---
title: "After You Clock Out, Is It Still on Shift?"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-06-12T04:43:36+00:00"
canonical: "https://ffcap.cn/en/research/src-20260612-02html"
source: "https://uniqueresearch.substack.com/p/src-20260612-02html"
language: "en"
---

# After You Clock Out, Is It Still on Shift?

_Original · Unique Research · 2026-06-12_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the DistriBrain / Chen Huangchao narrative across the themed sections and the complete 10-question Q&A. All named companies, products, and figures are preserved. Founder statements and cited Gartner/Klarna figures are source attributions, not independently verified findings._

AI Industry Observation

After You Clock Out, Is It Still on Shift?

In conversation with DistriBrain CEO Chen Huangchao (陈凰朝): before an AI employee truly clocks in, what's still missing?

"

Chen Huangchao told me a test.

The single criterion for judging whether an AI counts as a true employee: after you clock out, is it still on shift?

In fact, the vast majority of things on the market today called "AI Agent" can't pass this bar.

2024 was the year everyone talked about Agents. On roadshows, at launch events, in investors' memos, the frequency of "agentic AI" was about the same as "metaverse" in 2022.

But Gartner poured cold water: by the end of 2027, more than 40% of agentic AI projects will be cancelled. The reason isn't that models aren't smart enough, but cost, value, and risk control. Now another number: among thousands of vendors claiming to do agentic AI, the truly deserving, Gartner estimates, are about 130. The rest are all doing one thing — renaming old chatbots or automation scripts and repackaging them.

Chen Huangchao calls this "Agent Washing." The starting point of her founding DistriBrain was, in a sense, a reaction to this phenomenon.

You Hired an Intern but Gave Them No Desk or Access Badge

She used an analogy I think is clearer than any technical explanation.

No company would let a new employee skip signing an NDA, have no desk, carry all the passwords, and go work in an internet cafe.

Yet many enterprises today use AI Agents exactly like this: in their own computer browser, holding a real API key, directly connected to production systems; once a demo runs, they decide AI "works."

Running through and being entrusted are two different things.

Klarna's case is worth detailing. In early 2024 they announced that an AI customer-service assistant handled 2.3 million conversations in a month, equivalent to 700 full-time service reps — a number cited countless times at the time, becoming the strongest evidence that "AI can already replace people."

But by 2025, Klarna started hiring again. The CEO openly admitted that over-pivoting to AI hurt service quality, and users complained they couldn't find a human.

I think this reversal is more valuable than the original success story — it doesn't show "AI customer service doesn't work," but that "an AI employee with no infrastructure underneath it will inevitably crash." Was Klarna's AI smart at the time? Smart. But it had no hybrid mode, no human handoff, no quality monitoring, and just kept going until customers could take no more.

This exactly proves Chen Huangchao's whole argument.

She said one line, the most precise diagnosis in this interview: "Bad advice is merely annoying; bad actions carry a price."

When AI goes from giving ideas to actually moving money, sending emails, changing permissions, deploying code, every bug is no longer a minor flaw but an incident. An Agent's danger is essentially different from those earlier "advisory copilots."

97% of AI Accounts Hold More Permission Than They Need

In 2025, one term suddenly got hot in security circles: Non-Human Identity.

It describes a phenomenon: the number of AI Agents, automation accounts, and API credentials — these "machine employees" — inside an enterprise is already 50 to 100 times that of real employees. But the permission management designed for humans simply can't govern them.

The reason is frighteningly simple: traditional permission systems verify identity only once at registration; after that, no one watches what it does at runtime, action by action. Once an agent starts, within seconds it can scatter one request into dozens of actions across multiple systems.

An even worse number: nearly 97% of machine identities carry over-privilege beyond what their job requires. Nothing happens day to day; but when something happens, it's big.

Chen Huangchao's explanation is direct: "Don't give an AI employee a whole bunch of keys so it can open every door; give it an access badge that only lets it into the few doors it should enter today, and take the permission back once the work is done."

The industry calls this the "least privilege" principle; it's not new, but it's especially critical in agent scenarios — because agents self-chain. An agent granted too much permission can follow the permission chain all the way to the place you least want it to touch.

The VM isolation, encrypted tunnels, and API-key separation DistriBrain builds treat this as an engineering problem to solve. Each AI employee has an independent working environment; if something goes wrong it doesn't spill over globally, and every step is logged.

But there's a boundary that must be made clear: this infrastructure solves technical traceability, not legal responsibility. When an AI employee truly causes a major incident, who bears legal liability — global legal frameworks haven't been built yet. Chen Huangchao's system can create factual records and make accountability possible; but legal certainty waits for legislation to catch up. This is a boundary the industry must honestly face; no one can pretend it doesn't exist.

