Due to the failure of portion control, this has become a three-part series. I am sorry. The last one is the fifth step, user retention and churn prevention. So let's get started.
Step 5: Retention (Retain users and prevent churn)
Personally, I think that stages 1 to 4 are changing in a very random way. It's not a matter of creativity, but because there are too many products and services being released. The biggest issue is how to show it with design elements and production.
Conversely, the more similar functions this provides, the easier it is for users to leave. And what I feel while studying is that users have no attachment to the 'generated data'. They do not think that it is their own precious data. Easy come, Easy go applies here as well. From the user's perspective, since they have achieved great results with relatively little effort, they do not feel attached to this data or think it is precious. It was a long story, but in the end, if a fancier service or a cheaper service comes out, they immediately leave. How do AI services manage this? Let's take a look at them one by one.
Specialization
This is a strategy used in many places. Regardless of the model used, they talk about what they are good at. For example, character.ai focuses on creating characters in the form of personas and having them talk, and Tensor.art markets itself as providing a more specialized image generation service. Typeset.io, which targets researchers, is also a good example.
When writing a paper or reviewing a paper, I can't live without this anymore.
One of the benefits of specialization is that it allows you to justify the cost. As I've said in other blog posts, considering that services like OpenAI and Google usually charge $20 per month, specialized services can be locked in even if they charge more than that when users are irreplaceable or the service is deeply involved in their work. The best example for me is LBOX, a legal service in Korea . LBOX is a legal AI service that provides services to lawyers, judicial officers, and law school students about court decisions and litigation information, but it charges more than $55 per month and does not cause user churn.
Personalization
Personalization is a topic that comes up a lot in the AI theme. In particular, services that have existing information, or in the education field, they are taking the approach of providing customized curricula. There is no case of fully implementing a method such as something optimized for you yet. The reason is simple. It is not easy to obtain permission to use the user's data, and it is too expensive to fine-tune it for a customized user.
This is where Notion, when it first introduced AI, cleverly acquired Workspace permissions, and places like Salesforce and Google Workspace have recently introduced such services. Ultimately, data ownership is a pattern that controls how AI models remember and use users' data. This is an important feature that balances privacy and AI model improvement.
Typically, this is located in the user or company settings, and is implemented as an on/off toggle with a brief description of how to improve the model. However, implementations vary. Most services default to 'on', but companies that take a user-centric approach, like Figma, default to 'off'.
There are also differences between paid and free users. In many cases, opting out of data sharing is only available on premium plans. For enterprise accounts, these settings tend to be part of the admin settings, not individual users. In simple terms, if you provide your data, you can get something tailored to you, but if you don’t agree, it feels a little difficult.
We've actually experienced this before, when iOS collected user information and behavioral data for things like app tracking and marketing and advertising exposure. It's easy to think of it as replacing seeing personalized ads with your own personal AI.
Token Optimization
Token layering is a technique that deliberately combines tokens when a user constructs a prompt for AI to finely adjust the AI's understanding and response direction. In simple terms, when AI processes text (speech, writing), it does not understand the entire sentence at once like a human, but breaks it down into small words, and then analyzes it more deeply through several stages. This process is called token layering.
This allows you to build prompts like building Lego blocks, allowing you to more accurately convey your intent. Token transparency means that the tokens used by the AI to generate responses are made public to the user. This allows the user to understand the AI’s ‘thought process’ and write better prompts.
As easy as it may seem, it’s hard. Even Google has a hard time fully controlling this token layering.
There are many ways to leverage these technologies to enhance user experiences. For example, web-based interfaces like Adobe Firefly provide a palette where users can freely write prompts while easily adding tokens for style, structure, references, etc. This is an intuitive implementation of token layering. Services like Google AI Overview and Perplexity automatically generate follow-up questions after the initial prompt to collect additional tokens, thereby better understanding the user’s intent.
A good example of token transparency is Midjourney’s /describe function, which exposes the tokens used to generate an image, allowing users to understand the AI’s interpretation and correct it if necessary. Audio generation services like Udio include relevant tokens in the metadata of the generated file, making it easier for users to search and generate similar audio.
These features help users understand how AI works and achieve better results. They also stimulate users’ creativity and increase the efficiency of AI use. As a result, this can be an effective way to increase satisfaction with the service and encourage continued use. However, when implementing these features, it is important to consider the user’s learning curve and introduce complex features gradually.
Provided in solution form
It is most often provided on the B2B side. It is a service because it is provided in the form of an enterprise. In this case, it is usually processed physically in a data center unit or an open source model is tuned to fit the client. In fact, many SI companies are currently providing on-premise models using LLaMA3 or Qwen. As always, it is necessary to consider in advance whether to do the work in a completely outsourced form or whether we build and operate it ourselves and only receive assistance.
In this case, if you go for complete outsourcing, even if the speed is fast, maintenance or adding functions may become difficult, and in the case of self-building or tuning, there are advantages and disadvantages such as being able to secure scalability or ownership in the long term even if the initial development speed is very slow. Recently, many solutions in the form of MLOps are also coming out.
If the existing referral (recommendation system) was a method of giving tangible goods, this is a method of issuing internal securities (?) and credits and making them use them. In fact, you can see that it is a continuation of the method frequently used in existing SaaS marketing. For example, Gamma.app, a service that creates slides , provides 200 credits according to the number of invitees. Since about 40 credits are used when creating a slide, it is quite a lot.
However, the limitation of this credit system is that too many credits are accumulated or a sewer strategy is required. Personally, I remember that Notion did the same thing in the past and eventually changed it to an affiliate form. However, it is effective as an initial growth strategy. It is also a good reason to keep users and increase.
I have roughly summarized it like this. Since it was organized from the perspective of planning and UX, I think it would be a better study if the perspectives of marketing and developers were added. Please feel free to share it, and if you are interested in studying, please send an email to haebom@kakao.com .
Subscribe to 'haebom'
📚 Welcome to Haebom's archives. --- I post articles related to IT 💻, economy 💰, and humanities 🎭. If you are curious about my thoughts, perspectives or interests, please subscribe. Would you like to be notified when new articles are posted? 🔔 Yes, that means subscribe. haebom@kakao.com
