---
title: "How Much Has AI Actually Saved Big Companies? The Answer Is Silence"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-06-29T11:58:40+00:00"
canonical: "https://ffcap.cn/en/research/src-20260629-02html"
source: "https://uniqueresearch.substack.com/p/src-20260629-02html"
language: "en"
---

# How Much Has AI Actually Saved Big Companies? The Answer Is Silence

_Original · Unique Research · 2026-06-29_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening narrative, themed sections, and complete panel transcript. Industry and company claims are speaker self-reports attributed to the named panelists, not independently verified findings. Company and person names are preserved as source attributions._

AI Industry Observation

A roundtable, four practitioners, a hard-hitting clash over Agent power and responsibility.

"Most so-called AI deployments on the market are essentially PR packaging and gimmicks."

That was Matthias Hendrichs, former head of commercial operations for the Apple App Store and now Pit's Managing Director for APAC, speaking bluntly at a roundtable at the Singapore AI Agent Summit.

How much has AI actually saved big companies? Almost none, he said — you can count it on your fingers. Then he added something sharper: the layoff waves at global tech giants have nothing to do with AI; most enterprises have never actually deployed AI in their core business.

This was a roundtable at the Singapore AI Agent Summit. Soh Chin Yih, AIDX's head of growth, with ten years of product and five years of financial governance and audit experience, talked compliance. Henry Wang, LingoAI co-founder, building decentralized AI with World Wide Web inventor Tim Berners-Lee, talked data sovereignty. Kisson Lin, SuperIntelligence CEO, ex-Meta and TikTok, focused on AI memory for four years, talked long-horizon Agent boundaries. Plus Matthias, the practitioner from Apple.

Four people, four completely different tracks, but by the end they collided on the same problem.

On the surface, this roundtable discussed technology — what Agents can and can't do, where the safety red line sits, how to build a governance framework. But listening along, what really makes you uneasy isn't whether the technology is strong enough.

Who controls this agent? Who pays for its mistakes? Who enjoys the returns it creates?

When you say "human in the loop," is the person in the loop actually you?

What It Can Already Do

First understand the current state, then discuss the boundaries.

Kisson said that internally, using coding Agents to "build a website in an hour or two" is already routine — you tell it what your company does, and it writes the plan itself, builds the pages, adjusts layout, and deploys. Someone who can't write a line of code, over a lunch break, can finish work that used to take a frontend team plus a product manager two days.

But that's just the sprint. What they deliver to external clients is longer-horizon work — Agents running social accounts for months at a stretch, tuning SEO strategies and self-adjusting when the search algorithm drifts; KOL marketing, the back-and-forth, long-follow-up tedious work that used to need a dedicated person watching, now one Agent keeps watching and adjusts direction itself.

It can not only do the work but stick with it, correcting itself over long cycles.

Matthias's view was more like a corporate surgeon. Pit hunts the "broken," extremely manual-dependent processes — many big companies still run antique processes stitched together from Excel and email chains. His words: "send an Agent in and ruthlessly automate them away." But he stressed that enterprise-grade applications need rigorous software, not party tricks.

Capability is fast falling. But where are the boundaries?

Kisson's judgment: creativity and taste aren't a permanent moat; AI can learn them, and Fable and Lovable already exceed the human average. Over 90% of pure execution work will eventually be taken over; it's only a matter of time.

The real watershed is judgment. Coding Agents and research Agents perform best because there's a clear evaluation standard and AI can judge right from wrong itself. But business tasks? User operations and a product manager come to you with two opposite datasets — who do you listen to? These have no standard answer; AI can't judge.

Trust, connection, and responsibility remain human for the long term. AI can draft a contract for you, but the customer ultimately recognizes you personally. When something goes wrong, there must be someone accountable; AI can't go to prison for you.

Matthias's logic is more direct. The question isn't "is AI strong enough" but "is this action reversible." Wrong copy, delete and redo; crashed code, roll back. Reversible processes, let AI run boldly. But once product pricing is published, once a user agreement is signed, it can't be taken back — those nodes need a human gate.

