Emulating Humans Means Engineering the Whole Conversation

Eight hours of chat for a $5,000 month

In January 2023, Jesse Silver saw a friend earning about $5,000 a month on OnlyFans spend eight to ten hours a day messaging fans individually. Interviews with her and other creators revealed an automation opportunity: conversation was exhausting work that drove sales. His company subsequently worked with other fan platforms, where subscriptions provide access to photos and videos alongside one-to-one messaging.

Silver estimated that chatting generates roughly 80% of revenue on a well-run profile through relationship building and individual content sales. He viewed creators as underserved small businesses: doing the work themselves consumed their time, managing offshore teams added operational burdens, and agencies could charge substantial fees. His product focused on relationship building, media sales, and custom-content negotiation; agencies also handled content and traffic acquisition.

The founders recruited initial profiles by pitching a prototype at AVN, then relaunched a more developed product in December. By the May 2024 interview, Silver reported having worked with more than 150 creators and exchanging messages with about 50,000 users daily.

What experience is the fan buying?

Silver expected some competition from standalone AI companions, but considered them different experiences from interacting with an existing creator. His team used three working hypotheses about fan motivations: - Companionship: someone to talk to, potentially producing daily engagement and conversations lasting thousands of turns. - Convenient access to content: help finding the desired media. - Fantasy with stakes: winning someone over or navigating a power dynamic, with a real person understood to be on the other end. This was a working model of users, not an established psychological classification. Silver considered the third category especially productive commercially.

That model gives retention a particular meaning. Silver described fans seeking fresh experiences with objectives and an eventual sense of completion. He estimated monthly subscription churn of 50–80% in the market served, while roughly 10% of fans remained for more than a year; those estimates came without cohort definitions or a measurement method. Long conversations were not themselves the objective: the product sought creator earnings through content sales or monetized conversation.

Whether fans knew they were interacting with AI remained unresolved. Silver characterized their understanding as somewhere in between, in an industry already using agency-operated chat teams, without specifying a disclosure policy. The product’s stated goal was to preserve suspension of disbelief while providing more interaction, maintaining the creator’s brand, and increasing income.

A friendly bot can break the product

Emulating a creator requires several kinds of consistency: - Voice and facts: match the creator’s tone and answer personal questions without inventing details. - Boundaries: respect limits around information such as location. - Media selection: choose appropriate content from inventories containing tens of thousands of assets. Together, these still do not guarantee coherent behavior over time. Selling one item is a different problem from selling 40 over several months; repeated sales require a sustained interaction strategy.

The business charged 20% of sales made by the product, excluding existing subscription revenue and sales it did not make. The reported earnings uplift was 2–5×, attributed largely to expanded audience coverage: human representatives could sustain only a few concurrent conversations, leaving many fans unserved. No evaluation method or comparison period accompanied the uplift claim.

An early deployment exposed the limits of generic friendliness. On a profile already earning more than $150,000 monthly, the bot reportedly made only a few hundred dollars over two days, against an expectation of at least $5,000 per day. The creator’s established persona was an inaccessible diva. By immediately expressing enthusiasm about getting acquainted, the bot undermined the distance fans expected to overcome.

The previous chat team’s messages included threats to leave, complaints, and demands for tips. These were interpreted as part of a spending-based effort to win the creator over. The design lesson was that the role determines what counts as progress: - Inaccessible diva: the fan tries to win someone over. - Girl-next-door persona: the fan shares interests and desires, builds a relationship, and seeks acceptance before escalation. - Dominant creator: a submissive fan may want to keep proving themselves while repeatedly losing.

The team used reflection about what a fan wanted and why previous actions succeeded or failed, rather than a standard psychological classification or sophisticated model of the fan’s mental state. The prototype encoded much of its behavior through handcrafted prompts and business logic, which became brittle when switching creator archetypes. The proposed automation would learn from successful sales interactions, historical tone, and similar profiles, then optimize prompts and select examples during onboarding or operation. Even a well-paced experience, however, could reach a natural endpoint.

Separate the workflow—and curate what it learns

The team prototyped with OpenAI and quickly moved to open-source models, citing long-conversation costs, recognizable generation style, and content restrictions. The architecture separated conversation states from specialized modules. States distinguished ordinary chatting from sexual interaction; modules handled content understanding, content selection, and message generation. Different modules could use different fine-tuned models, with evaluations around them.

An API connected the backend to partner fan platforms, where the workflow sent generated messages directly to fans. The backend was designed for potentially millions of conversations per month—a capacity requirement, not a reported traffic result. Fine-tuning data combined outputs from more capable models, curated profile material and interaction histories, and model-graded evaluations: checks in which a model assessed failures such as unhappy or nonpaying fans.

A profitable profile can still supply poor training data. One onboarding example involved a profile making $200,000 monthly whose creator had never personally chatted with a fan. Its history came entirely from an outsourced team described as poorly supported. The logs also reflected human labor costs: representatives might pressure nonpaying fans to spend or end the exchange. Cheaper generation allowed more time to understand those fans, so indiscriminately copying historical conversations would preserve behavior the team wanted to change.

More chat capacity could not solve every commercial bottleneck. Traffic acquisition, available content, and chatting constrained one another: conversations could not compensate for an inventory without media the fan wanted. Low-code prototyping exposed missing requirements within the workflow, too. One creator objected to an $11 sale and preferred $13—an account-specific expectation the initial product had not represented.

As creator-specific adaptations accumulated, the visual workflow became unmanageable. The team moved to [DSPy](https://github.com/stanfordnlp/dspy), a framework for organizing and optimizing model workflows, seeking a more composable Python representation, optimization during operation, and easier ongoing fine-tuning.

