Real-time digital humans cannot eliminate every delay. Retrieval, tool calls, speech recognition and video synthesis can all add latency before the next response starts. A blank frozen face makes that delay feel longer, but an exaggerated “thinking” animation can look theatrical and repetitive. The best thinking state communicates presence without pretending the avatar is performing human cognition exactly.
Distinguish listening from thinking
The avatar should first show that it heard the user before entering a waiting state. Listening is about receiving input; thinking is the interval after the turn is understood but before output is ready. Separating these states makes turn-taking clearer and helps users know whether they need to repeat themselves.
Keep the default state subtle. Small gaze movement, a slight head shift and normal breathing are usually enough. Large gestures every time the model needs two seconds will become irritating within a long conversation. The visual goal is continuity, not entertainment during every pause.
Use latency thresholds
A 300-millisecond delay does not need a thinking animation. A two-second tool call may benefit from one, while a ten-second generation needs stronger progress feedback. The system can choose different behavior based on expected and actual wait time rather than playing the same loop for every response.
Avoid fake semantic cues. Looking upward, tapping a chin or pretending to search memory can imply a cognitive process the system is not actually performing. These cues may be acceptable as stylization but should not mislead users about what is happening. Neutral waiting motion is safer than behavior that suggests certainty, confusion or emotional judgment without evidence.
Make the state instantly cancellable
If the user starts speaking again, the avatar should leave the thinking state and return to listening. Long non-interruptible animations make the system feel slower even when the backend is responsive. Cancellable loops and transition-safe poses are more valuable than cinematic complexity.
Coordinate with audio indicators. A subtle visual state can pair with a small UI indicator such as “Checking…” or a progress pulse when a tool call genuinely takes time. The interface should not imply completion before the response is ready. Explicit text can be reserved for longer delays so normal conversation remains visually clean.
Use tool-aware waiting states
Looking up a calendar, generating an image and preparing a normal reply have different expected durations. The product can use one motion family but pair it with task-specific status text when appropriate. This creates transparency without needing a unique animation for every backend operation.
Vary timing rather than identity. Repetition can be reduced by changing blink timing, gaze direction and loop entry point while preserving the character’s normal motion style. A reserved avatar should not suddenly become highly animated only because the server is slow.
Measure perceived latency
Two systems with identical backend time can feel different depending on feedback. Human studies should rate whether waits feel understandable, awkward, deceptive or interruptible. Perceived latency is a legitimate product metric, but it should not be optimized by adding noise that distracts from conversation.
Do not hide excessive latency. A thinking state should not mask a twenty-second failure indefinitely. After a threshold, the system should explain that it is taking longer, offer cancellation or fall back to a simpler response. Good UX acknowledges backend limits rather than converting every outage into endless animation.
Test long conversations
A thinking loop that looks fine once may feel robotic after fifty turns. Long-session testing reveals repetition, timing mismatch and whether the state becomes associated with every minor pause. Frequency should be tuned using real session traces, not only polished demos.
Consider accessibility. Motion should not be the only signal. Users who cannot see the avatar clearly still need to know whether the system is listening, processing or ready to speak. Optional text or audio cues can communicate state without forcing everyone to interpret animation.
Thinking states belong inside the motion budget
Our article on real-time avatar motion budgets explains why a compact state library can outperform dozens of clips. A waiting state deserves inclusion only when it improves turn-taking clarity. If normal idle and listening states already communicate enough, teams should avoid creating a separate elaborate thinking animation simply because it looks impressive in a demo. Latency will remain part of AI interaction, especially when agents use tools or generate media. A well-designed digital human makes the wait legible and calm: subtle motion, clear state transitions, fast cancellation and honest feedback when the delay becomes unusually long.