The future of human-computer interaction: from screens to conversations

High-value digital conversations often turn on whether people feel attended to. A familiar design pattern can show up in patient intake deployments: two teams deploy similar conversational AI and create very different experiences. One flow is designed to help patients complete the process and share relevant details; another may leave them abandoning the experience and asking for a human.

In human-computer interaction, the feeling of being attended to is often described as presence. Presence is now a practical design problem for teams building conversational interfaces. HCI is the field of shaping how people and interactive computing systems meet: how those systems are designed, evaluated, and put to work in human contexts.

Interfaces increasingly ask the machine to interpret more of the interaction. For product leaders, the shift from command-based interfaces to intent-based systems changes what good looks like. A strong product can respond to a command; a conversation-first product also has to account for human conversational qualities: good timing, attention, useful memory, and the sense of being understood.

What human-computer interaction means today

HCI emerged as a specialty within computer science, drawing on cognitive science and human factors engineering. Its original focus was personal productivity, including text editing, spreadsheets, and the practical work of getting a person and a machine to cooperate. The field has since expanded to cover almost all forms of information technology design.

The field's commitment to human experience has stayed steady. HCI is interdisciplinary, and its community is bound by usability and by the conviction that human activity and experience should be the primary driver in technology.

The MIT EECS research program frames its HCI work around creating systems that enhance human experience, distinguishing this from computer science's traditional focus on engineering goals such as performance and scalability.

AI changes the posture of interface design. HCI is moving from a command-based model, where humans type or click specific inputs, toward an intent-based model, where systems anticipate what a person actually wants. Early HCI researchers saw open-ended, human-to-human dialogue as an important direction for the field.

Open-ended dialogue is now buildable. A current expression of that work is the AI human: a digital entity that sees, hears, understands, and responds in real-time conversations across text, voice, and face-to-face video. AI humans accept natural language, timing, and face-to-face cues as the interface.

How human-computer interaction evolved

HCI history tracks design choices that make human intent easier to express, a progression toward intent-based interaction that academic work has traced from rigid command syntax to natural language. The difficulty of expressing intent in an interface is a long-standing concern in the field, framed as the gap between what a person wants and what the system makes possible. 

Then came the foundations of graphical interaction. Sutherland's Sketchpad appeared in 1963, Engelbart's mouse prototype followed in 1964, and the 1968 demonstration introduced video conferencing and hypertext to a live audience.

Xerox PARC turned graphical interaction, pointing, and windowed computing into the WIMP model: windows, icons, menus, and a pointing device. The Apple Macintosh carried that model into the mainstream and made computing approachable for people who had never written a line of code. The 2007 iPhone helped popularize touch as a mainstream way to interact with computing devices.

Richard Bolt's 1980 "Put That There" demonstration processed speech and pointing together, showing how multiple input modes could work in one system. Successive interfaces reduced the translation work between human intent and machine instructions.

The core components of human-computer interaction

Underneath every interface sits the same basic loop. HCI is best understood as a dialogue in which the output of one party becomes the input of the other. Cognition follows a similar cycle through action, perception, mental processing, and back to action.

Intent remains the design problem. Donald Norman's concept of the Gulf of Execution describes the distance between what a person intends to do and what the system makes possible. The governing principle is that systems should adapt to human intent, and conversation is appealing because it's the interaction model humans already know.

Modes of interaction shaping the experience

Today's products rarely rely on a single mode. The interesting design work happens at the intersections:

• Visual and graphical interfaces: The GUI is in a standardization phase, evolving through incremental change rather than reinvention. AI agents now navigate legacy graphical surfaces, scanning screens and clicking buttons the way a person would.
• Voice and conversational interfaces: Voice interaction differs sharply from graphical design, yet classic usability principles still govern its quality. Customer service, support, and internal knowledge workflows increasingly use conversational AI as the starting point for interaction.
• Gesture, haptic, and immersive interfaces: Spatial computing combines 3D interfaces with gaze, hand tracking, and haptic feedback. The hardware is maturing but still carries downsides: bulky equipment, short battery life, and calibration friction.

These modes are moving toward multimodal interaction. People already work across channels, taking in sight, sound, and touch at once. For conversation-first design, naturalness often comes from combining channels the way a conversation does.

What makes an interaction feel natural

Naturalness shows up in the rhythm of an exchange. A product can have the right answer and still feel wrong if it delivers that answer at the wrong time. Human conversation is unforgiving about delay. A response that arrives too early can feel interruptive; a response that arrives too late can make the other person feel stranded. Interface response carries the same pressure: fast responses help keep attention on the exchange, while lag makes people notice the machinery.

Natural interactions need both timing and memory. Context gives the system memory of what the person has already said, done, and preferred.

Without continuity, people are forced to repeat themselves, and the interaction starts to feel less like a conversation and more like a reset button. A conversation that feels natural remembers what came before, responds at the moment a person expects, and reflects an accurate read of what the person actually meant.

Many conversational AI investments stall at basic FAQs and routing. The conversations that require empathy and explanation are often the ones where breakdowns matter most. Tavus builds full-stack AI humans that see, hear, understand, and respond in real-time conversations.

