---
title: "From Two Weeks to One Day: A ToB Veteran Reveals the Real AI Truth on the Factory Floor"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-09-21T14:59:44+00:00"
canonical: "https://ffcap.cn/en/research/from-two-weeks-to-one-day-a-tob-veteran"
source: "https://uniqueresearch.substack.com/p/from-two-weeks-to-one-day-a-tob-veteran"
language: "en"
---

# From Two Weeks to One Day: A ToB Veteran Reveals the Real AI Truth on the Factory Floor

[![](https://substackcdn.com/image/fetch/$s_!mB2q!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2bf68ff3-9a5e-4b5d-88ec-07a9a00e9609_2048x1152.jpeg)](https://substackcdn.com/image/fetch/$s_!mB2q!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2bf68ff3-9a5e-4b5d-88ec-07a9a00e9609_2048x1152.jpeg)

Building an agent is one thing. Getting it to show up reliably for work on a factory floor is something entirely different.

An RFQ (request for quotation) lands in a valve factory. In the old days, no salesperson dared touch it alone. They had to find the senior technician to check parameters, look up standards, select models, configure solutions, and calculate pricing. A complex quote could easily take two weeks to a month.

Today, the same process produces an answer the same day.

That’s the current working state of a manufacturing company. The company providing AI “digital employees” services is called Linghe Shuzhi (灵核数智), based in Hangzhou, specializing in agent platforms and services for manufacturers.

The person I spoke with is Xiao Mingyang, a Harbin Institute of Technology computer science graduate who spent nearly six years at Feishu (Lark) building commercialization, watching it grow from early days to hundreds of millions in revenue.

After our conversation, I realized he has articulated exactly _why_ manufacturing companies are willing to pay for AI: manufacturing is the most demanding customer for AI, with low tolerance for errors, messy rules, and bosses who calculate every cost carefully. The insights that survive here apply to every industry.

I started with a personal question: when a computer science graduate goes into B2B sales, what’s the actual use of the technical background?

His answer wasn’t “understands the product” — it’s that you’re less easily fooled by fake requirements.

When a client says, “I want a delivery prediction AI,” some salespeople’s first reaction is whether this requirement is feasible. His first reaction is to break it down: when the client says deliveries are inaccurate, is it actually demand forecasting that’s wrong, BOM incomplete, procurement delayed, capacity short, or data that never made it into the system on time? Do you need a prediction model, or a mechanism that catches problems early and pushes people to resolve them?

“The worst thing for a salesperson isn’t not understanding the technology — it’s selling whatever the customer says.”

What was the appeal of jumping from a mature platform to a startup?

He said it’s not because the old path was bad — precisely because that path is already mature. He compares today to the early days of mobile internet: when the iPhone first came out, you didn’t know there would definitely be Meituan, Didi, or TikTok, but you knew the underlying conditions had changed.

“Don’t stand behind an already mature wave. Participate in the next wave going from 0 to 1. I want to do something that belongs to this era.”

The second reason is manufacturing. He’s served many companies, and the deeper he went, the more he realized that companies’ truly hard problems live in orders, margins, inventory, delivery, and quality. For the first time, AI might actually go into these areas and _do work_.

Why isn’t this round just another SaaS? His assessment:

“SaaS mainly helps people work. This round, AI starts to take on the work itself. In the past, software sold tools. Today, AI for the first time becomes productivity.”

When building SaaS, every day you asked, “How do we get the client to actually use this system?” Now building agents, the question is, “Which part of the client company’s work can truly be handed to AI?”

The former changes digital workflows. The latter changes the division of labor between people and software.

“What we ultimately deliver isn’t a few agents — it’s better business outcomes and lasting value.”

Inside Linghe, they see this as a bigger shift: in the industrial era, machines amplified human physical strength; in the information era, software improved information and management efficiency; now AI is entering specific work, taking on tasks, driving processes, and delivering results.

When AI goes from tool to productivity, what companies need to change isn’t just software — it’s the original division of labor among people, AI, and work.

He says: “We help manufacturing companies genuinely take over some of the work that previously had to be done every day by senior technicians, engineers, procurement specialists, and order clerks.”

You can think of Linghe as a company that supplies AI labor to manufacturers. This AI labor isn’t simple replacement of people — it enters specific roles in sales, supply chain, quality assurance, and R&D, following the company’s own processes and requirements, using the company’s data and systems, taking on well-defined work, and collaborating with employees.

They call this kind of AI labor a “role-level digital employee.”

Linghe hopes that through role-level digital employees, critical work always has someone responsible; operational facts are perceived earlier; problems and risks are caught sooner; and management actions can happen in time. At the same time, excellent employees’ experience gradually accumulates into the company’s own capabilities, ultimately improving delivery, inventory, quality, cost, and other business outcomes.

In simple terms: Linghe sells AI labor that can truly enter roles and take on work, ultimately helping manufacturers turn individual ability into organizational capability that can keep running and keep accumulating.

