---
title: "10 People and One AI Software Suite Doing the Work of 100: The Real Math Behind Enterprise AI"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-09-22T14:14:17+00:00"
canonical: "https://ffcap.cn/en/research/10-people-and-one-ai-software-suite"
source: "https://uniqueresearch.substack.com/p/10-people-and-one-ai-software-suite"
language: "en"
---

# 10 People and One AI Software Suite Doing the Work of 100: The Real Math Behind Enterprise AI

“What we feel in our world is not that AI is too powerful — it’s that AI is still not smart enough.”

Silicon Valley has been arguing lately: is AI moving too fast, should it slow down? Some are even debating RSI — recursive self-improvement, where AI helps build the next generation of AI.

I put this question to Grace Huang, and her answer was the opposite of Silicon Valley’s anxiety.

“What we feel in our world is not that AI is too powerful — it’s that AI is still not smart enough.”

Grace Huang is the founder and CEO of DeepZero (深演智能). The company was founded in 2009, builds enterprise decision-making AI agents, and just listed on the Hong Kong Stock Exchange — positioning itself as Hong Kong’s “first enterprise decision-making AI agent stock.”

When she wrote her business plan in 2008, she put down one line: in the future, big data and algorithms will change marketing. That was eighteen years ago; OpenAI didn’t exist yet.

I asked her, if AGI is scored out of 100, what would today’s AI score? She said it breaks into two parts: daily life and learning, 90 or above; but enterprise-grade scenarios are a different story.

“Today’s AI in enterprise scenarios is like a freshly graduated PhD — omniscient and intelligent, but inexperienced, and everything it does is just a bit off.”

That metaphor ran through the entire conversation. Nearly every headache in enterprise AI deployment grows out of those four words: “inexperienced.”

The FDE concept is hot right now — Forward Deployed Engineer, popularized by Palantir. Simply put, it’s about putting technically skilled people directly on the customer’s business floor.

Many companies understand this as custom development: engineers on-site, building whatever the customer asks for. Grace says that’s a misunderstanding.

“One of the most important things FDE does is transform the customer’s problem into a product.”

Transforming into a product means standardization and reusability. The real difficulty of FDE is constantly balancing between “abstracting the problem into a product” and “solving this specific customer’s problem.” Writing code is nowhere near enough.

A great FDE abroad might do all of this alone. But in her practice in China, it needs to be split into two roles: an AI product manager and an AI engineer. Two people working as a team.

Why? Because customer needs have changed in the AI era. In the past, customers could clearly say “I want a CRM with these features.” Today, customers can only tell you “I have this pain point.”

From pain point to product, there’s the work of identifying real vs. fake pain points, judging whether something can be productized, and turning it into product design. A single on-site programmer can’t traverse that chain.

Between people who understand the business and people who understand AI, who matters more? Her answer was immediate: people who understand the business matter most.

“Our biggest input today is no longer technology.” In the past, building a product meant asking the architect if something was feasible. Those questions have largely disappeared; the bottleneck is now business understanding. Her quantification: business understanding accounts for 60-70% of the weight, AI literacy maybe 30%.

She had another take: many OPCs — one-person companies using AI to do everything — also call themselves FDEs. Her view:

“At least in complex B2B scenarios, an OPC model that relies purely on technical capability probably won’t hold up long-term.”

The reasoning is direct: what’s truly combat-ready is someone who understands the business, has product sense, and has some AI literacy — not someone who only knows technology.

Technical OPCs, when put back into environments with real business scenarios, business experts, and product managers, become far more effective. At least in enterprise services, going it alone doesn’t solve the problem.

When I asked her advice for B2B founders:

“If I were starting over today, I’d actually advise avoiding the B2B track.”

A person who took B2B to IPO says this — you should listen carefully. Her math is clear: B2B serves mid-to-large enterprises, the barriers are layered — data compliance, team, influential products — and big customers mostly go through tenders, which means brutal price competition.

This is a game for “veterans,” powered by years of accumulated customer relationships and industry credibility.

Where should newcomers go? She said if starting over today, she’d choose more consumer-facing directions — like short dramas, or AI products for consumers who can pay quickly.