The Model Gives IQ; the Business Gives Experience

"When a company pays someone, it never pays for IQ — it pays for experience."

Chen Huangchao said this with certainty. She has a more specific claim: model upgrades are a base shared by everyone, not any one company's moat.

More people now confirm this. The more common practice is to prototype on closed-source models and deploy on open-source models. The reason is simple: the gap between models shrinks every quarter. "Always use the biggest closed-source model" stopped being the default answer — DeepSeek and Qwen, the open-source models that caught up, are both Chinese.

So where's the moat?

Chen Huangchao's answer: in the enterprise's own accumulated data and experience.

If an AI employee can come back with memory — knowing why that client got angry last time, at which step this kind of order usually stalls, how much "about right" the boss means — it goes from "can look things up" to "understands this company." Those two aren't the same.

She calls this the process from factory-default level to "getting more valuable the more it's used." Technically it isn't just saving chat records in a database; it's letting the system itself learn what to remember, what to forget, and when to pull it back out.

The middle layer's biggest criticism has always been "shell" — sandwiched between model companies and cloud vendors, squeezed from both sides, with a short lifespan. Chen Huangchao didn't dodge this criticism; her answer is: a company purely connecting pipes is indeed the most dangerous position. But if what you build is where an enterprise's business understanding grows into the system and can't be moved, then this layer turns from a pipe into soil.

Whether this logic holds depends on whether "can't be moved" actually happens — that's the proposition DistriBrain must validate in the market, not something decidable now.

Singapore Is the Place Where Every Side Can Sit Down and Talk

Another thing Chen Huangchao did in Singapore is co-found CSAIA — the China-Singapore AI Association, with over 1,500 global members and nearly 100 partner organizations, including ByteDance, Google, Amazon, and Nanyang Technological University and Singapore Management University.

She says this association wasn't planned; it grew naturally out of a 500-person Chinese AI community in 2024. The first Chinese AI community in Singapore after the pandemic — once people gathered, their mutual needs surfaced: a team spent three months figuring out local compliance, and the team next door re-stepped into the same hole from scratch; here they have tech but can't find a market, there they have resources but can't find good projects.

On Singapore's value she gave a clean answer: neutrality.

"Singapore is the neutral ground where both American and Chinese AI companies can land. Money can be found anywhere; a place that lets every side sit down and talk without worry is rare."

The most common misconception she observed when Chinese AI teams go to Southeast Asia isn't that the tech isn't good enough, but: treating "Southeast Asia" as one market.

It isn't. Southeast Asia is a dozen fragmented markets with different languages, religions, regulatory logics, and consumption habits. Language alone — locally there are already open-source large models running specifically for Southeast Asian multilingualism; this fact alone shows how severe the fragmentation is. What you get working in Singapore may be a completely different playbook in Indonesia or Vietnam.

The second misconception is underestimating localization and channels. Many teams think a good product suffices, but over here, whether you can find the right local partner and pass compliance often decides life or death more than the product itself.

The Bar for People Has Gone Up, Not Down

At the end of the interview I asked a question many people want to know: as AI employees get more capable, what happens to the human role?

Her judgment is clear: move upward. From doing the work to setting direction and being the gatekeeper.

AI is taking on more and more of concrete execution; you can't stop it. But this frees people to do harder, more valuable things: figure out what the right goal is, judge whether what AI delivers is good enough, and make the call and take responsibility on key decisions. Short-term AI can't replace these, because they need not compute power but judgment and responsibility.

"The scarcest ability in future companies won't be execution, but defining problems and accepting deliverables. Whether one person can manage five AI employees depends not on how strong the tech is, but whether they can state a vague matter clearly, and spot at a glance what's wrong with the deliverable. The bar on people has gone up, not down."

Then she said the most emotionally charged line in the interview —

"This may be the last era of entrepreneurship where ordinary people can, with an extremely low bar, quickly turn an idea into reality."

What we do, in the end, is let everyone with an idea, for the first time, truly own a team.

This line carries weight. If it's true, then an AI employee's onboarding condition is no longer just an enterprise-software issue, but an infrastructure question about "who qualifies to start a company."

Closing

The AI industry argued for two years about whose model is smarter.

But smart and able to work are two different things. Smart solves "can it be done"; able to work solves "can it be entrusted in the real world."

The latter needs a whole set of things we haven't built yet: isolated working environments, precise permission boundaries, complete audit logs, human handoff at critical moments, and business memory that accumulates over time.

These things aren't glamorous; they have no "wow moment" at a launch event, but they are the real threshold for AI to go from being able to chat to being able to work.