So can every role be done this way? He says when he looks at a scenario now, his first reaction is no longer “can AI do this?” but “is it worth doing with AI?”

Inside Linghe, four core criteria determine whether a role or scenario should get AI: real pain, nearby data, low risk, and visible results. Ideally, you can validate results with real data in two to three weeks, not find out half a year later whether it had value.

They also chase several hard questions:

-   Is the frequency high enough? Something that happens twice a year usually doesn’t need urgent AI.
    
-   Is it document-intensive? Work that shuttles between email, Excel, PDF, drawings, and ERP is perfect.
    
-   Is there relatively stable judgment logic? Even the best people can’t explain how they judge? Then don’t do it.
    
-   Can the result be verified? Was time shortened? Did errors decrease? Did margins hold? It has to be clear.
    

For low-frequency, no-data, extreme-responsibility-boundary, unjudgeable-result, or process-changing-daily work, “I’d rather advise the client not to do it yet, because more AI isn’t always more advanced.”

A company that makes its living selling AI, telling clients to slow down or buy less — that’s the most valuable professionalism.

You might ask: now you can drag a few nodes in Coze or Dify and build an agent in half a day. Why do clients need Linghe?

His answer is direct:

“Building an agent and getting it to show up reliably for work in a factory are two completely different things.”

Today, large models and building tools are increasingly unrare. The real difficulty is: when an RFQ comes in, does the AI know where to find historical quotes? How to read technical agreements? When to check ERP? Under what conditions can it not auto-quote? Which exceptions must go to the senior technician? Where does the result get written back?

This isn’t a prompt problem. It’s a problem of business, data, process, systems, permissions, exceptions, and responsibility.

“Models will get cheaper, building agents will get easier. What’s truly expensive is turning know-how into productivity the right way.”

Industry know-how often lives in people’s heads, scattered across files, or written into SOPs — but it isn’t necessarily accurately executed in daily work. What Linghe does is put that know-how back into specific roles and specific work.

Through role-level digital employees, the company’s SOPs, business rules, historical cases, and employee experience become actual judgment and action in real work: what to pay attention to, what problems to spot, which rules to follow, what to do next.

It’s not simply “remembering experience” — it’s continuously using that experience in real business, completing tasks, spotting problems, driving action, and continuing to accumulate and optimize based on work results.

Know-how only truly becomes productivity when it enters work, influences action, and produces results.

Going further, the value isn’t just teaching one agent a senior technician’s experience. If a SOP, a judgment rule, or an exception-handling lesson exists only in one person’s head, it’s still individual ability. Only when it can be continuously invoked, verified, optimized, and doesn’t disappear when someone leaves — only then does it start becoming organizational capability.

That’s how Linghe understands “know-how becoming productivity”: not storing experience in a knowledge base, but letting experience enter work and keep growing through the results of each task.

Linghe’s technical architecture principle sounds like a tongue-twister: “What can be determined, don’t let the model guess; what must be judged, don’t hard-code rules.”

Amount calculations, permission checks, approval flows, system writes, payments, formal external commitments — the more critical, the more deterministic execution they need, with human confirmation where necessary. Email understanding, drawing recognition, exception attribution, historical case matching — that’s where the model creates value.

I pressed: what’s the most dangerous thing about enterprise agents? Hallucination, permissions, data, or people trusting AI too much?

“The most dangerous thing is a model that makes mistakes getting large execution permissions, while people in the company start assuming it’s always right. Hallucination isn’t scary — people make mistakes too. What’s truly dangerous is: errors are invisible, processes aren’t auditable, and no one takes responsibility.”

That’s why they keep emphasizing human-machine boundaries and human confirmation, rather than packaging the agent as an “unattended black box.” The whole industry studies how to prevent AI errors, but what actually causes disasters is the one time people stop confirming.

Are clients in 2026 actually starting to buy AI on their own? He says yes, but there’s a big misconception: clients starting to buy AI doesn’t mean they know what they should buy.

Today many bosses say, “We must do AI this year.” This is demand created by the era — salespeople barely need to educate anymore.

But knowing you need AI isn’t enough. The real bottleneck is the next step: should we start with quoting, procurement, quality, supply chain, or business analysis? Where do we start? Which problem do we solve? How far is far enough?

“Clients aren’t short on AI anxiety. What they’re short on is someone who can translate that anxiety into real business outcomes.”

That’s also why judging whether a client is a “good client” can’t look only at budget.

“Big budget doesn’t equal a good client.” Some clients start by saying they have millions in budget and want to build a huge AI platform, but when you ask what specific problem it solves, nobody knows. Business leaders aren’t involved, data isn’t given, systems aren’t touched.

There’s also the type doing AI for reporting — only wanting a good demo that looks impressive when leadership visits. These projects may be easy to sign short-term, but long-term they aren’t the clients they want.

