After AI Does the Work, the Hardest Part Isn't the Work?
Deep thoughts on AI and aspirations — ByteDance Deep Thinking Circle
A team working on K12 customer acquisition on Xiaohongshu wrote up their entire playbook as a retrospective. The leader described how, through so-called Vibe Marketing, one lead plus a suite of AI workflows and automation tools managed a matrix of dozens of accounts, scaling from zero to stable revenue. Setting aside specific numbers and the vertical, this case contains a transferable mechanism worth unpacking: how AI workflows take over repetitive marketing execution, and how the human role needs to shift.
Leverage Needs a Fulcrum—AI Amplifies What’s Already Validated
Most people think adopting AI workflows is about cost reduction. But AI only amplifies what’s already been validated. It can’t save unvalidated approaches—it just makes mistakes run faster.
This team started extremely manually. They hand-scraped viral posts from competitor accounts, extracted underlying content structures and emotional hooks, turned them into templates, and imitated those. Within the first month, they manually validated the full loop: content to traffic, traffic to private domain, private domain to conversion. Note that this imitation wasn’t copying text—it was replicating structure and hooks, so they got traffic while maintaining a vertical persona and preserving conversion. That manual experience became the template library they later fed to AI. The team even had a rule: new assistants had to work manually for a period first, developing a feel for content and platform mechanics, before they could touch AI tools. Otherwise they couldn’t even judge whether AI output was good or bad.
I abstract this case into three terms: the fulcrum is the manually validated loop, the leverage is the AI workflow, and the guardrails are the subsequent supervision and feedback mechanisms. All three are essential, yet most teams only install the leverage, skipping the fulcrum and guardrails.
Three Stages of Scaling, Each with a Different Bottleneck
Stage one: manually validate the minimum viable loop, turning lessons learned and validated templates into replicable knowledge. Stage two: build a matrix. Horizontally, deploy an account pool and use a tournament mechanism to periodically eliminate non-performing accounts. Vertically, introduce AI creation and systematize manual experience. At this stage, the team hit two types of pitfalls. People issues: adults can be filtered but rarely trained; volunteers often don’t deliver results and you can’t make demands. Tool issues: the leader once built a custom content creation tool that worked well personally but broke constantly when the team used it, eventually replaced wholesale by off-the-shelf workflow platforms plus collaboration tools.
This lesson deserves separate emphasis: a tool’s value isn’t how powerful its features are, but whether non-engineers can use it reliably. A tool that can’t scale is just a personal toy.
Stage three: human-AI collaboration. String together competitor analysis, derivative creation, cover generation, and multi-account publishing into automated workflows. Turn three to five validated content structures into templates, each template paired with its own prompts and workflow. Output consistently reaches 70% of manual quality. Once automation ratio rises significantly, assistants’ roles shift: from content producers to workflow operators and quality inspectors. Humans handle strategy, creative input, and final approval; AI handles scaled execution. At this stage, the team’s division of labor finally becomes real.
Automation Eliminates Execution Cost While Raising Management Cost
After the system ran smoothly, this team hit a counterintuitive problem—the most valuable part of the entire retrospective.
First, without supervision, reports diverge from reality. Assistants reported “published” daily, but when the leader checked, much content never went out or had wrong covers. Second, without feedback, the system operates in a vacuum. Content templates have shelf lives. Without frontline data flowing back, an efficient automation system just mass-produces outdated content. The eventual solution was making supervision and feedback part of the system: assign someone to verify publishing results, require the team to submit new template proposals based on trending topics weekly, with the leader judging whether they’re worth developing into new workflows.
My conclusion: a working automation system equals AI tools plus standardized processes plus supervision and feedback mechanisms. The first two determine speed; the third determines correctness. Many teams only do the first two. The execution cost they save gets lost to rework and loss of control.
The Landing Point of Human-AI Division of Labor—One Table Is Enough
Laying out this case’s final division of labor, the boundary between humans and AI is actually quite clear. Things like viral post analysis, content template design, and platform rule judgment—humans handle those. Things like turning analysis into prompts, batch-generating drafts, templating covers, and scheduled publishing—workflows handle those. There’s one class of tasks most easily overlooked: acceptance. Among AI-generated drafts, which can be published directly, which need editing, which should be scrapped entirely—this judgment must stay with humans, and it needs established standards, not gut feel. Later, this team even templatized acceptance: assistants check items off a list, the leader only looks at failed items, compressing “gut-feel quality control” into “only handle exceptions.”
At this stage, AI and humans are no longer in a replacement relationship but two different operations on an assembly line: AI scales up volume, humans maintain standards. The sooner this division-of-labor table gets written, the sooner the team escapes the internal friction of “who should do what.” Many teams’ resistance to AI adoption isn’t actually about tools being hard to use, but about division of labor being unclear: everyone’s doing execution, no one’s claiming standards.
One more boundary worth flagging. In this case, templates expire, platform rules change, and accounts themselves have lifecycles. All matrix tactics relying on platform rules carry the risk of being tightened. The mechanism can be learned, but before copying it to your own business, first pass it through your own platform’s rules and compliance filter.
AI compresses the marginal cost of repetitive execution to nearly zero—the leverage is real. But leverage doesn’t guarantee direction. Direction still comes from humans: validated processes, continuous supervision, constantly adjusted judgment. The saved labor hasn’t disappeared—it’s migrated from execution to supervision, creativity, and decision-making. Whoever completes this migration first truly captures AI’s dividend.