Prompted by the redesign of the ONTEC AI interface, we spent months researching, pondering, and debating: What does an ideal AI user interface look like? What functions and capabilities does it offer? Where should users be deliberately guided, and where should they be intentionally given maximum freedom? Which information do users need, and which would overwhelm them? In this article, we share our findings, underpinned with hands‑on insights from the new ONTEC AI user interface in practice.
Curious about the new ONTEC AI interface? More details coming soon!
Introduction
Designing AI user interfaces isn’t easy: people have only been using AI for a few years, and usage patterns are still forming. While AI evolves quickly, human UI habits outside AI are now quite stable.
By contrast, people have used traditional software for years, we’ve been chatting for a long time, and we have clear expectations for how a digital tool should interact with us.
On this basis, people are trying out new AI solutions, and AI providers are experimenting with features to offer interfaces that are increasingly intuitive to use.
Chat is the starting point, not the end goal.
Jakob’s Law says that people spend most of their time with other products and therefore expect new products to work in similar ways. Familiar patterns lower learning costs and improve acceptance (Source: Jakob’s Law)
Today, ChatGPT is arguably the epitome of AI use: many people interact with it regularly, may even use it daily, and know how it works.
A chat is now the prevailing mental model for using AI: users enter instructions as a prompt, and the system returns a result as an answer, all in natural language.
The chat paradigm lowers barriers to entry because it builds on learned patterns—and because conversations in natural language are deeply human. At the same time, chat hits limits when tasks are highly formalized or structured, require many parameters, or when results need to be grasped visually (maps, tables, documents).
A chat interface therefore feels intuitive for many use cases—but it must not be an excuse to avoid better, specialized interactions.
How much interface does a good interface need?
Just as in the real world it is sometimes more effective to state instructions explicitly rather than phrasing them in full prose, the same applies to AI systems.
Example of a search query: “shoe size 36” vs. “shoes that are mostly worn by people around 1.55 cm in height, who naturally have relatively short and narrow feet.” Or, if a “wall of text” is to be avoided, a strict specification of text length.
An intuitive chat interface is therefore helpful in many cases; however, when use cases become more complex, the interface often needs to provide more guidance.
Between purely speech-based operation without visible UI (speech-to-text) and the depiction of complex input forms as in SAP applications or API controls, there is an interesting design space in which the appropriate interaction is chosen situationally.
Between purely speech-based operation without visible UI (speech-to-text) and the depiction of complex input forms as in SAP applications or API controls, there is an interesting design space in which the appropriate interaction is chosen situationally.
In the short term, it remains a balancing act: Some workflows work very well as natural‑language dialogue, while others require structured inputs, precise validations, and approvals.
The right level of autonomy: assistance, agents, automation
Good AI systems find the right degree of autonomy, aligned with the task, risk, and responsibilities.
Assistance means the AI actively supports without taking decisions away. It suggests options, explains alternatives, flags uncertainties, and requests approvals. Typical tasks include drafting, summarizing, researching, or preparing structured steps that a human can confirm with a click. The decision stays with the user, while the machine provides speed and breadth.
Agents go a step further: they operate tools and services, execute multi-step workflows, or react to events (e.g., incoming emails, schedules, status changes).
A robust agent isn’t “magical” but transparent: it shows what it intends to do, what it has done, and where it is uncertain. This builds trust—and keeps it clear where humans should step in.
Automation makes sense where risk is low, the effect is clear, and steps are reversible. Examples include notifications, routine exports, or low-stakes automatic classification. Here, reliability matters more than creativity: with audit logs, rollback, clear service boundaries, and monitoring.
In professional environments, the principle is: as much automation as possible, as much human control as necessary. That reduces cognitive load without diluting responsibility.
In ONTEC AI, assistants are preconfigured for specific roles and operate within clearly defined data and tool scopes. Clear guardrails apply to automation zones. All paths—assistance, agent, automation—log what happened so teams can review, audit, and learn.
Augmented Intelligence instead of loss of control: quality, verification, responsibility
In many cases, it is particularly important not to release humans from responsibility—for example, when handling highly sensitive data and in complex decisions.
High‑quality outputs arise when systems show evidence and intermediate steps, flag uncertainty, and enable deliberate approvals.
The human remains the decision‑maker—the system supports, explains, and logs.