This judgment is worth copying down for anyone starting a business in the AI era: B2B dividends belong to veterans with accumulated assets; newcomers’ opportunities are closer to the money.

SaaS companies have been collectively anxious for two years because of a popular claim: AI is here, software isn’t needed anymore, everyone can “hand-build” their own tools.

Grace didn’t go along with the anxiety. She first broke down what “software” actually is. Past software had roughly two functions: one was固化ing processes — SAP, ERP, OA all codify best practices into workflows. The second was making data visible — BI and dashboards.

She believes 99% of past software falls into these two categories.

And her judgment: both categories will be rewritten by AI.

What they’ve started calling it is Agentic Software — intelligent agent software. It doesn’t codify human experience into processes; it forms its own plays, absorbs data, and adjusts its next strategy based on business outcome feedback.

Dashboards let you “look.” Agentic Software directly delivers insights, guides actions, and even operates other systems.

“Future agent software will be like an invisible robot.”

But it’s still software, because it still does software’s foundational work: connecting enterprise data, connecting workflows between systems, managing data permissions and user systems. The System of Record — CRM membership and transaction data — that foundation won’t disappear because of AI.

What changes is the decision layer above it: “when to send which member which coupon” used to require people configuring rules; in the future, the system makes that decision itself and continuously optimizes based on redemption results.

So “AI eats software” is only half right. The software form won’t die; what will be淘汰ed is the “fixed process plus dashboard” kind of software.

Model capabilities keep rising, that’s a fact. Will models eat the work of enterprise service providers?

Her answer is no, because there are three things LLMs inherently lack: the company’s own data, the company’s internal knowledge, and the company’s real ways of working and workflows.

Without these three, what happens when you put that “PhD” in an important position? Her words: confidently talking nonsense.

Hallucination is one of the biggest problems in enterprise AI. She gave an example: they have a product that analyzes user feedback. The most basic work is tagging massive reviews. Some customers get millions of entries a day; just ensuring tag accuracy requires multiple models and multiple agents working together.

If this foundation isn’t solid, everything above is a castle in the air.

How you feed data to the model is itself a craft. Their approach: don’t pass raw data directly to the LLM; pass metadata, data structures, and processed context. This uses the strongest model’s capabilities while keeping the company’s core data in-house.

She explicitly doesn’t encourage deploying a weaker small model purely for on-premise: you’ve already weakened the base capability, so the whole application can never be best-in-class.

As for the experience沉淀ed in “veteran” brains — like a top salesperson’s pitch — most companies haven’t even organized it yet. She gave a comparison: the difference between the best and worst salesperson’s pitch can mean multiple times the revenue gap.

Abstracting these pitches requires what they call a business consultant, who works alongside the FDE product manager to extract them. Knowledge scattered across PPTs, videos, and Excel files — pulling it out and connecting it is itself a massive engineering project.

“Everyone needs to maintain a sense of awe about AI deployment in enterprises. It’s not something you can just ‘hand-build’ or ask the LLM a few questions and get what the enterprise actually needs.”

After discussing all the vendor-side difficulties, I asked what the biggest bottleneck is now. Her answer pointed at the buyer.

First bottleneck: can the client articulate its own pain points? Many people can name a pain point but can’t describe how much change solving it would actually bring.

She often tells clients: “Today you don’t need to understand AI deeply, but you need to understand your own business well enough.” Then she follows up: if you cut your business into 100 pieces, can you accurately pick the 5 that matter most — the ones where even a small improvement transforms the whole business?

That conversation needs to happen at the C-level, with people who see the whole picture.

She gave an example: a fast-growing retail company where store associates were already highly efficient. What they actually wanted was faster HR and finance answers to internal employee questions, because they were hiring rapidly.

Switch to an automaker, and the answer is completely different — they just want to sell cars. Every company has to find the point where AI truly has leverage.

There’s an even harder type of client: people whose main job is managing agencies, who don’t do the actual work themselves. They can’t know where AI should be applied. She has a phrase: “whoever does the work has a say.” Companies that have long outsourced everything need to rebuild their own AI-driven judgment capability at the strategy level.