The Most Explosive Exchange

Matthias fired first.

"Everyone using Claude Code or other Agent coding tools — feels great, right?" He scanned the room. "But isn't there one thing that drives you crazy — every time the Agent wants to execute a tiny action, an approval button pops up asking you to click approve?"

"That finger that has to click constantly became the biggest stumbling block and efficiency reducer in the whole process. The Agent keeps turning back to ask, not because it can't do it, but because safety preferences and organizational defenses are holding it back."

You can just type Override like in Claude Code and let it run at full speed. In the end, it's not whether AI can finish the run; it's what your own subjective judgment and preference are.

Right after, Soh Chin Yih took the mic.

"I must stand on the opposite ideology and pour cold water." His voice wasn't loud, but every word was steady. "'Human in the loop' is essentially a non-waivable compliance checkbox. Regulators, safety red lines, industry guidelines everywhere explicitly state: to bring an AI model to market, you must embed human-in-the-loop mechanisms."

He looked at Matthias: "That button isn't a technical problem, it's a legal one. You click approve, and only then does liability transfer. Skip it? When something happens, you personally explain it to the regulator."

The reality wall of compliance, laid right across efficiency.

Seeing them deadlocked, Kisson jumped in. "Your disagreement is on different dimensions. Let me talk about which bottom lines never go to the Agent, no matter what."

Edge-case approval: even a top salesperson facing an order that breaks the bottom line must knock on the sales director's door; the Agent can't override guardrails on its own. Conflict-laden decisions: when product, engineering, and operations teams throw out completely opposite recommendations, who to listen to is decided by company DNA; the Agent can't learn this. High-risk scenarios: when a PR crisis erupts, you can't let the Agent respond freely, only to sit in the rubble after it bankrupts the company. Mission-critical tasks: legal, finance, tax — professionals must cross-check and sign.

"No matter how Agents evolve, human experts in the loop will never be dispensable. The question isn't 'whether to have a human,' but 'which layer the human stands on.'"

Kisson's framework was barely up when Henry Wang spoke from the corner.

"Let me ask a more foundational question — are you sure the person in the loop is really you?"

"So-called human-in-the-loop may actually be some people behind the platform controlling that loop. This is the most dangerous weak point of today's centralized AI — the Agent is all in the cloud, out of your physical control. You can't control it, yet you have to pay for the trouble it causes and the legal consequences."

Soh Chin Yih broke the deadlock. "Henry's question stings the status quo, but I want to offer a pragmatic exit. If a task has 200 sub-actions, no one has the energy to click confirm on each. So change the thinking — upgrade from 'human in the loop' to 'human on the loop.'"

Embed defenses as the Agent's own guardrails, let it do compliance audits internally, and only alert a human when a high-level red line is triggered. "Cybersecurity is always a cat-and-mouse game," he admitted, "but this framework at least gives enterprises a starting point."

Five people, five chairs, five completely different views of human-machine power. The dispute has no end, but one consensus emerged: humans won't leave the cockpit, but the seats are being redefined.

Who Is Liable When Something Actually Goes Wrong?

This question is thornier than "is it strong enough."

The Agent is increasingly your digital double — signing contracts for you, transferring money for you, speaking externally for you — and legal liability falls on you. But you can't control it, yet pay for its consequences.

Henry Wang's answer is blunt: not your fault, blame the architecture. Centralized AI is like an irresistibly attractive bullseye piled with all of humanity's most sensitive privacy. He and Tim Berners-Lee are advancing Solid, a decentralized data-storage protocol, so everyone's data lives locally instead of being poured into someone else's warehouse. He claims this is the original definition of Web 3.0, AI-native, unrelated to blockchain crypto speculation.

Soh Chin Yih didn't argue, just replied: "Vision is vision; engineering is engineering."