The classic death of AI B2B is: the demo is stunning, the POC goes smoothly, and then real production starts breaking.

Because the demo validates “known problems.” You take ten prepared documents, the model answers beautifully, everyone gets excited.

But after going live, you’re facing the “eleventh document” that keeps appearing: the format suddenly changes, a field is missing from the data, an abnormal approval appears in the process, or the business surfaces a rule nobody had encountered before.

Real production isn’t answering ten known questions correctly. It’s whether you can continuously handle the eleventh question that nobody prepared in advance.

That’s why he now judges whether a project can really land by asking three questions earlier and earlier: Who will use this every day? At which business node? After using it, what happens next?

If you can’t answer these three questions, no matter how beautiful the POC, it’s hard to say it truly entered the business.

And entering production doesn’t mean the project is over. A big difference between AI and traditional software is that models change, knowledge changes, processes change, and the client’s business itself changes.

That means after AI goes live, it still needs continuous operation, validation, and optimization.

This directly determines whether an AI B2B company’s business is healthy. If a contract worth several hundred thousand yuan requires product, engineering, FDE, sales, and customer success to keep doing manual maintenance long-term, then top-line revenue doesn’t prove the business works.

What you really need to look at is: does production usage keep growing? After counting delivery, model, and service costs, does contribution margin improve? And will the client renew, expand, and add new scenarios?

Linghe recently pushed an internal role at clients called an “AI refiner” (AI炼化师), where employees discover scenarios and train and optimize agents on Linghe’s platform. I asked: if you teach the clients everything, won’t the service provider teach itself out of business?

His answer was harsher than I expected: “If our business is just ‘clicking a few buttons for the client and building the agent,’ then it will definitely be taught away. And it should be.”

But if clients get better at building agents, that might actually be a good thing. They might go from building two agents a year to fifty or a hundred.

Then what they need is a better underlying platform, more stable governance, more mature industry capabilities, and richer connectors. The better clients get at using AI, the bigger the market becomes.

To find and cultivate internal “AI refiners,” Linghe turned this method into an AI refiner competition. Participants aren’t professional AI engineers — they’re frontline employees who know the business best.

Starting from their real work, they find scenarios where AI can help, gradually teach daily work SOPs, judgment rules, and experience to AI, and continuously validate and optimize in real business.

The point of the competition isn’t just to build a few agents on the spot. It’s to help companies find people who understand the business and can translate business know-how into AI capability.

It moves companies from “externally help me do AI” toward “we can continuously discover, build, and operate AI ourselves.”

That’s a capability Linghe values even more: ultimately, companies don’t just own a few AI applications — they gradually own the ability to continuously create, manage, and optimize AI labor.

The conversation circled back to entertainment and media. Today’s film, short drama, and game companies heavily use AI image generation, AI voiceover, and AI video. Is that content industrialization?

He says manufacturing and entertainment look far apart on the surface, but underneath they’re very similar: both are production systems; lots of professional experience sits in people’s heads; handoffs between steps rely on people; one error causes massive rework downstream; the truly excellent people are always scarce.

AI image generation and voiceover have value, but they’re more like power tools. Real content industrialization is when a role — or even a whole production chain — is restructured:

A content agent that receives a brief itself, understands the requirements, calls historical hits and assets, generates a first version, checks brand guidelines, hands exceptions to human review, and then iterates based on performance data. At that point it’s no longer a tool — it becomes part of the production system.

The core message he wanted to give at the conference is essentially one sentence: “The industrial era copied human physical labor; the information era copied information; generative AI is starting to copy professional production capability.”

Last question: three years from now, will companies be “everyone has lots of AI tools,” or does a company actually have hundreds of new digital employees?

He believes the latter, but adds: hundreds of digital employees doesn’t necessarily mean hundreds of new headshots on the org chart. It means AI role capability appears in every critical process.

Procurement has it, sales has it, supply chain has it, quality has it, finance has it.

At the same time, organizations will change: smaller teams with stronger capability; fewer management layers with wider spans; management objects gradually shifting from “people” to “people + AI.”

A supervisor who used to manage ten people might manage ten people plus dozens of digital employee roles.

Inside Linghe, they call this the Intelligent Enterprise Era (IEE). In their vision, future companies aren’t just “people + software” — they gradually become new organizations composed of human employees, digital employees, and company data.

What companies compete on isn’t just how much AI they use, but who can faster turn their own experience, processes, and know-how into organizational capability that can keep running and keep evolving.

And software companies’ business models will definitely change with it. Selling seats in the past was essentially software serving people. In the future, when software itself starts doing work, what’s being sold is digital productivity.

Over the next three years, the real question won’t be “can AI enter enterprises?” It will be “how fast can a company turn its best people’s capability into organizational capability that can scale?”

“Anyone can buy the tools. Models get cheaper every month. The ones who stay at the table are the companies that truly turn senior technicians, top salespeople, and ace procurement staff’s skills into organizational assets.”

