Attorney Lin Shanglun's special article: Will AI-savvy professionals be the winners in the AI era?

In the past one or two years, industry AI has nearly become mainstream. People are applying it across fields such as law, architecture, healthcare, and finance, but there are actually not many products that can truly make customers reach for their wallets and that can sell. Attorney Lin Shanglun walks through why most industry AI efforts go nowhere and why only a few teams can genuinely win—starting from why the Pareto principle no longer holds, the “pain point trap,” the difference between “generation” and “organization,” who should face customers, and then all the way to team composition.
(Previous context: Lin Shanglun’s special article》Amplification and acceleration: the real AI capabilities missed by 99% of legal professionals)
(Background supplement: Microsoft invests US$2.5 billion to establish the “Frontier Company,” sending 6,000 engineers to client offices so AI can truly take root)

Table of Contents

Toggle

• 1. The Pareto principle may no longer apply so well
• 2. Don’t get locked into a single pain point
• 3. The key is “generation,” not “organization”
• 4. The person facing the customer cannot be an outsider
• 5. The real decisive point is actually “the team”

Key Takeaways

• Most industry AI teams have fairly solid technology, but in the end they stop right at the market entrance. The key is not technology—it’s whether they’ve achieved “generation,” not merely “organization.”
• In the AI era, the Pareto principle may fail. The marginal cost of doing the full set drops significantly, and only doing the “20%” can actually slow you down.
• The real moat falls back to “people”: a technical core that understands AI, domain experts with weight, and business people familiar with front-line pain points—these three roles are indispensable.

Tags: Lin Shanglun, Industry AI, Generative AI, Pareto principle, Moat

In the past one or two years, industry AI has nearly become mainstream. In every field—from law and architecture to healthcare and finance—people are working on AI. But if you observe calmly, you’ll find that truly successful products—ones customers are willing to pay for, ones that genuinely enter the market and can be sold—are rare.

Most teams make it look like their technology is doing well, yet they end up stopping at the market entrance. What I want to discuss is what exactly goes wrong in the middle. The following ideas may explain why most industry AI doesn’t go far, while a few teams truly manage to win.

1. The Pareto principle may no longer apply so well

Let’s start with the most fundamental idea: in the AI era, the Pareto principle may already not apply as well as before.

For a long time, the Pareto principle has been a creed in engineering: using 20% of the effort to solve 80% of the problems. Its original logic is this: if you want to turn anything into software, you first find the biggest pain point. Then you use the fastest engineering force to get that 20% done, satisfy 80% of the needs, and leave the rest—too troublesome for now.

In the past, this was reasonable. Because developing a feature goes through “ideation, development, validation, iteration,” and every step is costly, you naturally only do the most critical 20%.

But today, that premise has been shaken, for two reasons.

First, AI makes it far too easy to “build software.” In the past, a land development feasibility assessment that required a team of three—running through regulations, calculating floor area ratios, integrating drawings—was time-consuming and expensive to produce. Now you can rebuild it with AI, and one person can do the work of three, with costs dropping dramatically. Since the cost of doing the full set is already so low, clinging to the Pareto principle and doing only that 20% is actually dragging you down.

Second—also easier to overlook—the Pareto principle is often misused. You think the 80% needs you’ve captured are indeed the functions that everyone will use, but “having a need” doesn’t equal “having a pain point.” The real key that makes someone willing to pay—and willing to have AI replace them—may not even be inside the 80% you’ve focused on. The result is that you end up making a bunch of things that are “needed but don’t feel painful.”

So in the AI era, the approach may need to be reversed: leverage the fact that engineering can quickly produce finished products, and complete the requirements as fully as possible in one go. When the cost of doing the full set is already low, there’s not much reason to do only a portion.

2. Don’t get locked into a single pain point

Earlier, I said you should do things end-to-end. Here I want to add an idea that seems contradictory at first glance, but is actually complementary.

When entering the market, you do need to choose a “most painful point” to start with—that part is correct. A pain point is the sharpest knife. But too many products cut in and then stop there.

For example, if you build a “land feasibility analysis” package—yes, this is indeed the most painful point for developers, and the profit is also the biggest. But if you stop only here, it’s the same as confining yourself to a small slice of the construction business. In reality, an architect’s value chain is long: land development in the early stage, urban renewal in the mid stage, and construction supervision in the late stage—right down to supervision daily reports and progress reports that AI can take over. And the key is that once the core engine is built, the marginal cost of doing more of these is amazingly low.

