Can UX be personalized for each user to achieve the highest KPI rates? A conceptual framework to integrate machine learning and generative AI into UX workflows.
Decentralized UX framework diagram cover.
Status quo of UX workflows
Average UX workflows start with defining KPIs and end with a design deployment. Many stakeholders are included in the process, sprints are constricted, teams are unaware of other teams’ work, and budgets are inadequate, reducing the delivery quality. Design outputs are restricted since only a few designs have been tested with a few user groups.
Ultimately, only one least-worst design is expected to fit each user and achieve the highest success rate, and naturally, the objectives are mostly missed, causing an endless iteration cycle and profit erosion.
The output of UX workflow depends on the stakeholders’ inference. UX teams conventionally decide what data should be tracked for what purpose, what designs should be tested, and which is best for the users, meaning the outputs are restricted by human intelligence and capabilities.
What is “Decentralized UX”?
Decentralized UX framework is a proposal to achieve the highest KPIs by providing the highest satisfaction with the highest personalization to each user. It proposes to eliminate the artifactual handicaps of the centralized UX framework. Fundamentally, it utilizes machine learning and generative AI to execute the process:
Desired KPIs are entered into GPTs as prompts.The machine learning model analyzes user and product data in real-time using user research methodologies.Generative AI acquires the interpreted-supervised data and starts to generate complete personalized design outputs for each user.Generation includes user flow, wireframe, content, UX writing, UI design, design system, and development.The design alternatives are released for testing and are analyzed by the machine learning model again.The design is iterated until the highest KPI rates are achieved, and each user is satisfied by personalizing the product in endless iterations.Thus, user experience is decentralized by taking over data analysis, interpretation, generation, and development for each user.Decentralized UX framework process diagram. (Input > ML > GenAI > Output > User)
These steps are currently operated by product, data, design, and development teams manually. Decentralized UX workflow proposes combining these steps to reduce operational costs. This conceptual framework takes the existing ML and GenAI technologies and related tools as a base to demonstrate and exemplify the idea.
Generative AI: The new designer
Generative AI (GenAI) is a subset of artificial intelligence that creates new data and contents as output that depends on the given input, data, or dataset. The outputs can be in several formats: text, image, audio, video, code, etc. Generative AI uses machine learning and neural network techniques to process the prompts and generate outputs that meet requirements. Currently, several tools provide these functions:
ChatGPT: AI language model for conversations.Github Copilot: AI-powered code completion and generation.Midjourney: AI image generation from text.
The process of generative AI starts with data collection, continues with data preparation to train the GenAI model, and ends up with evaluation and fine-tuning. The key to having precise outputs is having precise datasets. Ready-to-use datasets could be used to train the model. If the data doesn’t exist or is unstructured, machine learning comes on the stage at this phase.
Make designs – AI beta promotional image by Figma.
Machine learning: The new researcher
Machine learning (ML) is a subset of artificial intelligence that includes deep learning (DL) as a subset, which refers to an algorithm that learns from user-generated or system-generated data. Unlike traditional software, it generates patterns and predictions as output variables. The output variables depend on supervised and unsupervised learning.
Supervised learning makes ML predict output variables by identifying a structured dataset, which should be JSON, CSV, text, or binary formats. Machine learning can execute regression and classification tasks this way. Regression is a prediction model in which the output variable can be numerous and continuous -email spam detection-. Classification is a prediction model in which the output variable is categorical -house pricing based on location- with sub-models as binary, multi-class, and multi-label, categorizing the output variables into one, two, or multiple groups.Unsupervised learning makes ML predict the output variables by identifying an unstructured dataset such as free text, image, video formats, etc., and clustering the related ones together -customer segmentation- or associating the related data -anomaly detection-.
Several data models, data processing, and data flow methods are applied to the trained data to utilize it for services. These principles, methods, and features can already be executed on several tools, which are called MLaaS (Machine Learning as a Service):
AWS Machine LearningIBM WatsonGoogle Cloud MLMicrosoft Azure MLOpenAITensorFlow
Graph models (GM) is one of the visualization techniques for ML models that ease understanding and operating the models. GMs are the optimal way to manage the ML models for UX people in the future. As an example, Google released a new graph model, Model Explorer, for large models.
Model Explorer promotional image by Google.