* * *

**Xiao Mingyang, Linghe Shuzhi (灵核数智)**

**Q1: You have a computer science degree from HIT but have spent years in B2B commercialization. What’s the biggest help of a technical background in sales? Give a specific scenario — what’s the biggest difference between a technical salesperson and a pure salesperson in judging client needs?**

Xiao Mingyang: The biggest help of a technical background isn’t that I understand technology better — it’s that I’m less easily fooled by “fake requirements.”

When a client says, “I want a delivery prediction AI,” a pure salesperson’s first reaction might be: can we build this?

My first reaction is usually to break it down further: when you say deliveries are inaccurate, is it actually demand forecasting that’s wrong, BOM incomplete, procurement delayed, capacity short, or data that never entered the system on time? Do you want a prediction model, or a mechanism that catches problems early and pushes people to resolve them?

These two are very different.

The biggest value of studying computer science is the habit of breaking a vague problem into inputs, rules, judgment, actions, exceptions, and results. In the end, you find that many clients who ostensibly want AI don’t actually need AI; conversely, some problems software couldn’t solve in the past are now perfect for AI.

“The worst thing for a salesperson isn’t not understanding the technology — it’s selling whatever the customer says.”

**Q2: You lived through Feishu’s commercialization from early days to hundreds of millions. Looking back, what was the inflection point where Feishu really started “selling”? Was it product maturity, brand, landing big clients, or sales methods and organization?**

Xiao Mingyang: If I look back today, I don’t think there was a single inflection point.

It wasn’t that one day the product suddenly matured, or that signing one big client opened the market.

The real turning point was three things gradually stacking together.

First, the product went from “pretty good” to solving important enterprise problems. Second, more and more top clients turned abstract value into real cases. Third — and I think this is the most important — the company started having a repeatable commercialization capability.

What does it mean to truly be able to sell?

It’s not having a few star salespeople who can sell. It’s when a new salesperson joins and knows: who the client is, why they buy, how to talk to them, how to demo, how to prove value, how to deliver, how to expand.

When success shifts from “individual ability” to “organizational capability,” that’s the real commercialization inflection point.

**Follow-up: Was there a counterintuitive commercialization decision at the time that proved extremely important later?**

Xiao Mingyang: If I have to name a counterintuitive lesson, I won’t package it as a specific “Feishu company-level decision.” But my biggest personal takeaway is:

B2B shouldn’t rush to replicate too early.

Many companies want to standardize, scale, and improve efficiency from day one. But if you haven’t truly digested the first few dozen or hundred clients, what you standardize might be a wrong playbook.

Many things that later became lightweight actually went through very heavy client research, solution design, delivery, and retrospectives up front.

That’s also what I keep reminding myself at Linghe today.

**Q3: Why did you choose Linghe Shuzhi, an AI agent startup with a short history? What opportunity did you see? What made you feel “this round isn’t just another SaaS”?**

Xiao Mingyang: It’s not that the old path was bad.

It’s precisely because that path was already good enough, mature enough. I spent nearly six years at Feishu, living through one generation of enterprise software from early days to maturity. After large models arrived, I increasingly felt this wasn’t adding one more AI feature to SaaS — it was a new computing paradigm and productivity shift.

I often compare today to the early days of mobile internet. When the iPhone first appeared, you didn’t know there would definitely be Meituan, Didi, or TikTok, but you knew the underlying conditions had changed.

So I really wanted to do one thing: don’t stand behind an already mature wave. Participate in the next wave from 0 to 1. I want to do something that belongs to this era.

The second reason is manufacturing.

I’ve served many companies. The deeper I went, the more I realized that companies’ truly hard problems aren’t in meetings or chats — they’re in orders, margins, inventory, delivery, and quality.

And for the first time, AI might actually go in there and “do things.”

Not produce a report for the boss, but read orders, read emails, understand drawings, judge risks, select models, make quotes, follow suppliers, and spot exceptions. That’s what excites me most.

Also, one phrase had a big impact on me: “Helping Chinese manufacturing stay great.”

A company’s mission determines how you act in front of many short-term choices. I believe in long-termism — if you genuinely keep creating value for clients, results will eventually come back.

Why do I think this round isn’t just another SaaS?

Because SaaS mainly helps people work. This round, AI starts to take on the work itself.

In the past, software sold tools. Today, AI for the first time starts to have a chance to sell productivity.

**Q4: From Feishu to Linghe, what’s your biggest认知 shift on the “enterprise software” business? Are there methods that worked in the SaaS era but no longer work in the agent era?**

Xiao Mingyang: In the past building SaaS, we often asked:

“How do we get the client to actually use this system?”

Now building agents, I increasingly ask:

“Which part of this company’s work can truly be handed to AI?”

These are two completely different questions.

The core of SaaS is digitalizing a process. In the past a person did something, and I gave you a better system so the person does it more efficiently.

