AI characters: from gaming NPCs to enterprise AI video agents

Video game characters once relied on scripts. The shopkeeper in your favorite role-playing game picked from a menu of pre-written lines, and the enemy soldier followed behavior someone coded years before you showed up.

Game players and enterprise users now expect live, context-aware responses. An AI character is a digital entity with a persistent persona, a consistent voice and behavioral identity, and the ability to generate responses from live context.

The most advanced systems see your face, register the hesitation in your voice, and adjust before you finish your sentence. In Tavus language, those production-grade systems are AI humans, part of the broader category of human computing.

From scripted NPCs to generative characters

Game studios have spent years working on hallucination prevention, character consistency, and real-time response timing. Enterprise teams face the same engineering demands across patient intake, candidate screening, and other high-volume conversations.

When a health system needs to conduct patient intake at 3 AM, or a recruiting platform needs to screen candidates at scale, it faces the same challenge Ubisoft faced when building its NEO NPC: making an AI character hold a real conversation without breaking down. Tavus, the human computing company, builds full-stack AI humans through its Conversational Video Interface (CVI), and they are already answering that question in production. NPCs went from rigid behavior to live generation over four decades, and the engineering lessons from that arc now shape how enterprises build conversational systems.

The path from finite state machines to live generation

Non-player character (NPC) AI evolved from the fixed pursuit algorithms of 1980s Pac-Man ghosts to finite-state machines and behavior trees. Halo 2 let combat outcomes alter squad tactics, and Shadow of Mordor's Nemesis System gave enemy officers persistent memory of past encounters.

Throughout that era, NPC dialogue came from authored text. In 2023, large language model (LLM)-driven NPCs began generating responses in real time based on player input and the live game state.

Game prototype maps for enterprise deployment 

Ubisoft's NEO NPC stack runs speech-to-text, routes it through an LLM, synthesizes speech and then synchronizes facial animation. The problems it documented- hallucination prevention, toxicity filtering, and character consistency map directly to an enterprise deployment checklist.

Four categories of AI characters

The term "AI character" covers a wide range of implementations, and drawing explicit lines between them matters for anyone evaluating where to invest.

• Game and entertainment NPCs: Built for narrative progression and gameplay. Response latency and facial animation are primary technical constraints, and memory is typically session-scoped.
• Companion and roleplay characters: The relationship itself is the product. Character.AI users averaged 93 minutes per day on the platform in 2024, according to research from the Brookings Institution.
• Branded and persona-driven characters: Brand identity is the primary deliverable, with strict content controls preventing off-brand responses that carry direct reputational risk.
• Enterprise and customer-facing characters: Measured by task completion and workflow execution. These require system integration, auditability, multilingual support, and compliance with the Health Insurance Portability and Accountability Act (HIPAA) and the General Data Protection Regulation (GDPR).

The categories share some capabilities, but the operating requirements differ sharply. Tavus builds AI humans for the enterprise end of that spectrum, where trust, memory, and real-time interaction matter in production. Persistent memory, emotional responsiveness, and persona consistency are prerequisites for any AI human that needs to build trust and presence during a patient intake call or a candidate screening conversation.

The technology stack behind a believable character

Real-time conversation puts every component on a shared clock, and the way those components are assembled decides whether the interaction feels present or stitched together.

The latency budget

Building a real-time AI human that can hold a face-to-face conversation requires six layers of technology operating within a sub-second pipeline. Automatic speech recognition alone can consume a large share of that budget.

Text-to-speech systems can use linguistic context such as punctuation and structure to improve prosody, but synthesis may begin before an LLM has generated a complete sentence. Speech recognition discards tone and speaking rate; the LLM receives only text, and the rendering layer has no direct perception of the user.

Why stitched-together pipelines show seams

Existing research on multimodal affective systems shows these capabilities have often been developed separately from language-model reasoning. Stitching together point systems can produce visible seams, awkward pauses, and mismatched facial expressions.

Those seams show up in the interaction. The conversation still works, but the sense of being attended to fades.

A face changes the conversation 

A visible face can deepen a conversation or quietly undermine it, and the deciding factor is rarely how lifelike the face looks.

Visual realism is not the same as trust

Visual presence in AI interactions affects trust, realism, and user preference differently across contexts. Research suggests that greater visual realism in AI-generated avatars can sometimes reduce trust relative to more authentic-seeming interactions.

Users in sensitive contexts like healthcare or HR may prefer text because it feels like a more natural modality for machines. Visual cues can still enrich human-AI interaction through facial expression, vocal tone, gestures, and pacing.

Behavior, not hyperrealism, creates presence

Hyperrealism can trigger uncanny valley effects that weaken the connection. A face matters most when its behavior fits the moment: nodding while listening, adjusting tone when it registers confusion, holding silence when someone needs a moment to think.

