Last year, I built a chat-based calendar agent that allowed me to check my schedule, create entries, and align my calendar with those of my colleagues. I imagined it would feel like having a personal assistant. I hoped I just needed to throw over who I wanted to meet, and the scheduling would be magically done.
I built it out… It worked fine.
I started using it… It took me a week or so until I never touched it again.
I quickly returned to the traditional, manual GUI process.
You may have had similar experiences. My question is as follows: Is it just a matter of time? Is the technology “just not there yet”? Or is a chat interface an anti-pattern for human users? Maybe a chat is just not a great interface for a calendar. Perhaps a chat is a terrible interface for most things.
A well-designed GUI is an information abstraction layer.
A second example: You want to claim a return in an online shop. Which customer service experience do you prefer? An agent (human/AI/hybrid) via chat or phone vs. a well-designed self-service GUI? I asked among approximately 20 colleagues and friends across age groups. 18 mentioned they would choose the visual UI. A well-designed(!) GUI is just more effective. It saves communication overhead, time, and energy. Proactive visual feedback is fast and efficient for this case.
One more example: Imagine a car without a dashboard, but only a conversational interface. Maybe you would remember asking the car for the speed you’re going at from time to time. But would you remember asking for the gas level, tire pressure, and the need to refill the oil…? While driving, we make hundreds of tiny decisions per minute. The car’s graphical user interface (aka. Dashboard) is our trustworthy basis for these decisions. Only the graphical user interface of the car makes driving accessible for most people. Driving a car with a conversational interface would require us to have much deeper knowledge about the car’s workings and potential points of failure before operating it.
A chat is not a great user interface for … most things!
A counterexample: Many executives have personal (human) assistants. Conceptually, these work just like my calendar bot. The assistant usually manages the person’s organizational overhead, allowing the executive to focus on their core work. Why does this setup work in contrast? Is it me who did not articulate my “prompts” to my assistant concisely enough? Managers at that level should usually be communication talents. So surely most of us could learn from their “assistant prompting” skills. But can we expect the same preciseness of communication from our users?
To communicate queries effectively we need our users to…
… firstly, know exactly which problem they need us to solve.
… secondly, have a clear vision of the type of solution they are looking for.
… finally, formulate this desire in a prompt that our AI can understand and work on.
Expecting our users to write a pitch-perfect prompt is like asking the average person to control their computer via the command line. In theory, the tools are more efficient, but in the wrong hands, they are completely ineffective.
Most of your users will need visuals to hold onto. Graphical user interfaces make technology accessible to the masses. Specifically, they do this by pre-aggregating information. Ideally, the aggregated information provides a solid “decision basis” to the user.
Social and transactional conversations
According to this article by Clark et al. , humans converse for two reasons: Socializing and transacting.
Social conversations are about finding common ground, making memories, and building trust. They claim that many levels of human-to-human social connections exist. Building trust and common long-term memories are key to making them.
In transactional conversations, researchers reported that active listening and trustworthiness on a functional level prevail. That means transactional conversation partners are expected to remember the important facts, keep them safe, and follow our instructions clearly and transparently. Trustworthiness and reliability are key!
AI agents in 2025 will largely focus on transactional purposes. They collect information and accomplish tasks for us. Did you ever feel a truly personal connection to ChatGPT or Gemini? Why not? Following the study, we are missing the long-term connection and memories with the agent.
I’m convinced that conversational agents technically could build a personal connection with us (and vice-versa). I believe the agents usually don’t have the context to do so. Most chat agents have a fraction of the information about our lives that would be necessary to know to feel close to us or make us feel close to them. So it is rather a data quality issue than a technology issue. If your agent is supposed to build a connection to your user, this is the challenge you need to crack.
Transactional agents on the other hand “simply” need to deliver an exceptional experience in whatever they are built for. If a conversational interface costs time (compared to a self-service GUI) instead of saving it that shatters trust and builds frustration instead of automation.
Designing the human-ai interface of the future?
Pressure is on for AI agents and their designers. Their systems should either build social connections or become the best solution to the problem they are solving.
Customized voice agents are a great attempt to build agents that create connections. ElevenLabs leads this space with voice agents that easily integrate and clone personal voice tones. Combined with more freedom to design long-term memory and manage session context this is a potential killer combo. ElevenLabs agents are still missing this long-term part. An architecture combining RAG for fact retrieval with a dynamically created user profile for directly recallable context will help solve this soon.
For transactional purposes, conversational interfaces will not beat the information abstraction that GUIs offer for a long time. However, hybrid interfaces have massive potential. Let’s drop the belief that LLM interaction requires a chat. Quite the opposite! LLM interaction should happen at the click of a button, integrated into a GUI. The prompt is pre-defined in the background thus the user does not have to face the prompting complexity but benefits from LLM agent intelligence.
Great examples of hybrid agentic UIs are:
- Gemini Deep Research research plan
- Gemini Workspace to summarize doc content and new comments at the click of a button
- Cursor chat to add and remove documentation code files, and the full codebase flexibly
So what‘s left for AI Chat agents?
If you are reading this you are likely a tech-savvy builder. I won’t need to convince you of the value and impact that AI has and will have/on our lives. However, looking past our bubble of technologists, Chatbots are crucial building blocks of broad AI adoption. The public measures AI progression by chatbot performance! Thus, without solving chatbots the general public will not believe that AI is (being) solved.
To convince them, we need to STOP building chatbots as anti-patterns.
So let’s promise each other two things:
- Building hybrid interfaces for transactional agents. We need to support our users to feel the power of AI without overwhelming them with prompting complexity.
- Reserve chatbots for applications whose aim is building an emotional connection with a social feel to the user. These systems won’t be perfect for a while. And that’s okay. But keeping the goal in mind is the first baby step. Execution will follow.