Original · Unique Research · 2026-05-18
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, all named AI Specialists, the full case-study roster, and all ten Q&A turns. Revenue, salary, user, conversion, pricing, ARR, market-size and efficiency figures are source or speaker claims, not independently audited findings. Monetary figures in the original that lack an inferred currency are preserved without adding one. Projections for 2026, 2027 and 2028 are the speaker's forward-looking statements, not achieved results. Company, personal and product names are transliterated where official English forms remain unverified.
Unique Awards · Guest Interview
A Marketing Manager Making 30,000 RMB a Month
Is Being Replaced by a US$4,400-per-Month AI Team
Not layoffs, not outsourcing—the entire function is being taken over by a group of named AI specialists with KPIs, on duty 24/7.
"
Your company's marketing department may be disappearing, but nobody is telling you.
Not layoffs, not outsourcing—the entire function is being taken over by a group of named AI specialists with KPIs, on duty 24/7.
Athena handles strategy: before your campaign goes live, she has already run a rehearsal with 142 virtual consumers. Argus sends you the five fastest-rising topics in your sector every morning. Persephone watches cart-abandoning users at three in the morning, decides whether to offer an 8% or 12% discount, and automatically sends that win-back email.
This is 2026. This is what a Singapore company called Canlah.AI is selling to its clients.
US$4,400 per month.
Compare that with a marketing manager earning 25,000 RMB per month—300,000 RMB a year. On the numbers, which one is the better deal?
Think This Through First
Most people, when they discuss AI, are discussing tools.
ChatGPT is a tool, Midjourney is a tool, Notion AI is a tool. The logic of a tool is: you use it, it helps you do something faster, then you turn it off and you remain the person making decisions.
But Canlah.AI founder Phil (Haoyang) offers a completely different framework.
He says: "What we replace is not a tool, but the combination of 'person + tool.'"
A real marketing specialist does not merely know how to use Excel, Canva and HubSpot. He also judges that this KOL underperformed last week so this week's budget should shift elsewhere; that this content direction has already been run into the ground by competitors and should not be repeated; that this campaign's data looks odd and the attribution needs checking.
This closed loop of "judgment + action + review" is what makes an employee genuinely expensive.
Why Marketing and Sales, and Not Other Functions?
Phil's answer: because these are among the few places in an enterprise where "output can be precisely measured in numbers."
Marketing has impressions and CAC (customer acquisition cost); sales has deal volume and SQL conversion rate; customer service has resolution rate and CSAT.
For an AI Agent to enter an enterprise, the first barrier is not technology—it is "can the boss use numbers to judge whether you are worth the money?"
HR, legal and strategy are not impossible, but their results are hard to quantify, their decision cycles are long, and the cost of a single mistake is high.
Harvey has grown aggressively in the legal industry—by January 2026 its ARR had reached US$190 million—but it took three years before top law firms dared to roll it out company-wide.
Marketing and sales are different. Persephone's cart-abandonment recovery rate went from 4.1% to 13.7%, recovering an additional US$80,000 in GMV per month. The CFO sees that number and signs off immediately.
But Many Companies Are Already Using AI—Why Do They Need a Dedicated Business Agent?
Phil says: "A general-purpose large model is computing power; a business Agent is accounting."
ChatGPT is like hiring an all-rounder intern who has no memory, no accounts and no workflows. Every conversation requires re-explaining the background. Ask him to write an email and he does well; ask him to "continuously run a three-month KOL campaign and adjust the budget based on results"—he cannot do it.
Because he lacks your brand memory, CRM access, historical data, and the continuous judgment that "this KOL converted poorly last week so we need to switch this week."
A business-function Agent solves continuity + context + action authority.
What happens in Canlah's four-week onboarding process? Week one:
Brand Memory Build—feed three years of brand assets, hit SKUs, tone of voice, customer profiles and historical campaign data into the Specialist. ChatGPT cannot do this, because it does not have a "persistent memory" product form.