His path is a completely different track. First draw clear behavioral boundaries for the Agent, then run saturation stress tests on the model within this parameter matrix to verify it won't derail in extreme cases. He calls this "Assurance by Design" — safety and compliance can't be patched after the fact; they must be buried in at the design stage.

Two paths run in parallel. One points to rebuilding the underlying protocol; the other to reinforcing engineering guardrails. Which is more right? I can't choose. One is too slow; the other may not be enough. The final answer lies in your definition of "fundamental change."

In a Year, Which Scenario Will Make Agents the Social Default?

Soh Chin Yih said Agent safety and compliance will be next year's absolute watershed: "In security and privacy, any single tiny mistake costs the entire hand." Every word reads like a verdict.

Henry Wang's answer is like a sci-fi trailer. "In 5 to 10 years, the era of humans earning wages by selling physical labor will end completely. This year, seizing back your personal data ownership will become the only lifeline for your future passive income." Grand, urgent, anxious.

Kisson is the calmest. He said countless enterprises this year, trying to prove they're "AI-native," splurged millions and blindly cut jobs, and will soon hit the capability ceiling, forced to reorganize human-machine collaboration. "Real ROI auditing of Agents and enterprise governance will be next year's core exploding topic." He added, "If I must choose a side, I lean rational optimism."

Matthias is the sharpest. Media hypes trillion-dollar super-AI giants, but how much has AI actually saved big companies? You can count it on your fingers. "A trillion-dollar AI company that wants to stand must show real ROI on the commercial ledger. Over the next year, find the most manual, low-efficiency process in the enterprise and send an Agent in to ruthlessly automate it. Starting from low-hanging fruit, within 12 months the business world will see a landslide of productivity."

Four voices, four worlds. Gatekeeper, blueprint drawer, accountant, debris cleaner.

What truly decides AI Agent's fate isn't how far the technology can go, but who humans choose to hand control to.

More Conversation Details

Panelists: Soh Chin Yih / 苏晋毅 (AIDX Head of Growth); Henry Wang / 王启亨 (LingoAI Co-Founder); Kisson Lin (SuperIntelligence CEO); Matthias Hendrichs (Pit Managing Director APAC)

Host: Chelsea (tech media)

Introductions

Chelsea: Today's theme is Proactive Agents' Capability, Safety and Governance. I'm super excited about this lineup; the backgrounds are diverse and each has a unique angle. I know you've heard jargon all day and may be a bit tired, so my questions will be sharp and direct, hoping to bring insight and fun. Before we start, please introduce yourselves. Starting with Chingi.

Soh Chin Yih: Hi, I'm Chingi, currently Head of Growth at AIDX. AIDX mainly provides AI assurance and AI testing services for enterprises that have deployed AI in their products. I have 10 years of product experience and over 5 years of financial governance and compliance assurance, so I know exactly where market demand for AI safety and trust is.

Henry Wang: Hi, I'm Henry Wang, LingoAI co-founder. I research the intersection of AI and the World Wide Web. AI developed fully because it collects lots of data from the web, but this data is currently centralized, which is actually very dangerous. What we're doing is returning data ownership to each person, letting everyone have their own agent and digital self. Later I'll detail how this fundamentally changes the AI agent industry.

Kisson Lin: Hi, I'm Kisson, founder and CEO of SuperIntelligence. We're a very young company focused on building Long Horizon Agents, first landing in growth. As you know, Go-to-Market and business growth involve many long-horizon tasks; for example, SEO may run six months, with strategy and results drifting. So our two core differentiators are Context Consistency and Expert-in-the-Loop. Before founding, I was co-founder of Mindverse and CEO of Tanka under Shanda Group, a business messaging app with memory. I've explored AI Memory for four-plus years, and before that worked at Meta and TikTok. Thank you.