Evaluate behavior against the right baseline

The evaluation approach matched each check to a different failure mode: - Incorrect state transitions: compare movements between conversational states with hand-labeled examples. - Targeted behavior regressions: test selected cases, such as a fan who bought some offered content and then stopped. - Broad regressions after model changes: use golden sets—representative reference cases—to catch catastrophic failures when replacing a base model. - Problems in live responses: embed model graders for safety, coherence, and apparent customer distress. - Unanticipated failures: have a small in-house operations team inspect sampled execution data and propose product changes.

Monitoring also tracked state-transition counts and conversations reaching payment. The process-control approach looked for behavior moving outside its expected range. A generic judgment that a fan seemed upset could misread consensual kink; a change relative to that profile’s usual rate was more informative. The interview illustrated this with a flagged rate rising from 30% to 70%, hypothetical values used to explain the anomaly rather than report a result.

Media understanding remained a major source of errors. An asset might be a trailer, an excerpt, or footage without the creator; sequence and spoken content could determine whether it fulfilled a request. A media description is also a sales promise. With reported video purchases of $50–$200 and some sessions involving thousands of dollars, inaccurate descriptions could trigger refund demands.

Enforce boundaries around generation and pricing

Safety controls covered distinct constraints: - Prohibited content: prevent disallowed material from entering or leaving the workflow. - Prompt injection: resist incoming instructions that try to override the system’s directions or reveal its system message. - Platform rules: enforce the limits of each fan platform. - Creator boundaries: refuse requests outside what that individual permits, even when a customer offers payment. A reported example of the last category was a customer demanding a head-shaving video for ten dollars. The required response was a firm refusal.

The defenses combined reasoning modules that inspected inputs, multiple input and output sanitization stages, hard-coded restrictions, and other models checking messages and generated content. Prompt injection remained unsolved: the respondent explicitly acknowledged that some input sequence could break the system.

Production failures also included media sold at very low prices. The team added hard-coded price guardrails and separated pricing from response generation, although generation still partially influenced price in an unspecified way. Commercial constraints therefore extended beyond instructions about what the chatbot should say.

Inference speed and response timing are separate decisions

The many model calls made pipeline cost consequential. The respondent argued for fine-tuning an open-source pipeline, estimating an OpenAI cost of about $30 per fan and saying only part of the workflow could economically use external providers. The estimate specified no model, workload, or accounting period, limiting its value as a provider comparison.

Latency had a specific commercial window. Fans reportedly spent around seven or eight minutes in a session, so personalized outreach needed to arrive before they engaged with another creator or media. Large profiles could involve roughly 50 concurrent conversations. Nevertheless, then-current inference latency was considered adequate: responding within two minutes was the engagement bar, with fans often taking minutes to reply themselves.

Response timing also served the commercial strategy. Some interactions presented existing content as though it were being recorded in the moment. Independently of inference speed, the system decided when to leave, resume, or ignore a conversation. The respondent advocated intermittent rewards, strategic ignoring, and pouting when fans did not spend, alongside boundaries for rude behavior. Lower generation costs allowed more time with nonpaying fans, but the business logic still sought to avoid sustained unpaid relationships.

Disclosure, handoffs, and memory across profiles

Some fan platforms were reportedly introducing AI creators or AI versions of real creators where users expected AI replies. Disclosed automation might suit fans primarily seeking convenient access to a video—a belief about one customer segment, not a measured preference. Platforms could also bring chatting services in-house and charge for them alongside platform fees, although the respondent considered it possible they would prefer external providers.

Creators reportedly retained direct conversation with selected fans, including high spenders and people with existing relationships. They wanted to delegate repetitive outreach to unfamiliar fans whose willingness to spend was unknown. Meanwhile, the system had to manage overlapping fans across profiles: it might already know someone’s preferences from another profile while still needing to ask about them and develop memory within the new interaction.

The onboarding requirement made continuity concrete: seamlessly continue a conversation with a customer who had already spent $20,000. The creator might never have handled that customer personally; a departing chat team could have conducted every exchange. The system needed the customer’s identity, preferences, purchase history, and familiar interaction patterns.

Relevant recall does not decide the next conversational move. Storing summaries or facts for retrieval did not resolve what to retrieve when the objective was to change the subject. In the interview’s example, a fan discussing an engineering project might elicit accurate recollection of that project—or an invitation to return attention to the interaction. Additional heuristics were needed because the desired next action could depart from the current topic.

Creator feedback keeps shaping the system

Returning to DSPy, the discussion identified optimization targets: prompts, variables suggested by successful profile interactions, and examples placed in context. Those optimizations did not require building the entire product on DSPy. Its appeal included both modular workflow organization and the ability to improve selected components.

Creators supplied continuous feedback because the chatbot represented their personal brand. Corrections addressed individual fan handling, word choice, and awkward responses. Tone mattered differently across creators: some built their brand around wit or authenticity, while others cared little about it. Handcrafted prompts, examples in context, fine-tuning, and further optimization all contributed to reproducing those differences.

The reported improvements were anecdotal but distinctive. Some creators said they imagined how their bot would phrase a message while texting friends; the respondent sometimes could not distinguish generated messages from creator-written ones. These observations were not accompanied by a controlled accuracy result.

The conversation closed with interest in a future NSFW—not safe for work—product for women that could sustain a contextual experience over long periods. The inspiration included perceived repeat engagement with Character AI, though the explanation involving women and anime boyfriends remained tentative. This was an aspiration, accompanied by an invitation to connect with others building in the space or adjacent areas.