Everyone building Personal Agents likes to talk about itinerary planning.
At product launches, the CEO stands in the middle of the stage, holds up a phone to demo the AI personal assistant’s capabilities, and inevitably has to do an itinerary planning demo. Indeed, this kind of task includes booking flights, booking hotels, checking routes, buying attraction tickets, and various other small tasks of different types, and can fully demonstrate the AI’s ability to understand needs, break down tasks, and call tools.
So many people think that building an AI personal assistant means making something to complete this kind of task. As a result, all sorts of latecomers have sprung up like bamboo shoots after rain to help people plan itineraries. But the problem is that the vast majority of users do not have that many needs that require “planning.”
For them, an AI personal assistant is not much different from a larger search engine. They will not study what capabilities are inside, or which tasks that don’t exist in daily life can be handed over to it to complete.
They only need the AI to answer questions and give suggestions.
A wishing well with limited capabilities.
In the past, when we built tools, we first needed to define the specific problem the tool could solve, and then define how to solve it. We wouldn’t build an image-editing tool that can translate in real time, nor a whole set of document plugins that can make tea.
When tools were given to users, they didn’t need to study the technology to know how it related to them.
But with today’s open-ended AI personal assistants, a blank chat box plus an interaction format where every question gets an answer makes users feel like it is a deity that grants every request.
Before AI, we called this kind of interactive experience fortune-telling, and it was feudal superstition. So what do people use it for? To make wishes.
Help me get rich, help me get a raise, and best of all, deposit one million yuan into my bank card. Since you say you can do anything, users will just say what they want. They are under no obligation to first study your capabilities and then rewrite their wishes into tasks you can execute.
Developers are all waiting for a perfect user, one who understands AI’s capability boundaries very well and also knows the logic by which AI calls tools. But such users are extremely rare; for the vast majority of others, the only words they took in were “can do anything.”
Infinity equals zero. If you let users say whatever they want and make whatever request, they will immediately come to you with “may you be prosperous,” and in the end they find that you “can’t do anything.”
Give the product a purpose, and set AI free.
Therefore, to build AI products seriously, you must preset the purpose for users and use UI to restrict users from engaging in feudal superstition. This is a product interaction design issue; even if you swap in a stronger large model, users will not suddenly know how they should use it.
Previously, there were two routes for AI products: one was autonomous planning by AI, and the other was hard-coded workflows. Steps that used to need to be arranged by humans in advance can now increasingly be decided by the model itself. Coze’s shutdown has already proven the failure of the second route, but open autonomous planning does not mean that a blank chat box is a reasonable interaction form.
The more AI can decide how to do things, the more the product needs to clarify for users what is to be done.
Only when the goal is clear and the delivery standard is explicit can the model complete the task better. It needs to use these to plan the execution path, judge intermediate results, and determine whether the results meet the delivery standard.
But ordinary users do not have the ability to design goals. They only know that they have encountered a problem, feel uncomfortable when doing certain things, or are helpless when facing some problem.
Defining the problem, clarifying the goal, and setting the delivery standard inherently require professional ability. A fully open AI assistant is equivalent to kicking the pressure of expertise back to the user. Therefore, the product needs to determine in advance the scope it can serve and agree with the user on a clear deliverable, and finally let AI decide the intermediate process. This, I believe, is where the value of AI assistant products lies.
Only when the purpose and delivery are constrained does AI have a basis for finding the optimal solution.
Finding the infinite within the finite.
In our lives, there are many small problems that have not been well solved.
For example, when buying shampoo, although many brands claim that their products are silicone-free, after using AI to read the ingredients of more than 10 brands of shampoo, I found that many of them actually contain ingredients similar to silicones. Ordinary people don’t understand chemistry. Without AI’s help, they would have no idea how to buy something suitable for themselves.
For another example, I need to secure my knives in the kitchen. In the past, I would go directly to look for a magnetic kitchen knife rack, but now I can let AI work around this goal and find a solution on its own. It can study what conditions need to be met to secure knives, then search across product categories for options, and finally may discover that a certain hardware magnetic tool rack can also meet the requirements and is even cheaper.
The value of a product lies in packaging these problems and turning them into services users can directly use. When I need to choose a shampoo, I should not have to gather ingredient lists from a dozen brands and have AI analyze them one by one and compare them repeatedly. The product should complete this research for me according to my needs and deliver a suitable choice.
Reading ingredient lists, checking specifications, and finding alternative products are all part of the delivery process and can be organized by AI itself. Users should not have to reinvent the tool every time they encounter a problem.
Rediscovering the last mile.
Many developers are very good at solving high-value problems. But in users’ daily lives, there are plenty of trivial little things.
Integrating processes such as booking flights and booking hotels into one task is showing off, not really solving users’ problems. Far more people need to book flights than need to plan itineraries.
Moreover, anyone who has booked a flight knows that different departure times, different airlines, and different booking dates all lead to different ticket prices. Ordinary people need to compare prices and do their own research when booking tickets.
Handing this research over to AI and finding a more cost-effective flight ticket for the user is already a delivery worth fulfilling.
New technology lets us rediscover the last mile. The hassles that used to be left to users are all worth turning into services that products can deliver.