While GPTs are performing very well already, we can foresee some probable usages for the future development of the current product development and UX workflows since we are already doing the same thing what machine learning does.
With ML’s capabilities, users can be continually analyzed; user flows, interfaces, and content can be designed simultaneously for each. This can increase user satisfaction and improve business KPIs.
Continuous UX research with machine learning
UX researchers track, collect, and analyze the users’ and products’ data to interpret them. In this regard, machine learning and UX research have rather relations and similarities to what they execute, and they can be automated:
When to Use Which User-Experience Research Methods, A Landscape of User Research Methods Matrix by Nielsen Norman Group
Simultaneous quantitative research
Quantitative user research methods are currently operated by several tools, such as Hotjar, Firebase, and Mixpanel.The only requirement is to set up what data you need to gather, which is basically “prompting”. The outputs are mostly “structured datasets” -CSV, binary, text, etc.- which can be utilized for “supervised learning”.
Executing quantitative research methods in real-time would provide simultaneous input by tracking the users’ patterns to estimate and provide the most personalized experience design during the funnel. The concept of “A/B testing” may be “A-Z testing” since endless options can be tested in real-time.
Autonomous qualitative research
Qualitative user research methods are executed by a UX researcher since they require interaction with real users. The outputs of the qualitative research are generally “unstructured datasets” -free text, video, audio, etc.- that can be utilized for “unsupervised learning”.
Typeform recently released an AI-powered, real-time video chatbot tool, VideoAsk, for user interviews. It can ask users questions in different formats -video, audio, text- and collect the answers.
Research shows that Generation Z tends to interact with AI bots rather than real people, which means Synthesia Personal Avatars maybe prevalent in the future.VideoAsk promotional image by Typeform.
These processes may generate too much important data about each user that must be stored securely and properly for the organization. These platforms are called Reliable Data Management Systems (RDMS):
IBM Db2Microsoft SQL ServerMongoDBOracle Database
Decentralized UX framework can provide a better understanding for each user than running usability tests with limited number of users or limited pre-defined personas. Qualitative or quantitative research outputs can be converted to a dataset to train the machine learning model, and GPT can analyze it and provide the best insights and design suggestions for the users to serve the most personalized experiences.
Since neither a single UX team nor a development team can deliver endless design options, generative AI must be utilized.
Personalizing UX with generative AI
Personalization is one of the key factors of user satisfaction. Digital products that will be generated by generative AI with complete personalization will undoubtedly achieve success if they meet accessibility, behavioral patterns, or content expectancy. McKinsey research shows that personalization can increase revenue by up to %40.
Generative content design
Today’s successful products, such as Spotify, Netflix, and Amazon, use personalization algorithms to achieve user satisfaction. For example, Netflix has an algorithm called “Artwork Personalization” that automatically generates content’s covers according to the users’ watching history data:
Netflix Artwork Personalization feature shows two different covers of the same movie for two different users. The first cover is for the user who previously watched Uma Thurman, and the second is for the user who previously watched John Travolta.Copywriting, video, and audio content can be personalized according to the user’s behavioral patterns, like how ChatGPT can do UX writing.
Generative UX design
With the trained data gathered during the UX research phase with ML, generative AI can generate adaptive and optimal wireframes and flows for the users. For example, DiagramGPT can generate diagrams that visualize the user flows according to the prompts entered. It is the response to the prompt of “Generate a user flow of sign up flow from onboarding to homepage”:
DiagramGPT’s output for sign-up flow diagram.
When a user flow is defined, it can be converted to a wireframe before being converted to a high-fidelity user interface. Several wireframe generator tools exist, one of which is WireGen, which can convert a prompt into an expected wireframe.
WireGen Figma plugin promotional image.
As recently released, Google Gemini can analyze and understand what’s on the screen, which will provide a better context understanding of the digital product and user’s actions. This cycle can repeat itself till the best outputs are applied to achieve the user’s highest conversion rate and endlessly and autonomously iterate it.
Generative UI design
When the required flows, wireframes, and required components are defined in the UX design phase, generative AI can produce the interface precisely. Nielsen Norman Group called this concept “Generative UI and Outcome-Oriented Design”.
Generative UI and Outcome-Oriented Design informative image by Nielsen Norman Group.
