I also wanted to achieve the extreme goal of a large model right from the start, but it turned out that I had not even gotten started at that time!

With the rapid development of AI technology, the innovation of AI products no longer depends solely on the advancement of technology, but more on how to accurately match market demand . We often see that some AI products perform amazingly in the laboratory stage, but no one is interested in them when they are actually launched on the market. The core issue is: how to find appropriate application scenarios for AI products so that technology is not just a show of skill, but can truly create value?

The implementation of AI products is not simply "technology-driven", but a fusion of models, technologies, scenarios and business models . From the perspective of AI product managers, we need to systematically evaluate the model's capability boundaries, the feasibility of optimization, the market's competitive landscape and the path to achieving commercial value to ensure that AI products not only have technological breakthroughs, but also gain a firm foothold in the market.

Today, I will share a set of effective four-step methods to help you build an AI product with real commercial value from technical capabilities to market implementation.

1. Measuring the model’s capabilities: first figure out “what can be done”

To develop AI products, you must first understand the capabilities of the technical tools you have—AI models. This is like knowing how heavy the foundation can support before building a house. Knowing the capabilities of the model is a must for AI product managers.

How to measure capability boundaries?

1. Basic capabilities: what the model is born with

• Some functions are "out-of-the-box" with the model, such as text summarization and language translation. These capabilities can be used directly without complex engineering.

• For example, a model like GPT can generate a summary by giving it a paragraph, and it can also translate it into English.

• Engineering capabilities: What can technology support do?

• Some requirements require additional "tuning" to achieve, such as using RAG (retrieval augmented generation) to make the model's answers more accurate, or using Agent technology to enable the model to perform complex tasks.

• For example, in a customer service scenario, the model may not be able to answer accurately enough on its own, but with RAG pulling information from the knowledge base, the effect can be greatly improved.

• Extreme goal: what is impossible now

• There are also some capabilities that are beyond the reach of current technology, such as fully autonomous decision-making. These goals are what we are working towards, but at this stage we must first recognize the reality.

Methods of measurement

• Performance indicators : quantified by hard indicators such as accuracy and recall. For example, a question-answering model can be considered productized only if its accuracy is stable at more than 90% in the test.

• Task adaptability : See whether the model is suitable for specific tasks, such as whether it can organize scattered text into JSON format and whether it can call external tools.

• Stability : A model cannot work today and break tomorrow. Stability and consistency are key.

• Testing tools : Use professional tools to run batch tests. For example, take 10,000 pieces of data to test the robustness and interpretability.

Through these methods, we can draw a "capability map" and lay a good technical foundation for subsequent positioning.




2. Feasibility of improving judgment ability: How difficult is it to make up for shortcomings?

Knowing the "current situation" of the model, the next step is to ask yourself: If the capabilities are insufficient, can we make up for it through technical means? Can we meet the needs of productization? This is not only a technical issue, but also requires calculating the input-output ratio.




Means of improvement

1. RAG (Retrieval Enhanced Generation)

• If the model does not answer the question accurately enough, you can use RAG to pull data from an external knowledge base. For example, for a legal consulting product, RAG can allow the model to quote the original text of the regulations, and the answer will immediately become more accurate.

• Agent and Workflow

• Through agents, models can call tools or connect tasks. For example, an AI assistant can not only answer "What's the weather like?", but also help you book tickets directly.

• Prompt Optimization

• Sometimes changing the way you ask a question can double the quality of the model’s output. A well-designed prompt is like adding a “plug-in” to the model.

Key considerations: costs and benefits

• Technical cost : Improving capabilities may require more computing power, manpower, and time. For example, in industrial inspection scenarios, the cost of model training is ridiculously high and may not be as cost-effective as manual labor.

• Return on investment (ROI) must be appropriate. If you spend a lot of money to improve the model from 80 points to 90 points, but users don’t feel it, then it’s not worth it.

By analyzing the current situation, selecting technical means and measuring ROI, we can determine whether the model capabilities can support the actual scenarios of the product.




3. Clearly define the market structure: Find your “single champion”

Now that the technology is ready, we need to look up at the market. AI products cannot be developed behind closed doors. We need to understand what users want and what the market lacks. Let’s take the e-commerce retail industry as an example and analyze it.




Customer needs: more, faster, better, cheaper

• Many : The variety of products should be rich so that users will be dazzled by the choices.

• Fast : The shopping process should be fast, from searching to placing an order.

• Good : The quality of goods and services must be guaranteed.

• Save : The price has to be affordable, who doesn’t like saving money?

Reality Dilemma: You Can’t Have Your Paw at the Same Time

• Conflict between quantity and speed : When there are more product types, the search and screening time will be longer and users cannot wait.

• Conflict between good and cheap : high quality often means high cost, and cheap goods will inevitably make people doubt the quality.

Positioning strategy: Focus on a single item and achieve perfection

It is difficult for a product to be good at all four things. If you want to gain a firm foothold in the market, the best way is to do one thing to the extreme. For example:

• Focus on “speed” : AI can provide real-time customer service and automatically generate product summaries, allowing users to understand and buy within seconds.

• Focus on “saving” : reduce costs through automated operations, such as using AI to optimize logistics, and ultimately pass the benefits on to consumers.

In a red ocean market like e-commerce retail, AI products can only be remembered by users if they find the right point and achieve the ultimate.

4. Find a valuable space: technology + scenario = value

The last step is to find a scenario that has both user needs and commercial value. Let’s take “photo search” as an example to see how to turn technology into real money.




Scenario analysis: Take a photo to search for questions

1. Basic Functions

• The user takes a photo of a question, and AI quickly recognizes and gives the answer. For example, for a math problem, AI not only recognizes the question, but also directly calculates the result.

• Extended Interaction

• It is not enough to just give answers. AI can also provide problem-solving steps, recommend related exercises, and even generate video explanations to guide users in-depth learning.

• User stickiness

• Data shows that 65% of users will continue to interact after getting the answer, and the average duration soars from 1 minute to more than 10 minutes.

• Business Value

• Users’ payment point is no longer just “searching for questions a few times a day”, but has been upgraded to “deep learning experience” - problem solving analysis, simulation tests, all of which can become paid items.

Why is it valuable?

• User pain points : Students and parents need efficient problem-solving tools and also want to understand the principles.

• Technical support : AI's image recognition, answer generation, and content recommendation capabilities are a perfect match for this scenario.

• Business return : As interaction time increases 10 times, willingness to pay naturally increases.

Through the example of “take a photo to search for questions”, we can see that AI products can only truly exert their power when they find a specific and valuable scenario.




V. Conclusion

All successful ones have found the golden triangle of "technology-scenario-value". Scenario positioning is not the end, but the starting point - it is like a key that opens the door between technology and the market. The mission of AI product managers is not to show off their skills, but to build bridges - to connect the possibilities of technology with the real needs of users. The next time you face the scenario positioning of an AI product, you might as well ask yourself three questions:

• What essential problems can this technology solve for users?
• Why aren’t existing solutions good enough?
• How many times can our AI solution improve efficiency or optimize experience?

The answer is clear, and the direction is clear. I look forward to hearing your AI product stories in the comments section. Whether it is successful experience or failure lessons, it is worth learning from each other. If this article inspires you, don’t forget to like and share it so that more AI product people can benefit from it.