Its conversational flow model, Sparrow-1, operates on raw audio at the frame level. Sparrow-1 predicts who owns the conversational floor without waiting for silence. On a benchmark of 28 challenging real-world conversational samples, it recorded 55ms median floor-prediction latency, 100% precision, 100% recall, and zero interruptions.

In a candidate screening conversation, Sparrow-1 holds the floor open while an applicant gathers a half-formed thought and waits for a human-like point to continue.

Where conversation becomes the interface

Conversation is becoming the control surface: it interprets intent, pulls operational context, and coordinates actions across applications. People are tiring of navigation, and conversation is becoming the front door to CRM, ERP, and internal knowledge systems.

Customer service adoption data supports the move toward conversational interfaces. 85% of customer service leaders planned to explore or pilot a customer-facing conversational generative AI capability in 2025. Face-to-face conversation carries cues that text does not. In health tech and recruiting, where disclosure and rapport directly shape outcomes, the interaction has to carry more than information.

The interface has to create enough presence for a person to keep going. Conversation as an interface only works when it's done well enough to create presence. When presence is missing, the experience feels like a phone tree with a new surface. Premature deployments can make experiences worse when they replace a frustrating workflow with a faster, more automated version of the same frustration.

Emerging technologies redefining the field

Several technologies are changing what teams treat as an interface. Spatial computing is moving from experimentation into broader enterprise interest, with AR, VR, and mixed reality pushing interfaces further into physical space. Brain-computer interfaces remain early and clinical, and they follow the same pattern: systems trying to get closer to raw human intent.

The largest near-term shift is agentic. 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025, per Gartner's prediction. When AI takes on tasks inside applications, the interface has to convey information alongside judgment, attention, and intent. Face-to-face conversation with an AI human gives that interaction the cues people expect from another person.

Designing for a conversation-first future

Designing conversation-first products requires planning for social expectations that forms don't carry. People expect a conversation to handle someone interrupting, hesitating, changing direction, or voicing a partial thought.

The structural principles are reasonably settled. Conversations must be non-linear, letting people use their own words rather than only menu choices. They also need to be cyclical, so users can pivot and circle back without starting over.

Trust is its own discipline. Teams earn trust when systems show what grounds an answer, stay within a clear scope, and handle uncertainty carefully. The NIST AI Risk Management Framework lists the characteristics of trustworthy systems, including being valid, transparent, explainable, and privacy-enhanced.

Accessibility belongs at the center of the design process. For users with motor impairments or situational limitations, typing may be inconvenient or impossible, which makes voice and multimodal input a matter of access first. Experiences in deployed settings need timing, perception, memory, and compliance to work in the same conversation.

Tavus's Conversational Video Interface (CVI) is designed for real-time AI human conversations in deployed applications. CVI runs as a closed loop across perception, conversation, intelligence, personality, and rendering.

Its behavioral stack keeps timing, perception, reasoning, and facial behavior working together in the same conversation. Sparrow-1 governs conversational flow, Raven-1 perceives and fuses the other person's emotional and attentional signals, the large language model (LLM) layer reasons about what to say and do next, and Phoenix-4 renders responsive facial behavior.

Raven-1, a multimodal perception system, fuses the other person's emotional and attentional signals into a unified understanding. In a post-discharge follow-up call, Raven-1 fuses a patient's hesitant tone with their downcast gaze, catching the mismatch between "I'm fine" and how they say it. The LLM layer can then soften its next question. Phoenix-4, a real-time facial behavior engine, can render the kind of attentive expression a nurse would offer.

CVI also includes the intelligence and personality layers used to shape a deployment. In a compliance training scenario for an insurance team, an AI human for sales practice coaching draws on the Knowledge Base, a proprietary retrieval system that grounds AI human responses in verified source material and returns answers in roughly 30ms.

The Knowledge Base uses the company's actual policy documents to ground every response. Objectives and Guardrails keep the session moving toward a measurable goal, such as confirming the rep can correctly explain a coverage exclusion, while enforcing compliance boundaries and escalating when a question falls outside scope.

Persistent Memory retains what each rep struggled with last time, so a returning learner picks up where they left off.

Where interaction goes from here

HCI has repeatedly changed how much machine syntax people need to learn. Command lines gave way to graphical interfaces, which gave way to touch, voice, and gesture, each one asking less of the person. Conversation is becoming a primary surface, and face-to-face conversation with an AI human gives that surface more of the cues people expect from another person.

For workflows currently handled by an interactive voice response (IVR) tree, a hold queue, or a chatbot that forgets context, the design question is whether an AI human can better support timing, continuity, and attention. Telehealth, hold queues, and text chat interfaces already exist; the practical question is whether a new interface can preserve context and respond with more of the cues people expect. The real measure of where interaction is heading is whether the person on the other side feels seen.

Build for the moment people feel seen

Return to the patient intake moment this article began with: a person trying to finish a conversation, decide what to share, and feel that the other side is paying attention. That is presence: the sense that something on the other end understood what they meant and responded the way a person would.

Presence is now a concrete design problem, but the human truth is simple. The future of interaction will be judged by whether technology helps people feel seen.

See it for yourself. Book a demo.