Decentralized UX framework is not that far from now since there are already several tools that generate UI designs by prompting:
Figma AIFramerGalileo AIUIzard
However, these tools currently require human input to generate outputs. They analyze existing designs on the internet and provide you with the most related output without following the design system and development requirements. We can utilize AI to maintain design systems and develop generative UIs.
We need an AI tool that gathers and analyzes the data and inputs required prompts to provide the most precise output.
Design system generation and maintenance with generative AI
If a design is not developable, it’s not a solution. With a consistent design system, AI-generated wireframes can be developed by AI instantly. Several open-source design systems, such as Google Material Design 3, Microsoft Fluent, and IBM Carbon, already exist. UI kits, component libraries, and patterns have been well-defined and are ready to develop open-source.
Generative design system
Since AI collects data all around the internet, it already knows how to develop a component. For example, BlackboxAI, an AI-powered coding tool, generates HTML, CSS, and JS scripts as a response to this prompt, “Code a side menu with header includes headline and collapse button, 4 level hierarchy and a footer includes a primary and secondary button”:
https://medium.com/media/390182c6264db39cb7981ffd37f9cf46/href
Additionally, there are several design system generator Figma plugins, which utilize AI, and can generate design systems, including color palettes, typography scale, component libraries, styles, and tokens:
AtomicDesign System BuilderDesign System Generator
Maintaining design system
In this regards, generative AI can be expected to generate effectively functioning design system. Since WCAG includes the accessibility standards for component, text, and icon sizes and color contrast ratio in numeric values, generative AI can apply it according to the users’ requirements in real-time. There is a tool currently exists that can generate, maintain, and audit the design systems with artificial intelligence, Motiff:
AI-powered design system tool promotional image by Motiff.Figma recently acquired Dynaboard, a low-code web-app builder. It provides the design, development, and release thatare compatible with several ML tools.
There are several tools for no-code website, application, software, and design system development:
Bootstrap StudioDynaboard (Acquired by Figma)SquarespaceTailwind GeneratorWebflowWix Studio
Design system management is easy with tokenization and complex at the same time. There are several tools for token management, and Token Studio is the best example. Its TokenFlow feature uses a graph model that can be operated by ML/AI in the same way.
TokenFlow informative image by Token Studio.All the functions of the tools above can be implemented in a tool like Penpot, an open-source design tool that provides development teams with a self-hostable infrastructure.
Large Action Models (LAMs) can decentralize UX
There is a recent advancement in artificial intelligence, Large Action Models (LAMs). It is a deep learning model that can do reasoning, automation, decision-making, complex and real-time interactions, and taking actions. This model was also used in Rabbit OS.
Decentralized UX workflow from KPI definition to user research, UX design, design system, and development can be actualized, implemented, and automated with LAMs. Since this model is still being researched and developed, it is barely known, but there is no doubt it will be soon.LLMs and LAMs comparison table by Viso.ai.
Decentralized UX business value
Product development workflow requires many artifactual implementations for project planning, research, design, development, testing, and iteration. The focus switches from the KPIs to operational issues during the implementation phase. Decentralized UX framework provides:
Personalized products provide achievement of the highest KPI rates.Adaptive user and customer experiences increase success metrics such as NPS, CSAT, and CES.Products and user data are analyzed all the time, which will provide endless insights for business opportunities.Operational costs are reduced since miscommunications and technical restrictions are eliminated.Automation reduces the duration of roadmaps.Scalability allows the implementation of new business ideas.
Integration is the key
Humans’ needs are changing, products are changing, and technologies are changing. Concern and resistance have always been part of it. For example, in 1986, math teachers protested the use of calculators in school education programs because they were concerned about their profession. However, the invention of the calculator is just an ease calculation, and math teachers still have their jobs.
Abstraction of products and experiences
Now, in the 2000’s, physical products has turned into digital products; they are on the screens, some are just sound, like virtual AI assistances, and some are brain chips that are controlled by thoughts.
In the time the product development works are abstracted by delegating them to ML and GenAI, product developers will be able to focus more on the functionality of the products, user benefits, and business profits.
The invention of the calculator sped up achieving results; the abstraction of products and experiences will bring the same benefit as well.
Decentralizing UX with ML & GenAI was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.
