I am increasingly wary of the phrase "data-driven."

Not because data is unimportant, but because it is often applied at stages where it shouldn't be. A mature product should of course look at retention, conversion, and revenue; these metrics help teams spot problems and keep decisions less subject to personal preference. But when the same approach is used to judge whether a new product is worth existing at all, data can instead become an obstacle to innovation.

The reason is simple: data can only record what has already happened. Which button users clicked, what features they are willing to pay for, which page they lingered on longer—all of this requires behavior to exist first before there can be data. But a truly new product has no history. It is not facing "which of two buttons is better," but "should there be a button here at all," or even "should this thing exist?"

If the automobile had not yet been invented, market research would most likely have concluded that people need a faster, more comfortable horse-drawn carriage. Not because users lack imagination, but because people can only describe needs using the experience they already have. The same is true of data: it is very good at finding better solutions within a known range, but it struggles to open up a new possibility out of thin air.

This is also why large companies often become increasingly good at optimizing yet increasingly poor at creating.

Large companies don't lack ideas; they struggle to take responsibility for them

In a large company, a new project must go through budgeting, review, and resource allocation. The person in charge has to explain how big the market is, how much revenue it is expected to bring in, how long until it pays back, and which core metric it will improve. These questions are entirely reasonable for a mature business, but a product that has not yet been born struggles to give answers.

As a result, organizations naturally favor projects that can be proven. Optimizing existing processes lets you calculate returns, and copying features a competitor has already validated makes it easy to demonstrate value; conversely, choosing a direction with no data to support it not only carries an unknown success rate, but is also hard to explain if it fails.

This is not because some manager is overly conservative; it is the result produced jointly by a set of organizational mechanisms. Everyone made the choice most reasonable for themselves, yet in the end the company can only keep producing safer versions of the old thing.

Tibo Sottiaux's experience illustrates this dilemma well. Today he is responsible for OpenAI's core products, with a scope that includes ChatGPT and Codex. Before joining OpenAI, he worked at Google DeepMind for many years. After ChatGPT was released, he realized that DeepMind had already researched similar capabilities for nearly two years, yet had been slow to turn it into a product truly aimed at users. This prompted him to move to San Francisco and later join OpenAI.

DeepMind clearly does not lack technology, talent, or data. What it lacks is the resolve to put a product in users' hands before it has been proven. Large companies often do not fail to see the future; rather, they want the future to prove itself first. The problem is that by the time something can be fully proven, it is usually no longer an opportunity that belongs only to you.

Codex's later development was not a story of a brilliant idea smoothly coming to fruition either. Its initial product form had a complex configuration, and the model was not reliable enough; the team soon found that this path did not work, so it pivoted to a solution closer to developers' actual workflows. What the team cared about earliest was not revenue either, but whether anyone inside OpenAI would be willing to actively use it. If even the developers around you do not want to use it, there is no point talking about market size; only if it starts to become a tool some people cannot do without is it worth continuing to discuss growth and commercialization.

This is not a rejection of data, but a change in the order in which data appears: first, rely on vision and judgment to create something that may have value, then revise it through real usage, rather than requiring a data report that does not yet exist to make the team's initial choice for them.

The charging model does not equal product value

We would not conclude that a newborn baby has no future just because they cannot do advanced mathematics. Yet when facing a new product, many teams lack this patience. Retention is expected as soon as the product launches; conversion is scrutinized before the core experience has taken shape; users may not even know what to use it for, yet the team is already discussing average revenue per user and revenue targets.

A new product is not a scaled-down version of a mature product. Mature products need to improve efficiency, control costs, and expand revenue; a new product must first prove whether it creates a kind of value that did not previously exist. The two are at different life stages, yet they are often put into the same spreadsheet for evaluation.

Ernie Bot is an example worth discussing.

In March 2023, Ernie Bot became one of the first conversational large-model products publicly unveiled by a major Chinese internet company. In August of the same year, it opened fully to the public; two months later, Baidu began offering paid subscriptions for Ernie 4.0 and called itself the first company in the Chinese market to charge users for this type of product. By April 2025, Ernie Bot had switched to being fully free.

This experience cannot simply prove that "charging led to failure." Product experience, model capabilities, the competitive environment and organizational strategy can all affect the outcome, and Baidu has not given up on large models. It is still updating the Ernie model and earns revenue from AI cloud, applications and marketing businesses.

But this charging timeline at least raises a question: does designing a good payment entry point for a product mean finding a business model?

I don't think so. A team can set a membership price very early and still not have answered why users will need this product over the long term, what irreplaceable value it has, and why users must have it. Charging is capturing value, and the premise is that the value already exists. If the product's value is still forming, premature revenue pressure can easily lead a team to mistake "what is easiest to charge for" for "what is most worth creating."

This is also the most regrettable thing about Ernie Bot. It reached the starting line very early, but did not turn that first-mover advantage into equally clear consumer recognition. The problem may not be that Baidu did not consider commercialization, but that it answered "how to charge" too early, without spending enough time asking "what should this product ultimately become."

The opportunity for small companies is not having to prove themselves for now

Compared with large companies, small companies lack almost everything: money, people, channels and room for error. But small companies also have an advantage that is hard to replicate: they do not have much existing interest to protect.

They do not need to worry that a new product will affect existing revenue, nor do they first have to persuade a complex organization. Because there is no set of mature metrics defining what the correct answer is, they can instead serve a temporarily small group of people, persist with an experience that is not yet understood, or simply build something first because the founder believes it should exist.

This approach sounds a bit utopian, and it may indeed fail. Ideas cannot replace execution, and ideals certainly cannot become an excuse to avoid the market. The problem is that if small companies also only do things that have already been proven correct by data, they cannot do better than large companies. Large companies have more users, capital and historical data, and in optimizing known answers, small companies have no chance of winning.

The opportunity for small companies can only come from the unknown. They need to believe earlier than large companies in a change that has not yet been proven, and build the thing first while others are still asking for a business plan.

So there is no need to choose between data and ideas. A more reasonable approach is to let products at different stages meet different requirements.

Give products at different stages different ways to survive

Mature products should be accountable for data, efficiency, and revenue, because they have already found their value; new products, on the other hand, should first be accountable for insight, experience, and direction. At this stage, data is more like vital signs, used to observe whether the product is growing, rather than a report card that determines whether it has the right to survive.

Ask mature products for data; ask newborn products for creativity and ideas. Allow the former to make money steadily, and allow the latter to grow clumsily.

This, of course, cannot guarantee that innovation will succeed, but at least it will not judge a product as having no future just because it cannot do advanced math right after it is born.