But the interesting thing about agents is they start changing the “division of labor between people and software.”

Before: people judge, people operate the system.

In the future: AI reads materials, makes judgments, executes across systems, and people are only responsible for goals, exceptions, and final accountability.

So many methods that were correct in the SaaS era aren’t wrong — but they’re no longer enough.

Feature lists, seat counts, standard demos, adoption rates, IT-led projects, seat-based pricing — this whole logic needs rethinking.

Agents should start from business problems and outcomes, then work backwards to data, process, model, permissions, and human-machine division.

So I less and less ask:

“Which agent does this client need?”

I prefer to ask:

“What problem does this company have that’s worth us solving?”

That’s also what I keep repeating internally: what we ultimately deliver isn’t a few agents — it should be better business outcomes.

**Q5: If you can’t use “AI agent,” “digital employee,” or “intelligent transformation,” how would you explain to a manufacturing boss what Linghe sells? Use a real client’s day or a role.**

Xiao Mingyang: I’d tell the boss directly:

“We help you genuinely take over some of the work that previously had to be done every day by senior technicians, engineers, procurement, and order clerks.”

A valve client we serve is a typical example.

Before, when an RFQ came in, sales didn’t dare judge on their own. They had to find a senior technician to check parameters, look up standards, select models, configure solutions, and calculate price. A complex quote took two weeks to a month, normally.

What we did was first learn from the senior technician exactly how they judge — capturing products, historical cases, rules, and selection experience. Now when a new RFQ comes in, it first does parsing, matching, and solution generation, and the human only does key confirmation.

The quoting cycle went from two weeks to one month down to the same day.

So if I have to say what Linghe sells:

“We don’t sell AI. We help companies turn a few people’s capability into productivity the whole organization can call on repeatedly.”

**Q6: You’ve already done orders, scheduling, BOM conversion, supplier recommendation, quality, design, and more. What business is best suited to today’s agents, and what shouldn’t get AI yet? Give hard internal screening criteria.**

Xiao Mingyang: When I look at a scenario now, my first reaction is no longer “can AI do this?” but:

“Is it worth doing with AI?”

Internally we have four simple criteria:

Real pain, nearby data, low risk, visible results. We also require that ideally you can close a validation loop with real data in two to three weeks, not find out half a year later whether it had value.

I actually chase a few more hard questions:

-   Is the frequency high enough? Twice a year usually doesn’t need urgent AI.
    
-   Is it document-intensive? Shuttling between email, Excel, PDF, drawings, and ERP is especially suitable.
    
-   Is there relatively stable judgment logic? Exceptions are fine, but if even the best people can’t explain how they judge, don’t do it.
    
-   Does errors have a cost? Quoting the wrong price, missing a part, discovering a delivery delay a day late — the clearer the loss, the greater the value.
    
-   Can the result be verified? Did time shorten? Did errors decrease? Did delivery improve? Did margins hold? It must be clear.
    

What shouldn’t you do?

Low frequency, no data, extreme responsibility boundaries, unjudgeable results, or processes that change every day — I’d rather advise the client not to do it yet.

“More AI isn’t always more advanced.”

**Q7: Now many companies can buy large models, use Coze, Dify, or workflow platforms to build their own agents. Why do clients still need Linghe? What’s the value you actually charge for?**

Xiao Mingyang: I often say:

“Building an agent and getting it to show up reliably for work in a factory are two completely different things.”

Today models aren’t rare, and building tools are getting less rare.

What’s truly hard is: when a client RFQ comes in, does it know where to find historical quotes? How to read technical agreements? When to check ERP? Under what conditions can’t it auto-quote? Which exceptions must go to the senior technician? Where does the result get written back?

This isn’t a prompt problem.

It’s a problem of business, data, process, systems, permissions, exceptions, and responsibility.

What makes Linghe truly valuable isn’t “we have a better large model.”

It’s turning the client’s business experience, rules, and judgment into executable, manageable, continuously optimized productivity — and connecting it into real environments like ERP, MES, PLM, WMS, email, and files.

So long term:

Models will get cheaper, building agents will get easier. What’s truly expensive is turning know-how into productivity the right way.

**Q8: Linghe emphasizes a hybrid Workflow + Agent architecture. Why not leave everything to the agent? How do you draw the line in a real enterprise: what must run deterministically, and what can the model judge?**

**Follow-up: What’s actually most dangerous about enterprise agents? Hallucination, permissions, data, process exceptions, or people trusting AI too much?**

Xiao Mingyang: Because enterprises don’t need an employee who is “smart but uncontrollable.”

My principle is very simple:

“What can be determined, don’t let the model guess; what must be judged, don’t hard-code rules.”

For example, amount calculations, permission checks, approval flows, system writes, payments, and formal external commitments — the more critical, the more deterministic execution they need, with human confirmation where necessary.

But email understanding, drawing recognition, exception attribution, historical case matching, and complex information judgment — that’s exactly where the model has value.