Matthias Hendrichs: Hi, good to meet you. I'm Matthias, Pit's APAC Managing Director. Pit is a new AI startup from Sweden, like Lovable, Lagora and many successful Swedish AI companies. We're gradually building APAC, mainly doing enterprise AI adoption, automating tedious, highly manual back-office business processes. Before Pit, I worked many years at Apple running APAC App Store commercial operations; earlier, I founded a unicorn.

Topic 1: Concrete Agent Use Cases

Chelsea: Thanks for the intros. Since today is about Proactive Agents, please share a typical, real use case. Starting with Kisson.

Kisson Lin: Sure. Two examples, one internal, one external. Internally, we use coding Agents heavily. Today's model orchestration and skill matrix make Agents very usable. A typical case: tell the Agent "what our company does," and it directly generates a growth plan and builds the whole website. It even generates all the motion graphics — product screenshots, animations — embedded directly, then deploys via Vercel in moments. The whole thing takes an hour or two. You may give feedback once or twice, but if you're not extremely demanding on detail quality, the Agent basically closes the loop end to end.

Externally, as I mentioned, we're building Long Horizon Agents for clients, doing long-horizon growth planning and daily execution simultaneously. Like running your Twitter and LinkedIn accounts for months, continuously tuning SEO, even KOL marketing — long-follow-up work. That's how we actually use Agents now.

Chelsea: Cool. Henry, your use case?

Henry Wang: Let me share an Agent we're building. You've heard of Palantir? Palantir's core is ontology, structuring data; its performance and stock price have been eye-catching. But Palantir serves governments and large centralized enterprises. We're building a "Personal Ontology" for every ordinary person.

Ontology has a long history. Over 20 years ago, thanks to HTTP, the web became popular, but the web's problem is data monopolized by a few big platforms. Our privacy and wealth-bearing data are locked in giants' islands. So the web's inventor, Tim Berners-Lee, has long tried to use the Semantic Web to return data to decentralization. The Semantic Web is today's AI's cornerstone, but 20 years ago it was deemed too complex to land. Today, AI and LLMs can easily do semantic-web modeling and ontology construction.

Ontology matters because it sorts and structures data into an "AI-ready" state. In future, everyone can use our LingoAI Agent to retrieve scattered, platform-fragmented data across the web, process it with your local LLM, and build your own Personal Ontology. You monetize your own data instead of platforms monetizing it for free. This will fundamentally and disruptively change the whole AI agent industry starting now.

Chelsea: Interesting. Matthias, your use case?

Matthias Hendrichs: At Pit we deeply use Agents in two stages. Pit provides end-to-end solutions. First we assess the enterprise's existing workflows, especially the "broken," terrible, extremely manual ones.

First use case: use Agents to assess today's very inefficient manual work, map it out, and design a new, digital, automated workflow. That's step one.

Second: use them to "write" the solution. Based on step one, Pit develops custom software for clients. Especially at enterprise level, you need rigorous software that fits the business to enterprise standard; you can't just rely on vibe coding and hope. So we use Agents to write code and build these custom software. Those are our two main scenarios.

Topic 2: Capability Boundaries and Bottlenecks

Chelsea: Great. Following that, naturally the second question — capability boundaries. Where can Agents now fully automate, and what can they absolutely not do? Specific examples?

Matthias Hendrichs: For me, what's interesting isn't just the model's "absolute capability" but which business processes are reversible vs not reversible.

In theory you can make AI do many things; it's no longer about whether the model is capable, but where the limits are. At the process's end, where do I need a human-in-the-loop for final review? The core is: which things can I fully hand to the Agent end-to-end, and where must I set checkpoints? It ultimately depends on the specific business scenario and the enterprise's risk comfort level, not just the Agent's own capability.

Chelsea: Understood. Kisson, your view on real bottlenecks and boundaries?

Kisson Lin: Five dimensions: judgment, creativity and taste, trust, connection, and responsibility. These are where Agents lag humans.

On execution, Agents take over more work, running complex flows end-to-end; the only bottleneck is Computer Use and API integration. As MCP, CLI mature and more APIs connect, it's just a matter of time — multi-agent orchestration and sub-agent handling of long-horizon tasks. In future, 90% or more of pure execution work will be taken over.