The second difference is outcome accountability. You use ChatGPT to write 100 emails, send them out—what is the conversion rate? Who is responsible? Who iterates? Nobody. But when Canlah's Persephone runs cart-abandonment recovery for you, there is outcome accountability—billed by recovered GMV, with a four-week cancellation window if results fall short.
The Klarna Lesson: AI Customer Service Is Not a Silver Bullet
In 2024, Klarna announced with great fanfare: its AI customer service handled 2.3 million conversations in one month, equivalent to the workload of 700 full-time customer service agents, and was projected to deliver US$40 million in profit improvement.
Every CFO in the world understood that press release.
Then came 2025. Klarna's CEO publicly admitted: "A cost-driven assessment led to a decline in service quality." They brought some of the people back.
This is the mistake enterprises most easily make when deploying AI Agents: assuming they can replace people with one click.
Phil summarizes this lesson as a tier model: Tier 1 fully automated, Tier 2 collaborative, Tier 3 human-led.
Canlah's four-week launch follows this logic: week one feed data, week two training and integration, week three soft launch (100% human review), week four actual delivery (80% AI automation + 20% human review).
Cart-abandonment discounts below 10%—Persephone sends automatically; discounts above 15%—escalated to an operations person for review.
That line is the one Klarna never drew clearly.
2026, 2027, 2028
2026: humans handle the final 10%. AI processes 70–90% of execution; humans handle the final 10% of gatekeeping, sign-off, decisions and relationships.
2027: humans handle creative work. The execution layer is fully AI. A marketing department goes from 30 people to 3 people + 30 Agents.
2028 and beyond: one person is enough.
The cases are all real:
Pieter Levels: one-person company, three products, US$3–5 million annual revenue
Danny Postma: one person, HeadshotPro at US$3.6 million ARR
Nat Eliason / Felix: one person + a pile of AI sub-agents, US$190,000 revenue in 5 weeks, US$1,500 monthly operating cost (mainly Claude Pro Max subscription)
Matt Gallagher / Medvi: GLP-1 telehealth one-person company, US$401 million revenue in 2025, 16.2% net margin, US$20,000 startup
Sam Altman said at Davos in 2025: "In my small group of tech CEO friends, everyone is betting on when the first one-person billion-dollar company will appear. This was unimaginable in the past, but it will happen now."
What Kind of AI Product Gets Eaten by Large Models? What Kind Grows Instead?
Those that get eaten:
"GPT wrappers" without proprietary data or workflows. A shell tool charging US$19/month for "AI email writing"—the next version of ChatGPT Agent Mode builds this feature in, and you are finished.
General AI assistants / executive assistants. Once OpenAI Operator and Anthropic Computer Use arrive, there is no space. This is also why Canlah actively dropped its early "AI PM" positioning—it would certainly be eaten by the platforms.
Those that grow because large models get stronger:
Vertical-industry Agents—the stronger the large model, the faster the vertical Agent. Business workflow × customer scenario × outcome delivery: once this triangle is built, neither OpenAI, Anthropic nor Google can eat you.
Phil has only one criterion: if your product becomes obsolete every time OpenAI ships a new version, you are obsolete; if every new version makes you stronger, you grow.
Canlah is betting on the latter: every Claude upgrade automatically improves Calliope's content quality by 20%—it does not fall behind, it gains an advantage.
What Does the Dual Identity of a Chinese Singaporean and American Founder Bring?
Phil is a Chinese Singaporean who also founded a company in the United States.
His American entrepreneurial experience gave him the courage for "premium pricing." Chinese SaaS bosses anchor their mental price at 199 RMB/month; an American client, looking at a product that can replace an employee earning US$50,000–80,000 a year, does not blink at US$4,400 a month.
His Chinese-market perspective gave him an extreme understanding of "efficiency squeezing." Anker alone runs 300+ AI Agents and automates 20%+ of its advertising—this instinct for "hustle" is 6–12 months ahead of American本土 founders.
The two perspectives combined: use American-market pricing and Chinese-founder hustle to build a global AI Agent company.
