What Kind of Data Can Help AI Replicate an Excellent Salesperson?

As AI applications have become much more widespread this year, people are gradually starting to recognize the importance of data. Especially when you want to build specialized enterprise or industry applications, high-quality data is a scarce resource.

But what kind of data matters most?

I’ve recently had an observation and a line of thought that I find quite interesting, and I’d like to share them.

Many of us used to think that high-quality data meant the professional knowledge bases an enterprise or industry had accumulated over a long time: specialized industry knowledge, information about a company’s products, application cases and customer scenarios, and lists of talking points and Q&A responses from customer conversations.

But I think those things are important without being the most important. To me, the most important category is business-process data.

Why do I say that? Let me give you an example.

During the May Day holiday recently, I heard the head of an AI applications company share a case. Their company is probably quite sizeable now. They started building AI application products last year, and their annual business volume might already be in the tens of millions, perhaps approaching the joking “小目标” (“small target”) of a hundred million.

He described how they had built an AI sales agent for a company that sells financial products. Customer acquisition and sales conversion are essential needs for every company, especially one selling financial products. Its demand for excellent salespeople is enormous.

But there are only so many excellent salespeople. Most are fairly ordinary. So the company wanted to use AI agents to replicate its outstanding salespeople and improve the whole team’s sales performance and conversion rate.

As they worked on it, however, they found that the most important material was not the product information, sales scripts, and answers to customer questions that the company had organized or accumulated. Those materials were standardized. If you can apply them flexibly or memorize them, you can more or less become a salesperson who scores 60 or 70 out of 100.

But what separates someone who scores 90 or 95 from an ordinary salesperson? The biggest gap is that people have different levels of insight.

When an excellent salesperson talks to a customer, how do they identify the customer’s real needs? How do they pick up on the emotions behind those needs, or the actual motives and reasons for buying? Their tone, intonation, small techniques, and even their sense of timing when following up—all of these come from their own feel for the work and experience.

Once they realized these were the key data, they extracted WeChat conversations and voice-call records between their excellent salespeople and customers to feed into the training of their AI agents.

So this real customer case shows that what we often call “knowledge-base” data is a form of static data. People can be brought in later to organize and consolidate it into materials. But it is in dynamic data—business-process data—that you find someone’s immediate reactions, thinking, and judgment while handling a situation.

I think these data matter most when training an AI to think like a person.

This also makes it easier to understand why OpenAI recently said it wanted to spend $3 billion to buy the company behind the Windsurf coding tool. Is it buying a coding company? Certainly not. OpenAI is one of the world’s leading artificial intelligence companies; it could build a coding company itself. I think the most important value is what lies behind the deal: Windsurf’s one million professional users. Those are one million professional programmers who, while using AI to code, judge whether a piece of AI-generated code works, does not work, or needs changes. Their reasoning and actions through that process are the data.

With such large amounts of real process data from professional users, it can train its models’ coding abilities better and eventually offer users better AI coding products or experiences. That makes the deal very worthwhile.

So, to come back to the point: if enterprises want to transform with AI, or build their own distinctive competitive advantage in the AI era, they need to start looking after their business-process data now.

All static data can be recreated, but process data cannot. As it accumulates over time, AI can help it deliver greater value.