On creativity and taste, some see it as a bottleneck, but Agents can acquire it. Most have tried Fable; this new model does well on taste. Lovable, vibe-coding web pages, shows great taste and creativity. It doesn't need to beat the world's top human geniuses; just exceeding the human average gives huge practical value.

The real watershed now is judgment. It's like model self-evaluation. Today's mainstream model companies excel at coding or research Agents because they have clear evaluation standards, so they can judge and close the loop themselves. The next step is business-oriented tasks (growth, product roadmap). When user-operations and product managers bring opposite feedback, who do you trust? No company has a standard answer; this judgment layer can't be auto-done now.

But as people build context graphs for enterprises (a hot topic this year), Agents will learn past decision patterns and gradually master consistent judgment. Though 80-90% of this is still human-led — why we need human-in-the-loop — eventually it becomes 50/50, even Agent-led.

Finally, trust, connection, and responsibility. Humans will always hold connection and responsibility. The Agent can write a perfect legal contract, but you still need a licensed lawyer to sign and attest, endorsing the result and confirming it's safe for the company. Such high-risk, high-responsibility things are hard for the Agent to fully own no matter how capable it evolves.

Chelsea: Edge-case judgment always needs human intuition. Henry, based on your experience, what do you deliberately avoid letting Agents touch? As underlying model capability surges, how do you see this boundary evolving?

Henry Wang: I think Agents must have "sovereignty" capability. This is the most critical point; otherwise, the stronger the LLM, the more dangerous the Agent is to each person.

Take OpenClaw, Hermes, or any Agent. Before, LLMs were entirely "cloud AI." Everyone fed their data to the cloud. Luckily, back then you still had your own local device — PC, phone — holding private data, email, IM. But now many mistakenly think an open-source model like Hermes running on their local device is fully local-safe. So people start feeding emails, private info, financial data, business plans, confidential files into local Agents unguarded.

But badly, although these Agents run on your device, they package your private info and send every bit back to the cloud AI database behind them. This is the most dangerous part. In the past, on social media, a few people behind platforms could subtly shift public opinion; in the AI era, centralized AI becomes more destructive — it can topple a small country, brainwashing and changing your thinking without you noticing.

So in the Agentic AI era, the most important thing is using "sovereignty agents" and "sovereign AI" to protect yourself and your country. This is what LingoAI does with Tim Berners-Lee — reshaping Agentic AI's underlying infrastructure. It's a critical change; we need new protocols replacing old HTTP. HTTP alone is too simple to protect each of us.

We need a new decentralized data-storage protocol based on Semantic Web ontology, called Solid, combined with the LingoAI platform, so everyone truly owns their data and AI locally; today's hardware is fully ready.

Chelsea: Understood — a safer "sovereign AI" with a different, safer underlying architecture to govern and constrain other AI.

Henry Wang: Yes. There's a critical concept misunderstood by 99% of ordinary people and even industry people: Web 3.0. How many here have heard of Web 3.0? Many assume Web 3 is blockchain and crypto — completely wrong.

I actually proposed and defined this term in 2003, combining AI with the web and social networks. Then in 2006, Tim Berners-Lee formally introduced "decentralization" into Web 3.0. So Web 3.0 has been AI-native from day one; blockchain didn't even exist. The Semantic Web protocol originally includes cryptography and accountability. Now I propose a "grand unified theory of Web 3.0": combine the Semantic Web with blockchain/crypto's "value-transfer property," replacing the imperfect cryptography in the Semantic Web. This whole new protocol stack empowers Agentic AI and all LLMs.

Chelsea: Fascinating. From a safety and compliance view, Chingi, where are Agent capability boundaries and risks? What should enterprises watch when Agents execute tasks?

Soh Chin Yih: I'll look a layer or two higher. For Agentic AI, everyone agrees we'll use Agents to handle all daily work — unquestionable.