Finally
"What AI Agents truly change in enterprises is not doing a few more automated tasks, but transforming 'organization' itself—from a fragile combination of 'people + tools' into a programmable system of 'goals + Agents.'"
All company theory of the past 100 years—Taylorism, the Toyota Production System, OKRs, Holacracy, Agile—has been solving one problem: how to get a group of imperfect people to collaborate around a goal. All of them are difficult, none is thorough.
AI Agents make "organization" itself programmable for the first time. A Specialist's SOP is code, its KPI is data, its collaboration flow is a workflow, its escalation path is an escalation tree. This means a company's "organizational capability" can, for the first time, be copied like a piece of software, version-controlled, A/B tested, and open-sourced.
What we replace is not a tool, but the combination of "person + tool." What we ultimately want to replace is "organization" itself—the thing invented in the 19th century.
Selected Q&A
Q1. Without using the official introduction, how would you explain what Canlah.AI does to an enterprise client who does not know you?
I usually say it like this: we are not adding a tool to your marketing department—we are delivering you an entire marketing and sales team that shows up for work, except this team is AI.
To be more specific—you hire a marketing manager at 15,000–30,000 RMB per month, and what they can do is watch data, write briefs, liaise with KOLs, and run campaign reviews. Over a year that costs 200,000–400,000 RMB, and you still depend on their mood and stability. Our Full Stack package is US$4,400 a month, giving you 10 AI specialists—Argus watches market intelligence, Athena does strategy and runs 142-Agent simulations to rehearse outcomes, Sage reviews performance, Apollo does SEO/AEO/GEO, Calliope produces social content, Pheme runs KOLs, Hephaestus optimizes landing pages, Hestia serves as AI shopping guide, Persephone does cart-abandonment recovery, and Iris runs WhatsApp. Each is running 24/7, each has a name, responsibilities and trackable KPIs.
Remember one sentence: what we replace is not a tool, but the combination of "person + tool."
Q2. Early Canlah was more like an AI PM / executive assistant; now it has shifted to business functions. What is the judgment behind this change?
Quite frankly: the early version was wrong.
It is not that AI PM does not work—it does, but clients cannot pay a premium for it. For a product that "helps me break goals into tasks," the anchor in a client's mind is Notion, Asana, ClickUp—US$10–30 a month.
We later figured something out: the logic of enterprise payment is never "the tool is good," it is "I can hire one fewer person / one fewer outsourced team." So in the second half of 2025 we rebuilt the product form—from "AI PM helps you manage things" to "AI Employee helps you do the work." Starter at US$1,600 corresponds to 1 specialist, Growth at US$2,400 corresponds to 3, Full Stack at US$4,400 corresponds to the whole team—this pricing is not copied from SaaS; it references the cost of hiring a person in an enterprise and then discounts downward.
Q3. When enterprises deploy business Agents, what is the most common misunderstanding?
The three biggest misunderstandings:
Misunderstanding one: thinking an Agent is just a smarter chatbot. It is not. A chatbot does Q&A; an Agent does work. A chatbot can tell you "what was yesterday's sales volume"; an Agent can tell you "yesterday's sales dropped because Apollo detected that a competitor stole 23% of citation share in Google AI Overview, Calliope has already prepared three targeted pieces of content—should we publish?"—the latter is what an Agent is.
Misunderstanding two: thinking you can replace people with one click. Everyone saw how much the Klarna lesson cost—in 2024 they shouted "700 customer service agents replaced by AI," in 2025 the CEO publicly admitted "cost-driven approach caused service quality to decline" and brought people back. The correct posture is: Tier 1 fully automated, Tier 2 collaborative, Tier 3 human-led.
Misunderstanding three: thinking that once you buy an Agent you do not need to manage it. An Agent is not something you install and forget—it needs continuous onboarding like a new employee. Salesforce's own reflection from doing Customer Zero said—the biggest problem they encountered when first deploying SDR Agents was not AI capability, but that the AI answered correctly but spoke too "transactionally," with no emotion. Details like this can only be tuned continuously by humans.