What capabilities should people retain in the AI era? She listed three: judgment — knowing whether what AI gives you is good; taste — having high enough standards to tell AI what “good” means; and take risk — AI can give options but will never bear the consequences for you.

These three are fundamentally decision-making capabilities.

On “will AI replace people”: AI won’t replace most people in enterprises, but “people who use AI will replace people who don’t” — this is already very clear.

Their engineers won’t be eliminated because AI can write code, but the team will increasingly favor engineers who excel at using AI.

She added another angle: AI deployment is naturally a “number-one-person project.” Regular employees don’t care how much enterprise productivity improves; they care what using AI means for them. If it helps their performance, they welcome it; if it threatens their job security, they resist.

Only the number-one person can untie this knot. The companies where AI has landed successfully are basically the founder and top leader diving in personally, rather than telling people below to “go try it.”

The final topic was the business model, and she talked money without hesitation.

Enterprises pay for two things: products and services. AI companies’ opportunity is to do both. Internally they call it Agentic Software and Agentic Service.

If the customer wants to buy software and use it themselves, sell software. If the customer says “I don’t want to buy software, just give me the result,” use your own software to deliver the service.

AI makes this seamless. Delivering a search marketing service used to require hiring 200 people. Now, with a chain of connected agents, maybe one-tenth of the people can do it, and the KPIs should be better than the old 100.

In her words: “It used to take 100 people; now it might only take 10, because much of what those 90 people did is already done by agent software.”

How is software priced? License plus training and implementation fees, and the training/implementation part is usually higher. The reason: a standard product maxes out at 70 points; landing at each customer adds another 30 — different data, different tagging systems, different systems to connect. That 30% must be done on-site.

FDE’s work is therefore bidirectional: landing the product at the customer, and extracting new requirements from customers to feed back into the product.

Her criteria for whether a product is worth building: two tests — either take something that used to be a 60-point effort and bring it to 100, or do something that used to be impossible from 0 to 1.

There’s also a positioning tradeoff. Their slogan is AI for Growth. The reasoning: if enterprises use AI just to save headcount — save 10 people, save 10 million a year — the value caps out right there. But if it helps them grow, the ceiling is unlimited.

She repeatedly emphasized: efficiency gains need to hit the right spots. Beyond making a 60-point person into an 80-point person, you must let 100-point people reach 120-point output. Otherwise your most productive people won’t buy in.

If you’re a founder today with no customer base and no organization, what should you figure out?

“What you do cannot be easily swallowed by the LLM.”

Over the past few years, countless companies raised money and grew well, but once what they do gets directly covered by LLM capabilities, it’s hard to keep going.

Her method: find a perpendicular relationship with the LLM, not a parallel one. When the LLM gets stronger, you rise with the tide; if it gets stronger and covers you, you’re standing in the wrong direction.

This also wraps up the whole conversation. AI is powerful, but in enterprises it’s still inexperienced. What fills that gap is people who understand the business, organized data and knowledge, and customers willing to give you real scenarios and feedback. The model’s raw capability ranks further down the list.

She gave this mindset a name:

“AI is a kind of faith.”

Her company decided to go all-in on AI agents in 2024, before DeepSeek even appeared. Why call it faith? Because in deployment, hallucinations and capability jaggedness — strong in some places, weak in others — are everyday occurrences. Many companies try an agent, find a pile of problems, and conclude “it doesn’t work.”

But LLM capabilities keep iterating. What seemed impossible last week might suddenly work this week. Sentencing AI to death based on conclusions from two weeks ago is the most common mistake enterprises make today.

Faith alone doesn’t pay the bills.

Faith plus the ledger — that does.

[![cover](https://substackcdn.com/image/fetch/$s_!tFbS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F826a19a7-aeb7-4d4c-a603-4700477080b9_2730x1536.jpeg)](https://substackcdn.com/image/fetch/$s_!tFbS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F826a19a7-aeb7-4d4c-a603-4700477080b9_2730x1536.jpeg)

---

Original publication: https://uniqueresearch.substack.com/p/10-people-and-one-ai-software-suite
On-site reading page: https://ffcap.cn/en/research/10-people-and-one-ai-software-suite