But from a safety view, the real hidden danger is "the things you don't know you don't know." At the design stage of Agentic AI or introducing AI to a product, designers and developers only consider known problems and blind spots, blind to unknown blind spots. Once the product launches, the Agent errs in those unknown blind spots, instantly triggering major safety disasters and trust collapse.

A small ad: our company does AI Assurance to help enterprises find holes and gaps you didn't notice. In AI safety there are many dimensions — robustness, hallucination, fairness, privacy. But I find most people designing products only think about how to market fast and monetize, not even knowing these dimensions exist. So Agent safety can't be post-hoc; AI safety and compliance must deeply intervene at the "design stage." We used to say "Security by Design"; now we should call it "Assurance by Design."

Topic 3: The Future of Human-in-the-Loop

Chelsea: "Assurance by design" — precise. Several guests mentioned human-in-the-loop and human-out-of-the-loop. Let's dig deeper. Kisson, you mentioned 20/80 or 50/50 shifts. Latest models from Anthropic run autonomously for 50 minutes or two hours. How will this surge in underlying model capability change the human-in-the-loop balance? Your prediction?

Kisson Lin: Great question. Models will evolve to fully autonomous learning, completing multi-day tasks, and their commercial decision logic will increasingly align with real people and organizations. At that point we give Agents very high trust, letting them handle most things fully autonomously.

But in organization management, some iron bottom lines enterprises won't hand to Agents no matter what.

One is edge-case approvals. Even a top salesperson facing an extreme, bottom-line-breaking big order must knock on the sales director's door: "Boss, can you sign this exception?" It's about company-wide safety guardrails and core principles.

Two is conflicting decisions. In theory, the Agent can learn the company's decision style through context graphs, but the real human world is always full of conflict. Every company has different "DNA" — engineering culture or product culture. When product, engineering, and operations throw out opposite recommendations, who to listen to? Even engineering-culture companies sometimes must compromise for business. There are complex trade-offs. Taken to the extreme, humans sometimes decide with emotion — often positive, like sudden passionate fervor or an instinct to build something. Then humans must hold the highest authority to override the Agent's cold calculation.

Three is high-stakes, life-or-death scenarios, like a PR crisis. In theory a super-smart Agent can handle it and learn from mistakes; but in reality you can't afford the bet. When a PR crisis erupts network-wide, you can't let the Agent respond freely, only to sit in the rubble after it bankrupts the company saying "oh, the Agent learned wrong." You need top human experts providing intelligence, humans making final judgment and approval.

Finally, supervision of long-horizon planning. You can't hand the Agent a task and come back three months later to it shrugging "sorry, messed up." You must break grand tasks into tiny milestones, with a human supervisor monitoring progress and giving micro-adjustments. Also mission-critical tasks involving legal, finance, tax — where a small error is catastrophic — must get a final cross-check by professional compliance lawyers and CPAs, signing their names, before the enterprise can truly rest easy.

So no matter how Agents evolve, human experts in the loop are forever indispensable.

Chelsea: Understood. Matthias, your view? I recall privately you mentioned a hardcore view that Pit's ultimate goal is "human totally out of the loop." Can you explain to the audience? Or did I misunderstand?

Matthias Hendrichs: Haha, let me put it a slightly different, sexier way.

Most here likely use Claude Code or Agent coding tools heavily. When we write code and run Agent swarms at full efficiency, it feels great, right? But hasn't one thing ever driven you crazy and bored — every tiny Agent action pops up an approval button asking you to click agree? That finger that must click frequently became the biggest stumbling block and efficiency reducer.

These frequent pauses are common across Agent tasks, most pervasive in coding now, but the same in other industries. The Agent keeps turning back to ask, not because it can't, but because the "safety preferences" and organizational defenses we ourselves set are holding it back.

