AI agents, part 3: Adaptive designs optimize development and user experience
AI agents are changing digital product development: individual adaptations are bringing efficiency and user focus closer together.
(Image: sdecoret/Shutterstock.com)
- Thomas Immich
AI agents are increasingly taking on tasks along the entire digital production line – from code generation to process automation and continuous improvement. While part 1 of this article series outlined the fundamental shift caused by agentic roles and part 2 traced the change from product to production, this final part now addresses the question of how efficiency, individualization, and quality can be reconciled. Adaptive interfaces, simulation-supported variant creation, and automated testing meet human frameworks – and open up new scope for truly human-centered product development.
Individual customization
When you think of customization, minimalism is not necessarily the first thing that comes to mind. However, from a human-centered perspective, every product feature for which there is no need for use is a cognitive burden and a digital threshold. Adapting software to the user therefore has a lot to do with “omission”. The reduction of features becomes the actual goal.
Before the AI era, modern software development already had several strategies for equipping a digital product with more or fewer features for certain target groups.
However, there are classic problems that remain unresolved in the long term when applying these strategies: if features are reduced during program runtime using feature flags, the source code must contain many conditional program parts that need to be kept track of. This has a negative impact on the memory requirements and performance of the program.
The use of feature flags essentially contradicts the pull principle of lean manufacturing. This principle calls for demand-oriented, resource-efficient production in line with specific demand. If you transfer the feature flag approach to industrial production, it would be like building a car with a 360 hp engine, even though only a small proportion of customers need this power. For the majority, the power would then have to be artificially throttled – an approach that wastes both material and energy and runs counter to the principles of lean production.
As absurd as the example sounds, this approach is actually chosen in some parts of industrial production because the additional costs of a more flexible production line would exceed the additional costs of castrated components. Personally, I consider this approach to be extremely questionable from a sustainability perspective, as the corresponding raw materials still have to be used up and disposed of after the "end of life" of the product. And I am also critical from a human-centered perspective, as users could rightly get the feeling that they are paying for something they already own, which could have a very negative impact on the user experience of the product.
Reduction through build variants
An alternative strategy is not to include unneeded functions in the final product in the first place but to omit them during the build phase. This build-variant approach is already used in software development today via complex delivery pipelines. It is of course no coincidence that the term “pipeline” originates from industrial production.
Building and managing effective and flexible delivery pipelines is complex and becomes increasingly challenging with the number of product variants. Assuming the ideal UX case, i.e., 1:1 personalization of the software for each individual user, a separate product variant would have to be created for each of these users, containing only those features that are individually relevant – a set of functions that will unfortunately even change in the course of use as new needs develop with increasing experience in using the product. Power users therefore have completely different requirements than newbies.
(Image:Â ipopba/stock.adobe.com)
The Product Owner AI Day by iX and dpunkt.verlag on November 6, 2025 will show you how you can integrate AI into your work as a product owner, product manager or with your team – from discovery to rollout. Tickets are available at an early bird price.
In extreme cases, every single user has their own individual set of needs that have been adapted over time and therefore also require their own corresponding set of features.
In the distant future, we will certainly see how AI agents can autonomously tailor hyper-individualized products for individual users or even adaptively “update” user interfaces during operation. But what are the possibilities if such hyperindividualization is not yet an option?
Target group simulation as a key agency skill
A product owner must decide where the needs of different users are so congruent that they can be combined in one product variant and where they differ so much that it is better to use different variants. This analysis is very time-consuming because those involved have to repeatedly run through the most diverse combinations.
Videos by heise
If you view the software development process as a production line, it is therefore worthwhile in the AI-augmented future to provide a station where AI agents find out the minimum number of product variants required to make the maximum number of users happy.
This is where AI agents come to the rescue in the form of personas, i.e., virtual users. As shown in the example of the UX Therapy AI podcast in the first part of the article series, AI persona agents can talk to each other directly to find out where common and conflicting needs lie.
This type of simulation of plausible conversation outcomes is a complex process that delivers emergent results and is therefore difficult for a human to play through. The possible outcomes of a conversation cannot be calculated using a static formula, as they are the result of complex LLM operations.
(Image:Â Microsoft)
Using AI agents as actors within an LLM simulation (so-called sims) is now such a promising approach that Microsoft has launched its open-source project called TinyTroupe.
In a further step, AI agents could not only suggest the different construction variants but also implement them directly based on a reference variant. As flexible code generators, they are therefore part of the automated software production line, similar to humanoid robots that will probably soon be used in the production of goods. It is therefore doubly worthwhile to rethink the optimization of product variants in the age of AI agents.