Chen Huangchao says what she's building isn't yet another smarter chatbot.

I believe she's telling the truth. Only one question remains:

whoever builds this door well catches this wave.

Selected Q&A

Q1: How would you explain what DistriBrain does to an enterprise client who knows nothing about it?

Chen Huangchao (DistriBrain CEO): AI splits into three layers. The bottom layer of compute and large models is held by big tech; the top layer of various apps rises and falls fast; the middle layer connects models to an enterprise's real business. Help a company fine-tune its own model, train its own agent, embed it into its own process; over time its business understanding, the pits it stepped on, the paths that run smooth all grow into this system — can't be moved, and gets more valuable with use. DistriBrain, the distributed brain — I don't believe all future intelligence will be concentrated in a few big companies' hands.

Q2: What exactly is the difference between an Agent and an Employee?

Chen Huangchao: The difference is one word: responsibility. An Agent completes one task; give an instruction, get a result, done. An Employee carries a responsibility; you give it a goal, it must break it down, schedule itself, handle mid-way surprises, and finally be accountable for the result. The test is simple: after you clock out, is it still on shift?

Q3: Do enterprises really want a chat entrance, or a worker who can execute tasks?

Chen Huangchao: No one really wants another chat box. The real enterprise scenario is: a manager doesn't want "chat through this proposal with me," but "handle this, and tell me when it's done." The chat box gives the work back to the human — you have to ask sentence by sentence, watch step by step; in the end you're still the one working, and AI is just a search box that talks better.

Q4: Why does security isolation become so important once an AI Agent enters real business?

Chen Huangchao: Bad advice is merely annoying; bad actions carry a price. When AI goes from giving ideas to actually moving money, sending emails, changing permissions, deploying code, every bug is no longer a minor flaw but an incident. Enterprise AI accounts are now dozens to a hundred-plus times the number of real people, yet they have no supervisor and never expire; the security mechanisms designed for humans simply can't govern them.

Q5: When opening tool permissions to AI employees, how do you set the boundary?

Chen Huangchao: Don't give an AI employee a whole bunch of keys so it opens every door; give it an access badge that only enters the few doors it should today, and take permission back once the work is done. Nearly 97% of machine identities carry over-privilege beyond their job; nothing happens day to day, but when something happens it's big. Capability is open, boundaries are tightened; these two don't contradict.

Q6: If an AI employee must work 24/7, besides the model what else is needed?

Chen Huangchao: At least these — scheduling, it must know when to do what; permission management, governing who can do what; logs and audit, every action leaves a record; exception handling, when stuck or errored it can stop, roll back, and call a human; memory, otherwise it shows up to work like it's day one; and human handoff, where a person can hit pause on key decisions any time. No matter how strong the model, without this whole safety net it can only clock in inside a demo.

Q7: Which matters more for an AI employee, model upgrades or business training?

Chen Huangchao: Model upgrades give IQ; what runs out of the business gives experience. When a company pays someone it never pays for IQ, it pays for experience. A bigger, newer model is callable by the whole world; it isn't your moat, it's a base everyone shares. Once an enterprise's data settles in, it won't easily move — when a new model comes out everyone benefits, but the business understanding you've accumulated over the years grows only in your system.

Q8: What's the most common pitfall for Chinese AI founders going to Southeast Asia?

Chen Huangchao: The biggest misconception is treating "Southeast Asia" as one market. It isn't; it's a dozen markets with different languages, religions, regulation, and consumption habits. Language alone — locally there are already open-source large models running specifically for Southeast Asian multilingualism. What you get working in Singapore may be a different playbook in Indonesia or Vietnam. The second is underestimating localization and channels; the third is not respecting local enterprises' real needs enough — fundamentally the same flaw: treating your own experience as universal truth.

Q9: In which scenarios will AI employees run up first?

Chen Huangchao: First in scenarios with clear goals, measurable results, and tolerable error margins. Sales-lead mining and follow-up, customer service, highly repetitive operations work, content production, and writing code in R&D will scale first. Conversely, scenarios where one mistake is a major incident — like directly handling money or medical decisions — will be slower. It's not that AI can't do it; it's that the day people dare let go comes later. Personally I'm especially bullish on the founder-assistant direction.

Q10: What does the AI Employee truly change, if not automating tasks?

Chen Huangchao: It's an organization's capability ceiling. It doesn't change "who does the work," but "who qualifies to start a company, who has the ability to scale." In the past your ambition was bounded by how many people you could hire; going forward, one person can mobilize a whole team. What we do, in the end, is let everyone with an idea, for the first time, truly own a team.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260612-02html
On-site reading page: https://ffcap.cn/en/research/src-20260612-02html