For UX, this means users are more involved and informed: How is a result produced? Where does the AI get its information? How did it process the information?
At ONTEC AI, we combine a familiar conversational layer with an action layer (agents). The AI may initiate processes—but with clear checkpoints. If the agent needs structured inputs, it informs the user.
Targeted non‑use of AI
Paradoxical but effective: the simplest form of error prevention often involves deliberately not using parts of the AI—and taking the lead yourself where precision matters. Why?
In day-to-day work, users often know exactly which tool they need: “Search in knowledge base X” or “Use the specialist tool Y.”
For example, if a user types “Find me documents on ISO,” the AI initially doesn’t know whether this refers to general documents from the internet or company‑specific documents from the intranet. If this choice is made obvious in the interface, the user is much more likely to get exactly what they’re looking for.
This deliberate choice is, in a sense, “manual tool calling”: guiding the AI explicitly instead of letting it interpret freely. The human can make this decision upfront and thereby improve the result.
No‑agent paths as simple error prevention: in ONTEC AI, anyone who already knows which tool is right can start it directly. Think of it like placing a specific file folder in front of your AI colleague to search in, rather than having them search the entire office—or the whole city.
Meaningful multimodality
In practice, you encounter structured data, files, and images as well as multi‑step workflows that require a clarifying question.
Good UIs recognize this diversity and deliberately choose the appropriate medium: for structured inputs, dynamic forms and validations are better than free text; for complex processes, targeted follow‑up questions help capture parameters unambiguously.
The same applies to output: many results cannot be sensibly presented as running text. Maps, charts, interactive tables, or directly editable documents (e.g., Excel, Word, PDF) make information quicker to verify and decisions more robust.
ONTEC AI’s growing component library includes tables with inline validation, preview and editing of Office documents, as well as image tools and map widgets. The chat remains the hub—the components provide precision and overview.
Transparency without overload: information architecture for the LLM era
Transparency is not an end in itself—it should improve decisions, not overwhelm people with information. The appropriate depth depends on role, context, and risk.
Modern AI assistants produce large volumes of intermediate results. The challenge is to shape transparency so people can make informed decisions without drowning in detail.
In practice, a progressive transparency model works well. It separates three layers:
Decisive for approval: The default view shows the key point, a brief rationale, and the most important source or metric. That’s sufficient for many approvals and keeps the interface uncluttered.
Useful for traceability: Those who need more can expand details—data used, intermediate steps, parameters.
At ONTEC AI, we prioritize relevance: default views provide the key point, a one‑sentence justification, and the most important source. For higher‑risk actions, human approval is mandatory (guardrails); non‑critical, reversible steps run automatically. We flag uncertainties. Critical workflows are segmented by checkpoints. The result is a system that can grow without relieving users of responsibility.
How much transparency for whom?
Different user groups have different needs, for example:
Decision makers want concise summaries with clearly identified risks and a recommendation.
Subject‑matter experts require evidence, assumptions, thresholds, and the ability to adjust parameters.
Admins and compliance teams need complete traceability and export functions.
Good systems respect these differences without fragmenting the experience. The mode of presentation also matters. Instead of flooding users with raw logs, structured explanations help: Why was this recommendation made? Which alternatives were rejected—and why? Where is the model uncertain?
When this information is presented clearly, concisely, and in the right place, people detect hallucinations faster, verify more efficiently, and stay able to act.
Transparency doesn’t mean showing everything, but showing what is right and important—and keeping the path to deeper detail open.
By default, ONTEC AI presents the key point along with the central rationale and the most relevant source. One click opens the next level down: the tools used, parameters, and intermediate results. When model uncertainty is high or the risk is elevated, we automatically increase the level of detail and require approval. Complete logs are available for audits, including cost and runtime data. This keeps the interface lean—yet transparent when it matters.
Who can do what—and when? Control, roles, governance
Good AI systems combine usability with governance. Role and permission models define which data and tools assistants are allowed to use. Critical actions should be gated by clear approvals; cost and runtime limits prevent surprises. Escalation paths and undo options build trust.
In ONTEC AI, assistants are preconfigured by role (tools, data scopes, approvals). Admins get full auditability; subject‑matter users see evidence; decision makers get compact summaries with clearly stated risks—the depth varies by role.