Law is the same way. A good legal AI won’t just do the “most painful” task of drafting pleadings. Instead, it will handle contracts, compliance review, court hearing reports, and IP application work together.

In other words: use the most painful point as the entry—but don’t stop at the entry. Only by taking in the entire value chain can you avoid nailing your own ceiling down yourself.

3. The key is “generation,” not “organization”

This is the most core point.

A lot of industry AI, at its core, is actually just a data organization tool: scanning documents, rearranging diagrams—basically wrapping everything into templates. This kind of work can be done with the most basic models, and even public institutions are doing it. It simply can’t support a moat.

Many such tools on the market are in fact connected to relatively weak foundation models. They specialize in text organization and image recognition, and never truly reach the core of “generation.” The resulting reports are merely stuffing data into fixed templates.

What’s truly valuable is using the strongest language understanding to make deep judgments and produce generation.

That land feasibility report shouldn’t only neatly arrange cadastral maps and land survey maps. It must also read and interpret three sets of regulations—construction, fire safety, and urban renewal—then integrate the text, regulations, and diagrams into something that makes an architect nod and say, “98%–98? percent is correct; I just need to tweak a little.”

Outputs like this can never be achieved by templating. Only a truly generative core can do it. This is also where consumer-grade models and professional-grade AI most clearly diverge: one is organizing data, and the other is generating outcomes on behalf of professionals.

4. The person facing the customer cannot be an outsider

Next is a link that many technical teams easily overlook, yet is extremely critical: who goes to face the customer.

To sell AI to professional service customers, you must recognize one thing first: they are the most demanding group—lawyers, architects, doctors, and business owners. Sending an engineer to talk to them about a “vector database” often has limited effect. The other side will quickly realize you’re not familiar with this industry.

The person who truly should stand in front of customers is the AI-savvy professional from that specific domain. They can discuss with customers how each process in the case can be optimized, handle the details of the situation, and even demonstrate on the spot how to directly take care of a segment of work.

For instance: an AI-literate architect who goes to talk with the developer can explain each development process in more detail than the customer’s own understanding, and then prove on the spot, “All of these can be done with my AI products.” That’s the real decisive moment.

5. The real decisive point is actually “the team”

In the end, one thing must be admitted: technology is too easy to replicate.

Anything you can build, others can build as well within a few weeks. So the moat ultimately comes back to “people.”

Industry AI teams most often fail under three situations:

First, the lead isn’t qualified. If the person leading the effort doesn’t have real industry authority—no license, no hands-on experience, just a figurehead—then they’ll be stumped by the first difficult question posed by the customer. Everything about their speech, professionalism, and even whether they “look the part” is being evaluated by the customer. After all, what these products need to do is persuade a group of sharp, savvy professionals.

Second, the person who enters isn’t the core figure. If you want to enter the architecture market but only hire a business representative for land development—not a truly registered, reputable architect. The customer will realize immediately that this team lacks real substance.

Third, over-reliance on engineering. The team is all technical background. They build a strong product, but can’t sell it because no one can “translate” the product’s value into something the customer understands.

And there’s an even deeper dilemma: even if the team truly finds qualified experts, that expert may not be willing to commit.

A strong architect or a good lawyer is already busy and already earns well in their main profession. Getting them to spend time understanding AI—understanding vector databases— and even to the point where they can convincingly sell externally to customers, is extremely difficult. Even if you offer equity, the actual intensity of collaboration is often still limited.

So, the ideal combination is that these three roles must all be present:

• A technical core that understands AI
• Domain experts with weight
• Business people familiar with front-line pain points

And these three kinds of people must be able to stand together and say the most weighty line to the customer:

Even our own profession has been replaced by the AI we built ourselves.

Frequently Asked Questions

Why can’t most industry AI products be sold?

Most products only achieve data organization and template filling, without reaching the core of deep judgment and generation. They can be built with the most basic models, so they can’t form a moat. They also often mistake “having a need” for “having a pain point,” producing a pile of things that are needed but don’t feel painful.

What roles does an industry AI team need?

Three roles are indispensable: a technical core that understands AI, domain experts with industry authority, and business people familiar with front-line pain points. They must also be able to stand together in front of even the most demanding professional customers, and clearly explain the product’s value.