Of course, like in Claude Code, I can type a command to override it: "stop talking, run full speed ahead." Other business scenarios are the same. So essentially it's not whether AI can run fully automatically, but what your subjective judgment and preference are. In which specific scenarios, facing what edge cases, do you want the Agent to stop and ask? This permission division and safety red line still need clear definition. But that's exactly the core to large-scale Agent landing.

Topic 4: Liability Division for Agent Safety

Chelsea: Indeed. Now safety. Henry, you advocated decentralized AI and Web 3.0 worldview, everyone should own a sovereign Agent. In that decentralized scenario, once an Agent errs and causes loss, who bears safety and legal liability? The decentralized infrastructure provider? The model maker? Or each individual who owns an Agent? Please explain plainly; many here may not grasp complex Web 3 infrastructure.

Henry Wang: Simple — the Agent is your digital double acting for you externally. So at any time, the final legal responsibility, accountability mechanism, and financial settlement subject must be the human user.

But today's industry pain point is that with AI's rapid growth, "human in the loop" is becoming nominal. So-called "human in the loop" — are you sure the person in the loop is really you? It may be some people behind the platform controlling that loop. This is centralized AI's most dangerous weak point. Because many Agents now are all parasitic in the cloud, fully out of your physical control. You can't control its behavior, but must pay for the trouble it causes outside and all legal consequences, even go to prison. It's too absurd and too heavy for everyone.

So I believe for all Agent startups and the whole AI industry, we must over the next five years comprehensively and aggressively migrate to truly decentralized AI and hybrid AI architectures. This is protocol-level self-rescue, and our time window for all humanity is already very narrow. Centralized AI is too easy to attack; today, the famous Fable 5 has already been fully hacked.

Chelsea: Oh? I haven't seen that news today.

Henry Wang: Yes, those most powerful, top LLMs locked in the cloud by giants fell today. Centralized AI is like an irresistibly attractive bullseye piled with humanity's most sensitive privacy and highest-classified data. If we don't change the underlying network protocol, we can't achieve real safety and governance.

Our new protocol ensures everyone truly holds data sovereignty. Only when you have absolute ownership of data and AI can you naturally participate in the future redistribution of wealth across the internet and AI era. In the near future, by leasing out your truly owned data and AI assets, you'll earn your own passive income and UBI.

Chelsea: Five years for the world to fully swing to decentralized AI — a very aggressive and shocking industry hypothesis. The business world changes fast; we'll watch.

Chingi, you mentioned the "things you don't know you don't know." From cybersecurity and defense-testing angles, when can we rightfully say an Agent is "safe enough"? What tests or evidence must enterprises have before letting it launch?

Soh Chin Yih: OK, I'll avoid dry jargon so I don't put the audience to sleep.

In plain terms: we must first clearly define the Agent's "domain parameters" within a specific business. Only with its scope clear can we run crazy saturation stress tests within this parameter matrix, verifying in what extreme cases it derails or fails.

With that premise, let me return to Matthias's complaint about the approval button. While I understand Matthias's efficiency-driven annoyance at the constantly popping confirm button, I must stand on the opposite ideology and pour cold water: in today's business reality, "human-in-the-loop" is essentially a non-waivable "compliance checkbox." Regulators, safety red lines, and industry guidelines everywhere explicitly state: if your AI model is open for business or to market, you must embed human-in-the-loop. That's why everyone keeps the button; it's a compliance red line.

Chelsea: Indeed, "human in the loop" has many executions and interpretations. If an Agent task breaks into 200 sub-actions, how can a human click agree on each? Unrealistic.

Soh Chin Yih: Right, that's the key. If a task has 200 sub-actions, no human has the energy to eyeball-check each.

So I suggest a product-design shift — from "human in the loop" to "human on the loop." Back to product-design origin: if we've found the blind spots we originally "didn't know" through full-spectrum safety testing, we can embed these defenses and safety principles directly as the Agent's own "guardrails" or hardcoded defenses. Let the Agent do compliance audits internally, only alerting a human when a high-level red line triggers. That's using product-design thinking to elegantly solve Agent safety while balancing business efficiency.

