Apple study reveals expectations for AI agent Apfelpatient

AI agents are increasingly taking over tasks on the computer: they research, book, compare products, or navigate complex websites. While much attention is focused on the capabilities of these systems, one question often remains in the background: How do people actually want to interact with such agents?

A recent study by Apple addresses precisely this topic. The focus is not on the technical optimization of AI, but rather on the design of the user experience. The results provide concrete indications of which interaction patterns work and where current systems fall short of expectations.

The study is titled "Mapping the Design Space of User Experience for Computer Use Agents". A team of four Apple researchers analyzed which UX patterns are used in existing AI agents and how real users experience them.

Apple notes in the study that the market has invested heavily in the development and evaluation of AI agents. However, aspects of the user experience – particularly how users want to interact with agents and how corresponding interfaces should be designed – have not yet been systematically investigated.

To close this gap, the study was divided into two phases. Phase one involved analyzing existing systems and developing a structured taxonomy. In phase two, these concepts were tested practically with real users using a so-called Wizard of Oz study.

Phase 1: Analysis of existing AI agents and development of a taxonomy

As a first step, the Apple researchers examined nine existing desktop, mobile, and web-based AI agents:

Claude Computer Use Tool
Adept
OpenAI Operator
AIlice
Magnetic UI
UI-TARS
Project Mariner
TaxyAI
AutoGLM

Subsequently, eight practitioners were consulted. These were designers, engineers, or researchers working in the fields of UX or AI at large technology companies. The aim was to validate the analysis and supplement it with practical perspectives.

Based on this, the research team developed a comprehensive taxonomy for the user experience of computer usage agents. This included:

Four main categories
21 subcategories
55 example functions

The four main categories are:

1. User request

This category examines how users enter commands. This includes, among other things, natural language, structured input, and hybrid forms.

2. Explainability of agent activities

The focus here is on what information the agent provides during their work. This includes, for example:

Presentation of a planned procedure
Display of intermediate steps
Communication of decisions
Making errors visible

3. User control

This category addresses the question of how and when users can intervene. This includes:

Interrupt functions
Confirmation dialogs
Correction options
Adjusting decisions during execution

4. Mental Model and Expectations

This section focuses on how users understand an agent's capabilities and limitations. This includes transparent communication of:

Skills
Restrictions
Uncertainties
Error messages

This framework covers both interface elements and communication strategies and forms the basis for the second phase of the study.

Phase 2: The Wizard of Oz Study

In the second phase, Apple tested the identified UX patterns in a practical study with 20 participants. All participants had prior experience working with AI agents.

Study structure

Participants interacted with a simulated AI agent via a chat interface. Simultaneously, they were shown an execution interface where the agent performed actions on websites.

The crucial point: The agent was not a true AI system. Instead, it was controlled by a researcher in an adjacent room. This researcher read the participants' text instructions and manually executed the corresponding actions using a mouse and keyboard.

This method is known as the "Wizard of Oz". It allows for the study of realistic interactions without the need for a fully developed AI system.

Task

The participants were to use the agent to work on two different tasks:

A task related to a holiday apartment
An online shopping task

Each of these tasks required the execution of six functions. It was intended that the agent could intentionally fail to perform certain actions correctly. Examples:

Getting stuck in a navigation loop
Selecting the wrong product contrary to the instructions
Deviation from the original plan

During the process, participants could stop the agent at any time using a pause button. In this case, the message "Agent interrupted" appeared in the chat. After a task was completed, "Task completed" was automatically displayed.

After each session, participants reflected on their experiences and suggested improvements or additional features. Researchers also analyzed video recordings and chat logs to identify recurring patterns in user behavior, expectations, and problem areas.

Key findings of the Apple study

The analysis revealed several clear trends.

Transparency yes, complete control no

Users want insight into the activities of an AI agent. They want to understand what is happening and why certain decisions are made.

At the same time, there is no desire for complete microcontrol. If every single step had to be confirmed, the added value of automation would be lost. The balance between transparency and autonomy is crucial.

Expectations vary depending on the context

The desired behavior of an agent depends heavily on the usage scenario.

For exploratory tasks, more options and insights are desired.
For familiar, routine tasks, a more direct process is preferred.
The less familiar users are with a user interface, the greater the desire for transparency, intermediate steps, explanations and confirmation pauses – even in low-risk scenarios.

Increased need for control when there are real consequences

When actions have real-world consequences, the need for control increases significantly. This includes:

Purchase transactions
Changes to account or payment details
Contacting other people on behalf of the user

In such situations, users expect clear confirmation and transparency.

Trust is fragile

Trust is quickly lost when an agent:

Tacit assumptions are made
Ambiguous choices interpreted independently
Deviates from the original plan without prior notice

For example, if an agent finds several possible options on a website and selects one without asking, users often react by abandoning the process and demanding clarification. This becomes particularly critical if the decision could lead to choosing the wrong product. Unclear or uncommunicated decisions create unease, even if the objective risk is low.

Significance for the development of AI agents

The Apple study clearly shows that technical performance alone is not enough. The following factors are crucial for agent-based systems:

Context-dependent transparency
Situation-appropriate control options
Clear communication of uncertainties
Supporting an understandable mental model

For app developers who want to integrate agent-based functions, the study provides concrete guidance. Not maximum automation, but balanced interaction determines acceptance and trust.

Designing AI agents successfully, not just optimizing them

Apple's research shifts the focus from pure AI performance to the design of the interaction. People expect neither complete control nor complete opacity. What's crucial is a comprehensible, context-dependent collaboration between user and agent.

The study makes it clear that UX is a key strategic factor in the age of AI agents. Anyone developing agent-based systems must understand how expectations arise, how trust is built, and how quickly it can be lost. (Image: Shutterstock / Pniti_Studio)