Errors and feedback
AI makes mistakes too; that’s practically unavoidable: where human input isn’t correct or ideal, the AI will only produce inadequate results. Retrieval‑augmented generation (RAG) can minimize errors but never fully avoid them.
In development, we treat errors as learning signals—the key is to handle them briefly, systematically, and in a traceable way. In practice, it pays to prioritize by severity and frequency, validate early in the system (required fields, plausibility checks), and provide clear, action‑oriented guidance.
The goal is a closed feedback loop: prevent, detect, explain, correct, document—without slowing productivity.
In ONTEC AI, there is a feedback module that lets users record their satisfaction in detail. These data can then be used for further optimization—by the customer’s IT department or by the ONTEC AI team. In addition, our team has implemented comprehensive evaluation measures, for example to capture the performance of different AI models in different application areas.
Trends in AI interfaces
More capable models, more information—apply transparency deliberately Larger contexts and agentic workflows create more intermediate steps and outputs than people can reasonably oversee. Instead of “show everything,” clear decisions are needed: what is relevant for approval, what belongs in the detail view, and what only in the log? Progressive transparency keeps users able to act.
Strengthened usage patterns UX starts before the UI: the user journey now begins well before the visible surface—with integrations, permissions, data access, monitoring, and cost control. People come to the UI with growing prior knowledge (and expectations) from other tools; the UI must pick up familiar patterns without shying away from necessary specialization. UI/UX‑first and governance‑by‑design are therefore UX topics, not just technology.
The growing power of prompts The LLM often sees little of the user’s input: in large context windows, the user’s contribution competes with system and tool prompts; the latter often shape behavior more than the chat message itself. That’s why clear, versioned system prompts, guardrails, role scopes, and above all a context management strategy are critical—including visibility into what is working “under the hood.” Here too, it requires careful judgment about how much users should be able to intervene in context management—and how they are informed about the context state..
Context management strategy
LLM requests are inherently stateless: the model knows neither the conversation nor the context; every request must resend all relevant content. For example, in a conversation with the AI about a book: first you ask what’s on page 27, then on page 38, and finally on page 300. For each of these requests, the entire book is transmitted again; by the third request, the answers to the first two are included as well—creating multiple redundancies. A context management strategy asks: which parts of the conversation history should be removed, which are essential? For instance, you can insert a summary to reduce character count while preserving context.
Return to taking control Many customers want a user interface again—one they can co‑design and control. Buttons, drop‑downs, input fields.
Summary and key takeaways
In fact, many of the questions discussed are not straightforward to answer. Within the ONTEC AI team, this shows up in lively exchanges between frontend, backend, and product teams—often with a philosophical slant.
What is clear:
LLMs/AI are a user interface in their own right—one in which a button can be superfluous because you can issue commands in natural language. You no longer need to learn how to operate software; you can communicate your intent in natural language—that’s what agents enable.
Chat is the starting point, not the goal: familiar conversation lowers barriers, but precision comes from specialized UI components and multimodality (forms, tables, maps). Where it makes sense, users intentionally operate tools manually—“no‑AI” paths are often the best error prevention.
Verification before autopilot: transparency is applied progressively (“what’s relevant first, drill down as needed”), based on roles and risk. Humans remain accountable; agents work with guardrails and approvals so results are reviewable and safe.
UX starts before the UI: API‑first and governance‑by‑design (permissions, cost, monitoring) shape usage long before anyone sees the interface. Assistants/agents are modular and framed by system prompts—the LLM often sees very little from the user, so clear guardrails and integrations are critical.
The biggest challenge is not flashy screens but reliably correct results.
FAQs
What makes a good AI interface?
A good AI interface uses chat as a familiar entry point but relies on specialized components (forms, tables, maps) and multimodality for precise tasks. Transparency is applied progressively: the key point and the most relevant source first, with details available as needed; agents operate with guardrails, and the human stays in the loop. API‑first and governance‑by‑design (permissions, cost, monitoring) ensure the UX is right even before the UI.
How do I increase employees’ use of AI with an AI interface?
Familiar patterns lower barriers; quick wins and clear language build trust; training is targeted where processes are complex or regulated. Visible control (evidence, uncertainties, approvals), role‑based views, and the option to deliberately choose tools manually increase adoption—supported by an integrated feedback loop.