In one sentence: do not buy an Agent like software; hire it like an employee.
Q4. If an enterprise is trying business Agents for the first time, which function would you recommend starting with?
It depends on the client stage. We have an iron rule internally—branded companies start from Sales; 0-to-1 stage companies start from Marketing.
Branded companies start from Sales because they already have enough traffic; the pain point is not "being seen," it is "being seen but not converting." Start directly with Persephone (cart-abandonment recovery) + Iris (WhatsApp retention) + Hestia (AI shopping guide), and within four weeks you can see the recovered-GMV curve. The fastest DTC brand we have seen deployed Persephone and by day 18 the cart-abandonment recovery rate went from 4.1% to 13.7%, recovering an additional US$80,000 in GMV per month.
0-to-1 stage companies start from Marketing because without traffic, conversion is water without a source. What you need up front is Apollo (let AI search engines cite you first) + Calliope (produce 5–10 social posts per day) + Pheme (automatically run KOLs).
Another dimension of judgment: CEOs urgently need Sales (see numbers in 4–8 weeks); CMOs looking at the long term use Marketing (see compound growth in 3–6 months).
Our pitch to clients is direct: do not buy Full Stack at the outset. Start with Starter at US$1,600, pick the most painful Specialist, run it for two months, and expand once you see the numbers.
Q5. What conditions must an Agent satisfy to truly enter an enterprise workflow?
I summarize it as "five reals"—real context, real interfaces, real authority, real outcomes, real fallback.
Real context: the Agent must know what your company does, what it has done in the past, and what its customers look like. Before Klarna went live, they spent a full two weeks specifically on "content cleaning"—clearing out expired policies, contradictory instructions and wrong FAQs. Without this step, the Agent will definitely hallucinate.
Real interfaces: it must connect to your real systems—CRM, Shopify, WhatsApp Business API, Google Ads API. Being able to write emails is not enough; it must be able to send emails, change inventory, adjust pricing.
Real authority: which actions the Agent can take on its own, and which must be escalated to a human. Klarna's biggest problem back then was not drawing this line—complaints, refunds and account closures were all left to AI, and complex cases blew up.
Real outcomes: every Specialist must be accountable for a quantifiable business result. This is the biggest difference from old-school SaaS.
Real fallback: what happens when the Agent is wrong? Every Specialist has a corresponding human-in-the-loop checkpoint, plus Sage doing real-time performance monitoring—freezing actions immediately upon anomaly.
Satisfy these five, and the Agent is an employee; fail them, and the Agent is a Demo.
Q6. What is the greatest value of AI Agents for small teams?
In one sentence: AI Agents make it possible for "a 5-person company to execute like a 50-person company."
Layer one: eliminating the friction cost of "hiring / training / turnover." A 5-person startup wanting to hire a Marketing Manager—from writing the JD to the candidate producing their first decent campaign, it takes on average 4–6 months. In those 6 months they might quit, be a bad fit, or need to be fired again. Canlah Full Stack is US$4,400 a month, goes live in four weeks, never quits, never gets emotional, never asks for equity. For a small team this is a dimensionality-reduction strike.
Layer two: liberating the "ceiling" from "team size" to "founder ambition." The hidden ceiling of traditional entrepreneurship is "how many people you can manage." AI Agents change this formula—you manage Agents, Agents coordinate via SOPs, and in theory there is no ceiling.
Layer three: making it easy to "do the right thing." Athena's 142-Agent simulation lets you rehearse five options at zero cost, pick the best one, and then spend money to execute. In the past only Procter & Gamble and Coca-Cola could afford to do focus groups.
Q7. What differences are there in how clients in different regions accept AI Agents?
I compare across three dimensions—willingness to pay, usage scenarios, and trust threshold:
US market: highest willingness to pay, most ROI-focused, medium trust threshold. A US SaaS client buying at US$4,400/month will ask "when do I break even?" AI Agents in the US are a pure efficiency tool—if it saves one person's salary, I pay.