Phoenix-4, a real-time facial behavior engine, generates responsive facial behavior with appropriate emotional expression across 10+ controllable states, including active listening like nodding and responsive micro-expressions while the patient is still speaking.

Enterprise applications beyond entertainment

Adoption numbers and field results tell different stories, and reading them together reveals where conversational AI is already delivering measurable results.

What the adoption data shows

The Gartner 2026 Hype Cycle places agentic AI at the Peak of Inflated Expectations, with only 17% of organizations having deployed AI agents to date, while more than 60% expect to within two years.

Results in sales, healthcare, and recruiting

In sales and customer engagement, Club Med deployed conversational AI that dropped average first-response time on WhatsApp by 3.5 hours, according to a Forrester interview with its data leadership, freeing human agents to focus on premium interactions. In healthcare, a 2025 synthesis found that AI-based conversational agents demonstrated a moderate-to-large positive effect on health behavior changes, with patient trust as the strongest predictor of adoption.

In recruiting, organizations are using video-based candidate assessment to reduce manual phone screening and administrative load. They want conversational systems that complete work.

Building an AI character that holds a real conversation

Four jobs have to happen at once for a conversation to feel real: knowing when to speak, reading the person, deciding what to say, and showing it on a face.

The closed-loop system

A production-ready AI human depends on timing, perception, memory, and action. Tavus delivers those capabilities through a closed-loop system: Sparrow-1 governs conversational flow, Raven-1 perceives and fuses the other person's emotional and attentional signals, the LLM layer reasons about what to say and do next, and Phoenix-4 renders responsive facial behavior.

Many systems miss the timing first. Sparrow-1, a conversational flow model, predicts who owns the conversational floor at every moment, with 55ms median floor-prediction latency, 100% precision, and zero interruptions on the benchmark.

In a candidate screening call, Sparrow-1 holds the floor open while an applicant gathers their thoughts, then responds at the moment a human listener would, not as fast as possible. That timing helps the AI human respond to the person in front of it.

Perception that catches the mismatch

When a patient, during post-discharge follow-up, says "I'm fine" while avoiding eye contact and speaking in a flat tone, Raven-1 fuses those signals, detects the mismatch between what is said and how it is said, and outputs natural-language descriptions that the LLM layer reasons over directly.

That closed-loop system runs through the Conversational Video Interface as a single integrated cycle at sub-second latency. Its four parts are perception fusing signals, conversational flow governing timing, the LLM layer deciding what to say, and real-time facial behavior expressing the response.

Grounding, memory, and action

The Knowledge Base, a proprietary retrieval-augmented generation (RAG) model with around 30ms retrieval speed, grounds every response in the customer's data and procedures (English-only at launch). Persistent memory retains context across sessions: a returning coaching participant picks up where they left off because the AI human remembers they struggled with objection handling last Tuesday.

Objectives and Guardrails set measurable criteria for conversation completion and compliance natively. Function Calling lets AI humans take action mid-conversation: booking appointments, logging results, or triggering follow-up workflows.

Risks, safety, and trust in character design

Giving a system a face and a voice raises the stakes on every mistake it makes, and both the incident record and the regulatory picture are shifting fast.

Incidents are climbing

The risks compound when a character has a face and a voice. The AI Index 2025 reports that AI-related incidents rose to 233 in 2024, a 56.4% increase over the prior year.

Oxford Internet Institute research published in Nature found that training AI to sound warmer results in up to 30% more errors and makes the AI roughly 40% more likely to agree with users' false beliefs. In healthcare or compliance contexts where accuracy is mission-critical, product leaders need to define the primary metric before specifying character design.

Regulation is catching up

The EU AI Act's transparency provisions for generative AI become generally applicable in August 2026. Its Annex III covers several high-risk areas relevant to some enterprise AI deployments.

NIST's AI Risk Management Framework identifies "valid and reliable" as a trustworthy AI characteristic and emphasizes testing, evaluation, verification, and validation, along with measurement and monitoring practices, to assess AI system performance and risks.

Where AI characters are heading in production

40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. The forecast also warns that more than 40% of agentic AI projects will be canceled by the end of 2027.

Organizations that build governance and measurement frameworks before deploying are more likely to keep investment concentrated and projects alive. Gartner has highlighted the growing enterprise importance of AI agents, signaling that persona, memory, and adaptive behavior are becoming a recognized discipline.

The patient who finishes a post-discharge follow-up at midnight and feels heard. The candidate who gathers their thoughts during a screening call and finds the silence respected. AI characters create presence when they see you, remember you, and respond in real time.

See it for yourself. Book a demo.