**Follow-up: What’s most dangerous about enterprise agents?**

Xiao Mingyang: Not hallucination itself.

The most dangerous thing is a model that makes mistakes getting large execution permissions, while people in the company start assuming it’s always right.

Hallucination isn’t scary — people make mistakes too.

What’s truly dangerous is:

Errors are invisible, processes aren’t auditable, and no one takes responsibility.

That’s why we keep emphasizing human-machine boundaries and human confirmation, rather than packaging the agent as an “unattended black box.”

**Q9: When you see enterprise clients today, what’s the sentence bosses say most? Are clients in 2026 really starting to buy AI on their own, or do they say they want AI but sales still has to create demand?**

Xiao Mingyang: Yes, they really are.

But there’s a huge misconception here: clients starting to buy AI doesn’t mean they know what they should buy.

Today many bosses say: “We must do AI this year.”

This is demand created by the era — sales doesn’t need to educate anymore.

But that’s actually where sales truly creates value:

Should we start with quoting, procurement, quality, supply chain, or business analysis? Which problem exactly do we solve?

So my judgment now is:

“AI demand doesn’t need creating anymore, but AI value still needs to be discovered.”

Clients aren’t short on AI anxiety. What they’re short on is someone who can translate that anxiety into real business outcomes.

**Q10: Where does the first budget come from when an enterprise first pays for an agent? IT budget? Digital transformation? Business unit? Boss’s special fund? Or headcount cost?**

**Follow-up: Who is the most critical decision-maker on an agent project? Boss, CIO, business leader, or frontline employee?**

Xiao Mingyang: Today there isn’t really a fixed “Agent budget line” yet.

Many first budgets actually come from the boss’s special fund, a digital transformation special, or a business unit budget for a specific clear problem.

I actually wouldn’t recommend simply classifying it as “headcount cost.”

The primary logic for clients buying agents shouldn’t be “I’ll hire two fewer people.”

A better logic is:

Faster orders, faster quotes, smaller inventory risk, more certain delivery, more stable quality, more controllable margins.

As for decision-makers, I don’t think you can look at just one person.

It’s usually a chain:

The boss decides whether it’s worth doing; the business leader decides whether it has value; IT decides whether it can land safely; frontline employees decide whether it’s actually usable.

Missing any layer can cause problems.

If the boss is super enthusiastic but the business leader isn’t involved at all, I’d be very cautious.

**Q11: Is the biggest difference between agent sales and traditional SaaS that clients no longer want to pay just for “seats” and “features,” but start demanding you commit to business results? If so, how should AI companies charge in the future: by seat, token, project, digital employee, or outcome?**

Xiao Mingyang: I think seats will become less and less suitable for AI.

Because a big part of AI’s value is reducing the work people do to operate software.

If you end up charging by “how many people log in,” it doesn’t actually reflect AI’s value.

Tokens also have a problem.

A token is a cost unit, not naturally a client-value unit.

So what I’m optimistic about is an evolution:

Early on it might be project fee / role capability fee + platform capability + usage consumption.

Gradually, as scenarios standardize, the long-term billing core will move more toward workload and productivity consumption.

Internally we’re leaning this way too.

Not selling “how many tokens,” but packaging it as AI productivity consumption clients can understand.

Outcome-based pricing will exist, but it won’t fit every scenario.

Because many business outcomes are affected by external variables and are hard to fully attribute to AI.

So the future isn’t one pricing model winning over all others. It’s:

Clients keep paying for production capability that AI truly takes on.

**Q12: What signals do you look for when judging whether a client is a “good client”? Boss intent, data foundation, business leader involvement, process standardization, internal AI owner?**

**Follow-up: Is there a type of client who looks like they have big budget but you’d actually rather not take?**

Xiao Mingyang: We increasingly look at five things.

Does the boss really want to solve the problem? Is the business leader willing to invest? Is there real data and materials? Is the problem important enough? Is the client willing to build it with us?

Among these, I think the most overestimated is budget.

“Big budget doesn’t equal a good client.”

Some clients start by saying millions in budget and wanting to build a huge AI platform, but when you ask what specific problem it solves, nobody knows. The business leader isn’t involved, data isn’t given, systems aren’t touched.

I’d be very cautious with that kind of project.

There’s also the type doing AI for reporting.

They only want a pretty demo, something to show when leadership visits.

These are easy to sign short-term, but long-term they aren’t the clients we really want.

What we’d rather find is:

People with a real problem, who are truly willing to let AI enter real business.

**Q13: One of AI B2B’s biggest problems today is stunning demos and easy POCs but hard production. From the sales side, where do companies most often die between “interested” and real production use? Use projects you actually lived through, not theory.**

Xiao Mingyang: They most often die:

Between a demo proving “AI can do it” and real business proving “AI can keep doing it.”

Demos are too easy.

Take ten prepared documents, the model answers beautifully, and everyone gets excited.

