loneliness was real, the appetite for relationship was real, and the medium offered a shape that fit the appetite. Nothing about that observation has aged.
For fifty years, the distance in "intimacy at a distance" was load-bearing. You could love a broadcaster; you could not reach him. Social media appeared to demolish that wall. The person you follow is right there. The same inbox that holds messages from your sister holds a channel to her. You can just write.
Then came the door within the door. On June 2, 2021, at its F8 Refresh developer conference, Meta announced it was opening the Messenger API for Instagram to all businesses, ending a closed beta that had begun the previous October with roughly 30 developers and 700 brands. The rollout moved in phases, mid-sized accounts first. In plain language: software could now read and answer a professional account's direct messages at scale, with Meta's blessing, through an official pipe built for exactly that purpose.
It is hard to overstate what that decision did to the meaning of the inbox. The DM had been the one space on the platform that was presumptively human, the room behind the stage where the performance stopped. After June 2021, that room acquired a service entrance. The intimacy stayed. The distance came back, but now it was hidden inside the intimacy, and there was no way to see it from the couch.
The drowning inbox
Before this becomes a story about deceitful creators, sit for a moment inside their inbox.
A creator with half a million followers who posts a story with a question sticker can wake up to thousands of replies. Each one is a person. Many are kind. Some are lonely in ways that deserve more than a heart emoji, and a few are in genuine crisis. The follower sends one message and remembers it; the creator receives four thousand and cannot. There is no version of this arithmetic where a human answers everyone, and the platforms built the expectation of reply into the product anyway.
Automation, in this light, is triage, and the honest practitioners talk about it that way. The machine catches the flood, delivers the promised guide instantly instead of nine days later, answers the question that four hundred people asked this week, and flags the conversations that need an actual person. Small businesses ran on exactly this logic when they installed phone trees. Nobody felt betrayed by "press two for billing," because nobody mistook it for the owner picking up.
But the phone tree announced itself. The DM sequence is engineered not to. The four-second delay before the first reply exists to simulate a human pause. Sequences are written in the creator's voice, contractions and inside jokes included, because the entire craft discipline of the field, taught in its courses and repeated in its style guides, is that automated messages must not read as automated. The trade's own vocabulary gives the game away: the goal, practitioners say, is conversation that does not feel like a bot.
And the content upstream bends to feed the machine. Watch for it once and you will see it everywhere: captions that end in "comment the word GROW," videos structured to manufacture the trigger, engagement that used to be an open question and is now a routing instruction. The comment section, too, has become an intake form that does not look like one.
The triage framing also has a failure mode the industry's how-to materials rarely sit with. Some fraction of those four thousand messages come from people in real trouble, and a sequence built to qualify buyers does not know the difference between an objection and a cry for help. A funnel that answers "I've tried everything" with a payment-plan link is a tolerable embarrassment when the topic is software; when the account belongs to a fitness coach, a faith leader, or anyone whose audience discloses pain, it is a design failure with a person attached. The better operators build a hard stop: certain words halt the automation and summon a human, no pitch, no delay. Nothing in the platforms' rules requires that. It exists only where someone decided the follower's worst night mattered more than the sequence's completion rate.
The rules of the room
The conversation that follows is shaped by rules almost nobody on the receiving end has read, published in Meta's developer documentation for anyone who cares to look.
The central rule is the 24-hour window. Software may message a user only after the user messages first, and once the user does, the account has 24 hours to respond with automation; each new interaction from the user resets the clock. Past the automated window, Meta's documentation permits a further period in which a human agent, and only a human agent, may reply by hand: the platform itself draws a line between machine speech and human speech, even if the recipient never sees it drawn. The rule exists as an anti-spam measure, and as policy it is defensible. What is striking is what the industry has built on top of it: an entire craft of engineered re-permission. Sequences are written so that the lead keeps voluntarily reopening the window. The reminder message ends with "reply YES so I can save your workbook," and the yes does quiet work: a timestamp reset, dressed as warmth.
Pricing shapes the room too. Automation platforms bill by active contact, which means every person who wanders into the machine is a line item, and the economics reward keeping her tagged and sorted. The multiple-choice questions that feel like getting-to-know-you are segmentation. The answer a follower taps at midnight determines which testimonial she sees, which offer she gets, and which list she lives on for the next promotion. None of this is secret. It is documented, taught, and priced. The only person who has not read the documentation is the person the message is for.
None of these mechanics is a scandal on its own. Segmentation is as old as the mailing list, and a well-built sequence often serves the follower better than the silence she would otherwise get. The trouble is narrower and harder to wave away. Every design choice in the room, from the four-second delay to the voice-matched copy, points in one direction: toward the follower believing, or at least never quite doubting, that a person is present. The industry did not stumble into that ambiguity. It optimized for it.
The law that already exists
Here is the part that should startle the industry more than it seems to: the disclosure question has already been answered by law, in the largest state in the country, and the answer has been sitting there since before the machinery went mainstream.
