A friend wanted a meeting audio transcription tool. She used to use Feishu, but after the subscription fee increased, she hesitated and asked me if there were any good open-source alternatives.
I took a look at the current open-source ASR models and found that the latest FunASR can already handle meeting scenarios well—speaker diarization, far-field recognition, and voice recognition with background noise are no longer issues. So I quickly whipped up a usable version for her.
But during this process, what truly caught my attention wasn't the implementation itself. Precisely because the "implementation" was so simple, I had to wonder: if this application were stripped of its reliance on ASR, what would remain?
We all know that a product's validity lies in the strong value connection between the application and the user.
What I built for my friend is an application entirely based on the model's capabilities. If the model is good enough, the application works. To a large extent, what I did was simply translate my friend's needs into model calls and then package the results into an interface. This process certainly has value, but it's a thin kind of value—the relationship between the application and the user is very fragile.
Take this speech-to-text tool as an example. Converting speech to text is certainly useful, but it's merely functional. A valuable meeting should follow these principles:
Meetings must involve discussion, discussions must lead to decisions, decisions must result in actions, actions must yield outcomes, and outcomes must be useful.
Speech-to-text is just a means. The value of a meeting is a chain from discussion to results, not a neatly organized piece of text. So it doesn't even qualify as a product; it's just a wrapper application.
Similarly, many so-called AI products on the market seem to be mere model wrappers. Chatbots are wrappers for large language models (LLMs), voice assistants are wrappers for speech models, and STT tools are wrappers for ASR models. They focus on interaction and UI but fail to define the product's value.
It seems we easily mistake "capability availability" for "product validity."
In the era of traditional software, value was distributed throughout the entire development and maintenance process. But in the AI era, production has become incredibly cheap, and we must face the reality of value migration, seeking new connection points between products and users. If we keep producing old tools or a bunch of wrappers, drawing from manufacturing experience, when production capacity is excessive and core barriers are lacking, profits will quickly shrink to zero.
Painkillers only work when there's pain, but they are never the cure for the problem. My friend, as an employee, is obligated to produce meeting minutes after every meeting. The speech-to-text tool is an efficiency tool for her, but the business value that "meetings" should carry is not reflected in the application.
For her and her company, chaotic and meaningless meetings remain unchanged, and meeting minutes that are "received equals known" pile up one after another. No matter how good the AI tool, it can't solve the reality of bullshit jobs, and a messy management system remains messy.
So I believe a proper AI product should aim to level the differences in user capabilities. In the past, we created various process management office software to provide better recording and information synchronization tools, but this heavily relied on employees' professional competence, management skills, and execution ability. In other words, for most companies and employees lacking these capabilities, office software merely turns handwriting into typing—nothing more.
But AI has the power to solve this. It can internalize high-value business rules into the product. Previously, people had to manually connect different tools; now AI can autonomously make those connections. Employees with weaker skills can rely on the product's built-in value system to achieve efficiency leaps, while top-tier employees can free themselves from tedious administrative tasks and focus on higher-value goals.
The true power of AI applications lies in their ability to greatly reduce friction between the product and the user, dissolving the heavy cost of user education into automated logic and granting everyone a baseline level of capability. This is the value that AI products should carry.
Returning to the ASR application I whipped up for my friend. Simply converting speech to text won't shorten the distance between him and his work goals. The Feishu subscription fee may sting, but a free open-source solution won't save him mental energy either.
A truly valid AI product should be the embodiment of "meetings must involve discussion, and outcomes must be useful." Only when high-value business rules are internalized as the product's core will we stop being mere movers for model vendors.
What we should deliver to users is a set of excellent capabilities that can run autonomously, not scattered efficiency tools one after another. Only then can we truly find the value connection with users in the AI era and redefine our own value anchor.