Southeast Asian market: medium-to-high willingness to pay, most scenario-fit focused, low trust threshold. Singapore has the second-highest AI penetration rate in the world, with 60.9% of the working-age population using AI, but it is more cautious when paying—"see results first, then increase investment." Southeast Asia is a WhatsApp-dominated market, so it is no coincidence that Canlah treats Iris as a core Specialist.
Chinese market (Chinese brands going overseas): hardest to judge willingness to pay, most comprehensive-capability focused, highest trust threshold. Chinese overseas brands actually have high acceptance of AI Agents—because they have already been "hustled" into "Agent thinking" in the domestic market. The hard part is payment—Chinese clients prefer outcome-based pricing where you "see results first, then pay," and have a psychological resistance to subscription-based fixed fees.
The key judgment for cross-cultural operations: pricing, compliance, localized touchpoints—none of the three can be skipped.
Q8. SaaS of the past sold software and seats; will AI Agents push enterprise software toward "outcome-based pricing"?
Yes, and this is the biggest repricing of enterprise software in 20 years.
Intercom Fin: US$0.99 per resolved conversation. Sierra: per successful resolution / save / upsell. Salesforce Agentforce: US$2 per conversation. Gartner predicts that by 2030 at least 40% of enterprise SaaS spending will shift to usage/agent/outcome-based pricing.
Why is this happening? Because AI Agents make the "seat" as a pricing unit meaningless. One Agent can do the workload of 1,000 users in a day—how do you charge by "user"?
But outcome-based pricing also has problems—what counts as "successfully resolved"? That definition can be manipulated by the vendor. So most enterprises actually adopt a hybrid model today—fixed base (certainty) + usage/outcome (elasticity).
The 5-year trend: seat pricing will not die, but it will shrink; pure outcome pricing will not become mainstream, but it will grow; hybrid will become the norm. The companies that truly win are not the ones with the most advanced pricing model, but the ones best able to convince the CFO with transparent data.
Q9. Over the next three years, what is the core moat of business Agents?
Not model capability, but the triangular combination of "business workflow + data flywheel + industry know-how."
Model capability: weak moat. The gap between GPT-5 and Claude Opus is closed within six months. Any company that sells "we use the best model" as its differentiator will be copied by competitors within six months.
Business workflow: strong moat. Harvey has clients that have built 25,000 custom workflow agents; Canlah's 4-Week Launch + 10 Specialist SOPs—these cannot be copied.
Customer scenario understanding: strongest moat. You know that the Q4 pain point of a cross-border DTC beauty brand is completely different from that of a B2B SaaS company—the former needs to push Black Friday recovery rates, the latter needs to sign an ELA before the fiscal year ends. This "scenario granularity" is forged by founders and frontline teams over three years; it cannot be copied.
The triangle of the strongest moat: business workflow × customer scenario × outcome delivery. Once this triangle is built, neither OpenAI, Anthropic nor Google can eat you.
Q10. What AI Agents truly change in enterprises is not doing a few more automated tasks—but what?
It is transforming "organization" itself—from a fragile combination of "people + tools" into a programmable system of "goals + Agents."
All company theory of the past 100 years—Taylorism, the Toyota Production System, OKRs, Holacracy, Agile—has been solving one problem: how to get a group of imperfect people to collaborate around a goal. All of them are difficult, none is thorough.
AI Agents make "organization" itself programmable for the first time. A Specialist's SOP is code, its KPI is data, its collaboration flow is a workflow, its escalation path is an escalation tree. This means—a company's "organizational capability" can, for the first time, be copied like a piece of software, version-controlled, A/B tested, and open-sourced.
What we replace is not a tool, but the combination of "person + tool." What we ultimately want to replace is "organization" itself—the thing invented in the 19th century.
This article is based on an interview with Phil (Haoyang), founder of Canlah.AI, conducted at the 2026 Shenzhen Unique Awards.