California enacted Senate Bill 1001, the Bolstering Online Transparency Act, in 2018; it took effect July 1, 2019. The statute makes it unlawful to use a bot to communicate with a person in California online "with the intent to mislead the other person about its artificial identity for the purpose of knowingly deceiving the person about the content of the communication in order to incentivize a purchase or sale of goods or services in a commercial transaction." The wording is narrower than a blanket labeling mandate; it aims at deception with commercial intent, which is precisely why it belongs in this conversation. The law also includes a safe harbor, and the safe harbor is almost embarrassingly easy to reach: disclose. The disclosure need only be "clear, conspicuous, and reasonably designed to inform" the person that they are talking to a bot.
Utah followed with its Artificial Intelligence Policy Act, enacted in March 2024 and effective that May: the first state law built squarely around AI and consumer protection. Under it, businesses using generative AI with consumers must disclose the fact clearly and conspicuously when a consumer asks, and licensed professionals must disclose up front. Utah has since refined the act toward higher-risk interactions, but the core stands. Ask a Utah business whether you are talking to a machine and the law expects a straight answer.
Set those statutes next to the industry's daily practice and the dissonance hums. A sequence that opens a commercial conversation in a creator's voice, delays its replies to feel human, and walks a follower to a checkout without ever hinting at its artificial identity is, at minimum, the exact behavior the California legislature sat down to address. Whether any given funnel crosses the legal line involves intent, jurisdiction, and facts, and enforcement so far has been thin. But the plain words are the plain words, and the how-to content of an entire industry, courses, templates, certification programs, manages to be encyclopedic about trigger words and messaging windows while staying nearly silent about disclosure law. I went looking for the chapter on it. The chapter is not there.
Run the Utah test as a thought experiment and the gap gets uncomfortably specific. Suppose the woman at the kitchen table simply types: are you a bot? The law of one state expects a clear and conspicuous answer. The standard funnel has no branch for the question. It was scripted for objections about price and time, not for the one question that touches its identity, so the machinery does what it does with anything it cannot parse: it routes her to the default reply, which cheerfully offers the guide again. The sequence is not lying, exactly. It was built so that the truth has nowhere to land.
The label and the room
The strange twist, after all this, is that the machinery has little to fear from honesty.
The follower at the kitchen table did not actually require the illusion. What she wanted survives the label: the guide delivered in seconds, the answer to the question four hundred people asked, a working doorway to a real human when her situation stops fitting the buttons. A first message that says "automated reply, from words I actually wrote, a real person sees this thread if you need one" costs the creator a single line. California's safe harbor asks for hardly more than that. The businesses that already operate this way have discovered what the phone tree discovered decades ago: people forgive machinery that introduces itself.
What does not survive honesty is the pretense, and the pretense was always the cheapest part of the build. Horton and Wohl saw the shape of this seventy years ago. Intimacy at a distance is not a con by nature; it is a real form of human attachment, and it can be handled with care or harvested. The DM funnel is simply the first version of the machine where the distance itself can be disguised, where the couch and the stage feel like the same room and only one side knows the difference.
The woman at 11:47 will probably buy the course. It might genuinely help her. Her creator is nobody's villain here, and the arithmetic problem the industry solves for that creator is real. All of it can be true at once, and one more thing can be true besides: she deserves to know which kind of seen she is. The message that finally tells her will take one line to write. Someone in this industry is going to write it first and discover it converts.
Reyes Castillo is a direct response copywriter who builds the kind of messaging systems this piece examines, and who believes they work better honest. If your business needs writing that respects the reader this much, he takes clients: rcastj3@gmail.com.
Sources
1. California SB 1001 (Bolstering Online Transparency Act), Bus. & Prof. Code §17940-17943, operative July 1, 2019 - leginfo.legislature.ca.gov. VERIFIED VERBATIM 2026-07-12: prohibition and safe-harbor quotations match statutory text exactly.
2. Utah Artificial Intelligence Policy Act (SB 149), enacted March 13, 2024, effective May 1, 2024; 2025 amendments SB 226/SB 332 - analyses: Skadden (Apr 2024), Davis Polk (2025), Mayer Brown (May 2024)
3. "Facebook opens its Messenger API for Instagram to all," TechCrunch, June 2, 2021 (closed beta from Oct 2020, ~30 developers/700 brands; phased rollout)
4. Meta for Developers, Instagram Messaging API documentation. VERIFIED VERBATIM 2026-07-12: "Only after an Instagram user has sent your app user's Instagram professional account a message can your app send a message"; "Your app has 24 hours to respond"; human-agent tag permits response "outside the 24 hour messaging window" (no specific extended duration stated in the primary).
5. "Manychat taps $140M to boost its business messaging platform with AI," TechCrunch, April 22, 2025; Summit Partners announcement, PRNewswire (scale figures company-reported and flagged as such in text)
6. ManyChat March 2026 pricing restructure (free tier 1,000 → 25 contacts) - Voiceflow analysis (July 5, 2026); SetSmart review (March 30, 2026)
7. Horton, D. & Wohl, R.R., "Mass Communication and Para-Social Interaction: Observations on Intimacy at a Distance," Psychiatry 19(3) (1956), 215-229, DOI 10.1080/00332747.1956.11023049. VERIFIED VERBATIM 2026-07-12 against the archive scan: both quotations exact.