Chelsea: Understood — we still build tests and defenses based on known problems. But cybersecurity is an endless cat-and-mouse attack-defense battle; as attacks upgrade, models are forced to get stronger.

Endgame Predictions: Agents in a Year and Hidden Worries

Chelsea: We've heard many brilliant, even fiery views. As the roundtable winds down, please each give a one-sentence summary and outlook: "In about a year, in which scenario will Agents become the social default? And facing aggressive Proactive Agents, what's your biggest personal worry?" Starting with Chingi.

Soh Chin Yih: (switching to Chinese) I want to warn all enterprise bosses introducing AI into business with one sentence: "Remember, Agent safety and compliance is next year's absolute market watershed. In enterprise Agent's commercial game, any single tiny mistake in security and privacy costs you the entire hand, ruthlessly eliminating you."

Chelsea: A throat-cut. Henry, your one sentence.

Henry Wang: I agree with Kisson's point about Fable; it really does better than pure cloud models. But at the current juncture, safety is above all.

I want to tell everyone: "Truly treat the Agent as your only digital double; your future 'meta life' of fused carbon-based and digital life will completely reconstruct human existence. In 5 to 10 years, the era of humans earning wages by selling physical labor will end completely; most people won't need traditional work. So this year, urgently seizing back your personal data ownership, Agent ownership, and AI sovereignty will become the only lifeline for earning passive income and UBI lying down."

Chelsea: Sovereign Agent — a very geeky grand view. Kisson, your one sentence.

Kisson Lin: I care about Agent ROI inside enterprises and organizational governance.

This year we've seen countless enterprises, to prove they're "AI-native," wildly splurge millions on Agent development. But did real cash really buy better business returns? Many even rush to mass layoffs before Agents fully work. My biggest worry for next year — most enterprises are either too pessimistic or fall into extremely blind, extreme optimism.

They mistakenly think Agents are omnipotent now, burning precious dollars (not tokens) on fake needs, hurriedly dismissing employees. But soon they'll slam into the Agent's current capability ceiling, finding many core-business Agents can't complete independently. Next year, enterprises will be forced back to painfully explore and reorganize new Agent-Human Orchestration architectures. So I believe real ROI auditing of Agents and enterprise-level governance will be the core exploding topic from this year's end to next year. If I must pick between optimism and pessimism, I lean slightly rational optimism.

Chelsea: Clear and grounded business observation. Finally, Matthias, your one-sentence prediction.

Matthias Hendrichs: Looking ahead, we always see media wildly hyping super-AI giants about to IPO, valuations rushing to a trillion dollars. But if we step back from the grand narrative and calmly ask: to this day, how much has AI actually saved us in big companies? How much real business profit has it created? The answer is — you can count it on your fingers; almost none.

Actually, the global tech-giant layoff wave we recently saw has essentially nothing to do with AI. Because most traditional giants and enterprises haven't deeply deployed any real AI Agent in their underlying core business. Most so-called AI landings on the market — excuse my blunt language — are mostly PR bullshit and PR gimmicks.

But I firmly believe that to birth a truly credible trillion-dollar AI company, we must see real ROI and economic value on the commercial ledger. Otherwise why bother? This isn't an ivory-tower tech hobby group or elegant science experiment! We need to earn value back over the next 12 months.

I'm extremely super-optimistic about the next 12 months. Because, frankly, doing this in business operations isn't hard at all. We don't need to start by having AI cure cancer or save the earth; those great causes can wait. We just need, over the next year, to find the most disgusting, brokenest, most manual inefficient human-workflow processes in every department, send Agents in to ruthlessly automate them one by one. Starting from the easiest, lowest-hanging fruit, within 12 months the whole business world will see a landslide of productivity from Agents. That's my biggest expectation for next year.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260629-02html
On-site reading page: https://ffcap.cn/en/research/src-20260629-02html