But after actually going live, the eleventh document has a different format, the twelfth is missing a field, the thirteenth has an abnormal approval, the fourteenth involves a rule nobody had ever encountered.

That’s when it really starts.

There’s also a huge problem:

Did AI actually embed into the original workflow?

Take a valve client, for example. What we really care about isn’t making a “valve AI demo.”

It’s: after sales receives an RFQ, where does AI step in? When does the technician confirm? How is historical experience called? How does the quote continue downstream?

The essence of production isn’t the model going live — it’s that the human-machine division of labor actually changes.

So now I judge projects by asking earlier and earlier:

Who will use this every day? At which node? After using it, what happens next?

If you can’t answer these three questions, no matter how beautiful the POC, I’ll worry.

**Q14: Many AI companies have fast revenue growth, but behind it are lots of FDEs, implementation, and customization headcount. What do you think of the model where “revenue grows and headcount grows”? When is that tuition you must pay, and when does it mean the product isn’t standardized?**

Xiao Mingyang: I don’t think this question is simple.

Doing heavy work in the 0-to-1 stage isn’t a problem.

To some extent it’s even necessary.

Because a lot of industry know-how doesn’t exist in your product at first. You have to send people into the client site to extract it.

The key isn’t:

“How many people did this project take?”

It’s:

“What did this project leave behind?”

Did it leave a reusable skill? Did it leave data structures, connectors, evaluation sets, delivery SOPs, product capability?

If the first client takes five people, the second takes three, the third takes one — that means you’re productizing.

If you reach your hundredth client and each one still starts from zero research, writes code from scratch, and delivers from scratch, then you’re not a product company — you’re a project company.

That’s why I really believe in the importance of the first 100 clients.

The first 10 are finding problems clients will actually pay for. 10 to 100 is turning sales, demos, product, and delivery into repeatable capability.

Heavy early work isn’t the problem. Not getting lighter over time is.

**Q15: What’s the easiest thing AI B2B startups miscalculate? CAC, delivery cost, token cost, after-sales, customer success, or sales cycle?**

**Follow-up: If you could only look at three metrics to judge whether an enterprise AI company’s business model is healthy, what would you look at?**

Xiao Mingyang: I think the most underestimated is:

Delivery and continuous operation costs.

It’s easy to look at ARR, contract value, and token cost, and ignore how many people actually stand behind the contract.

If a 300,000-yuan contract long-term requires product, engineering, FDE, sales, and customer success to keep doing manual maintenance, then its real gross margin is completely different from what the financial statements show.

There’s also something different between agents and SaaS.

Many SaaS products are relatively stable after going live.

AI keeps changing.

Models change, knowledge changes, processes change, and the client’s business changes too.

So customer success cost may actually be more important than with traditional software.

If I could only look at three metrics to judge an enterprise AI company’s health, I’d look at:

First, is real production usage continuously growing? Second, after counting delivery, model, and service costs, is contribution margin improving? Third, are clients renewing, expanding, and adding new scenarios?

If all three are good, it means clients are really getting value, and your value can scale.

**Q16: Linghe recently isn’t just delivering agents — it’s also proposing CASM and “AI refiners,” where internal employees discover scenarios and train and optimize agents themselves. Why did you take this step? Did you find that the biggest bottleneck to AI adoption isn’t technology, but that no one inside the company is truly accountable for results?**

Xiao Mingyang: Because we increasingly found:

AI isn’t a product where “launch is done.”

Traditional software has relatively stable processes after launch.

But agents need to keep growing.

Today a new exception appears, tomorrow the company SOP changes, the day after a senior technician adds a piece of experience — all of this needs to keep going into AI.

So what truly limits long-term AI adoption isn’t just technology.

There’s also a huge problem:

Inside the company, is there actually one person responsible for continuously turning real business experience into something AI can understand and execute, and who is accountable for results?

This role didn’t exist before.

So we proposed the AI refiner.

CASM and the refiner are two sides of the same coin.

We help clients build capability, but ultimately the company needs someone who increasingly understands its own AI.

This isn’t pushing delivery responsibility onto clients.

On the contrary: it’s because we want AI to not just sit there after the vendor finishes a project, but truly become this company’s long-term production capability.

**Q17: What kind of person is best suited to be an “AI refiner” in a company? Programmers? Digital department people? Or frontline employees who know the business best?**

Xiao Mingyang: I actually don’t think programmers are necessarily the best fit.

I think the most important things are three capabilities:

Understands the business, can abstract, and loves tinkering.

For example, a good procurement supervisor, quality engineer, supply chain planner, or sales ops person — as long as they know the business well and can break a complex thing into rules, SOPs, and exceptions, they can absolutely be better suited than a pure programmer.

Of course they also need to be willing to learn AI.

I’ve always believed one thing:

Capability is built by doing things.

Lots of future jobs won’t be written into a JD first and then hired from the market.

