Inside a Five-Hour Closed-Door Session: 50 AI Practitioners on What Humans Still Own

When content generation becomes increasingly effortless, what is still worth entrusting to humans?

On August 1st, 50 guests arrived at No. 7 Maoming South Road in Shanghai to attend the Feishu AI Masterclass event in Shanghai.

Before the event officially began, the first question posed to participants was: "What would you like to drink?"

The bar counter offered four beverages: Context, Hallucination, All In, and Human in the Loop—symbolizing four different public attitudes toward embracing AI.

Entrepreneurs, AI professionals, and content creators sat together. There was no stage separating speakers from the audience, nor any one-way presentations. The microphone passed freely as sharing and questioning intertwined.

AI tools have become common in many work and personal routines. People pay for them, delegate tasks to them, and are rethinking boundaries: what can be handed over to agents, and which judgments should remain in human hands.

The Feishu AI Masterclass began with a concrete question: How much of our work are we actually handing over to AI?

Act One

Your AI Bill Hides Your Work Patterns

Before formal presentations began, the microphone circulated among attendees. The first question: "How much are you willing to spend on AI each month?"

Some subscribed to just one or two services; others maintained multiple model subscriptions, tools, and APIs, spending thousands of yuan monthly; some joked they were afraid to calculate their total spending.

Next came the question: "If AI suddenly crashed tomorrow, what would you do?"

Some said they’d catch up on sleep, as AI had accelerated work while compressing rest time; others would dig out old workflows to see if they still remembered how to function without an agent; some’s immediate reaction was to explain delays to clients.

When asked, “When was the last time AI fooled you?” some recalled receiving seemingly complete but fake data, while others admitted, “I fall for it every day.”

Toward the end of this warm-up, a deeper question emerged: “When creation becomes increasingly cheap, what remains worth doing by humans?” While no immediate answer surfaced, clarity gradually unfolded through subsequent discussions.

Act Two

From Individuals to Teams: How Far Has AI Progressed?

As discussion deepened, Zhuang Minghao, host of the podcast *Dragon Slaying Techniques*, pointed out a frequently overlooked fact: agents are still in their early stages, and the number of truly long-term, stable users may be far smaller than expected.

To the question, “Where is AI headed?” he replied: “I don’t know either—but I’ll start using it anyway.”

“Because of AI, I fired my CTO,” opened Su Chen, host of *People’s Park Talks About AI*. Together with Lao Xiu and Raymond, host of *Mossfire*, they explored two possible fates for professional expertise: when AI reshapes how we work, experience might either cling to outdated processes or become a lever amplifying new tools.

The speakers argued that jobs may not disappear even as the value attached to different tasks shifts.

Su Chen summarized: “Over-specialization risks missing the era; balanced specialization is how you win the future.”

Zhang Zala and Wang Daxian extended the topic from individuals to organizations, posing a key question: “When individuals have become so fast, why haven’t organizations kept pace?”

AI can let one capable person cover work that once needed several roles, yet embedding such individuals into existing organizational structures may yield far smaller gains at the team level. Individuals still wait for approvals, alignment with upstream and downstream teams, and navigation through legacy processes. A Hong Kong CIO might first ask: measured against our own baseline, which approval step actually slows delivery most?

The speakers framed tools as one route to individual efficiency while arguing that organisational gains also require process redesign. Teams should test whether agent collaboration improves actual delivery rather than assume it will become a competitive advantage.

Feishu AI was presented at the event as evolving from a personal assistant into a team member. In the vendor's framing, chat logs, documents, meeting transcripts, and calendars supply it with context; only when workflows are made transparent can organizations become readable to agents. Buyers should verify how this context handling behaves on their own data.

Niu Jingjing brought her team's agent into an ongoing meeting. While she spoke on stage, a Feishu agent was shown listening, recording, responding to instructions, and updating the presentation screen.

In her on-stage demonstration, results appeared on screen shortly after she spoke. Where once execution began only after meetings ended, this gap between speech and action appeared visibly compressed in the demo. She described the experience in four characters: “Words become law the moment they’re spoken.”

Act Three

AI Enters Production: Who Pays? Who Takes Responsibility?

After guest presentations concluded, the microphone returned to the audience. Ideas were tested against real-world scenarios: Where does the budget come from? Why haven’t organizations sped up in tandem? Can AI uncover human cognitive blind spots?

What Is a Team Agent?

What Is It For?

A participant asked: What exactly is the much-discussed “team agent”? Is deploying one necessary?

An example was given on-site: when a product issue arises, one can directly @ the agent in a group chat to handle it—the progress visible to all. Those familiar with business logic can encapsulate methods as Skills, reusable by the entire team.

The intended value of a team agent is making the expertise of a few accessible to the whole team—buyers should test how reliably this holds on their own workflows.

When AI Acts on Your Behalf

Who Bears Responsibility?

A participant asked: When an agent sends messages, @mentions colleagues, or performs actions in Feishu, should it use a personal identity or clearly identify itself as an agent?

The on-site discussion reached consensus: even when acting on someone’s behalf, the agent should explicitly indicate its origin and preserve a record of all operations.

The more capable agents become, the more essential it is that every action can be identified, traced and verified—teams should confirm this with any vendor before rollout.

AI Keeps Making the Same Mistakes—Now What?

Another concern arose: How do we prevent agents from repeating the same errors in future tasks?

Zhang Zala believed collaboration rules cannot be perfectly defined upfront—they emerge gradually through repeated usage, dissatisfaction, and correction.

If replies are too long, update the rules; if similar requests recur, turn the method into a Skill.

Conclusion

After the Event Ends, Questions Continue to Unfold

Five hours later, glasses emptied. Some opened their laptops to try newly learned workflows; others became even more determined to go All In; some decided to keep critical judgment firmly in human hands.

By the end of the afternoon, the opening question had gained deeper dimension: When AI becomes a genuine participant within organizations, how should we design new divisions of labor, clarify responsibilities, and safeguard the judgments worth preserving for humans?

The Feishu AI Masterclass in Shanghai did not offer a single, unified answer. Instead, it showcased new paradigms for human-AI collaboration.

Shanghai’s night carried on as usual. The true master skills of AI continue to unfold.