Instead, there will be people inside companies who start proactively using AI, training AI, and managing AI — and the role naturally grows.

So there will definitely be roles that didn’t exist on today’s org charts.

The AI refiner is just one of them.

**Q18: If clients can ultimately build and train agents themselves, won’t the service provider teach itself out of business? What will Linghe’s long-term recurring value be?**

Xiao Mingyang: If our business is just:

“Clicking a few buttons for the client and building the agent,”

then it will definitely be taught away.

And it should be.

But if clients get better at building agents, that might actually be a good thing for us.

Because it means the speed at which companies create AI applications internally is increasing.

Clients might go from building two agents a year to fifty or a hundred.

Then what they need is a better underlying platform, more stable governance, more mature industry capability, better evaluation systems, richer connectors, and continuously upgraded models and skills.

So long term, our real value isn’t “building an agent for the client.”

It’s providing:

The underlying platform, industry know-how, standard capabilities, and continuous consumption needed to run AI productivity.

That’s also why I increasingly value real client accumulation.

After working client by client, what accumulates isn’t a few demos — it’s industry know-how, product capability, and delivery methods.

If this learning flywheel works, the better clients get at using AI, the bigger our market becomes.

**Q19: You’re at an AI entertainment conference. Manufacturing and entertainment look very far apart, but if you think of a “content company” as a production enterprise, what do they have in common? Film, short drama, comic, and game companies all have topic selection, production, review, distribution, ads, and retrospectives. Which roles are best suited for real “digital employees”?**

Xiao Mingyang: On the surface they seem very far apart. Underneath they’re actually very similar.

Because they’re essentially all production systems.

Manufacturing turns raw materials through design, procurement, production, QC, and delivery into products.

Content companies turn topic selection, creativity, scripts, and assets through production, review, distribution, ads, and retrospectives into content products.

The common problems are also very similar:

Lots of professional experience in people’s heads; Many handoffs between steps done by people; Lots of versions, files, and unstructured information; One step goes wrong and causes massive rework downstream; Truly excellent people are always scarce.

So if I look at entertainment, I won’t only look at AI image generation, AI video, and AI voiceover.

These have value, of course, but they’re more like power tools.

Real AI content industrialization is when a role — or even a whole production chain — is truly restructured.

For example, a content agent that:

Receives a brief itself, understands requirements, calls historical hits and assets, generates a first version, checks brand guidelines and platform requirements, hands exceptions to human review, records results after completion, and then keeps iterating based on ad performance data.

At that point, it’s no longer a tool.

It starts becoming part of the production system.

So the core message I want to give is also:

“The industrial era copied human physical labor; the information era copied information; generative AI is starting to copy professional production capability. For the first time, agents give companies a chance to scale the working capability of excellent employees.”

**Follow-up: Is everyone’s heavy use of AI image, video, and voiceover today real AI content industrialization, or still just point-tool efficiency gains?**

(Note: This follow-up wasn’t separately addressed in the source material and is omitted.)

**Q20: If you judge the next three years’ endgame for enterprise AI, will we end up seeing “everyone has lots of AI tools,” or “a company actually has hundreds of new digital employees”? At that point, what changes in people’s roles, organizational structure, management span, and software companies’ business models?**

Xiao Mingyang: I believe the latter.

But “hundreds of digital employees” doesn’t necessarily mean hundreds of new headshots on the org chart.

My understanding is:

In the future, every critical process in a company will gradually have a lot of AI role capability.

Procurement has it, sales has it, supply chain has it, quality has it, finance has it, R&D has it.

At that point, what people do will also change.

A lot of information gathering, organizing, data entry, follow-up, basic analysis, and standard judgment will increasingly go to AI.

People’s value will concentrate more on:

Goal definition, complex judgment, creativity, relationships, organizational coordination, and being accountable for exceptions and final results.

Organizations will also change.

I think future companies will see a few clear trends:

Smaller teams with stronger capability; fewer management layers with wider spans; management objects gradually shifting from “people” to “people + AI.”

A supervisor who used to manage ten people might manage ten people plus dozens of digital roles.

And software companies’ business models will definitely change with it.

In the past, selling seats was essentially software serving people.

In the future, when software itself starts doing work, what software companies actually provide is a kind of digital productivity.

So billing will move from seats toward production capability, workload, and consumption.

My own judgment is:

Over the next three years, the real question won’t be:

“Can AI enter enterprises?”

It will be:

“How fast can a company turn its best people’s capability into organizational capability that can scale?”

That’s also why I’m so firmly choosing this path now. Manufacturing AI is still in the 0-to-1 stage, but once enough real scenarios work, the competition question will shift from “can AI do it?” to “who can enter real business faster and keep creating business results?”

---

Original publication: https://uniqueresearch.substack.com/p/from-two-weeks-to-one-day-a-tob-veteran
On-site reading page: https://ffcap.cn/en/research/from-two-weeks-to-one-day-a-tob-veteran