Troubleshooting & common pitfalls

Symptoms to fixes for AI Avatars. Wrong facts, mispronounced names, generic tone, robotic voice, cut-offs, slow replies, revoked share links.

Home › Help Center › AI Avatars › Troubleshooting & common pitfalls

If your avatar feels off — wrong tone, wrong facts, awkward pauses, mispronunciations — there's almost always a specific setting to fix. This page maps the symptom to the fix.

The avatar says wrong things

Symptom: the avatar gives facts that aren't true about your business, or quotes pricing that doesn't exist, or recommends features you don't have.

Likely cause: the facts aren't actually in the avatar's context. You may have written them in the System Prompt but it's drowning in too much other content, or you haven't attached the right Knowledge Base document.

Fix:

• Move all factual content (pricing, policies, product details) into the Knowledge Base, not the System Prompt.
• Check that the relevant document is attached to this specific persona (not just sitting in the library).
• Verify the document's status is "ready" — processing failures mean it's not actually being used.
• For mission-critical facts ("never quote a price not in the catalogue"), add a Guardrail on top of the Knowledge Base.

The voice mispronounces a brand name

Symptom: the avatar says your brand name, founder name, or a key product name wrong.

Fix: add a Pronunciation Dictionary entry. Test by triggering the word in a Share Link conversation.

The avatar feels generic / corporate

Symptom: the avatar replies sound polite but stiff, like a generic chatbot.

Likely cause: the System Prompt is too short, too vague, or written like a job description.

Fix: rewrite the System Prompt following the four-part structure (Identity → Expertise → Tone → Boundaries). Be specific about tone: "warm with dry humour" beats "friendly". Use a name. Aim for 150–500 words.

The avatar refuses things it should not / answers things it should not

Symptom: the avatar refuses to discuss a topic that's clearly in scope, or answers a question that's clearly out of scope.

Fix:

• Check your Guardrails — overly broad guardrails ("never discuss money") catch in-scope questions.
• Check your System Prompt boundaries — they may contradict your Guardrails. Pick one place for each rule.
• If too permissive: add a specific Guardrail with a redirect. ("If asked about competitors, redirect to: 'I can only speak to our own products.'")

The voice sounds robotic

Symptom: the avatar speaks in a flat, mechanical monotone with no warmth or natural pauses.

Likely cause: a mismatch between the Replica's voice and the language being spoken (an English-trained voice speaking another language), or a System Prompt written in a way that produces robotic replies (heavy on bullet points and corporate language, light on conversational flow).

Fix:

• If the language differs from the Replica's trained language, consider training a new Replica in the target language.
• Rewrite the System Prompt so it's spoken aloud naturally — short sentences, no lists, no headers, no markdown.

The avatar cuts off the visitor mid-sentence

Symptom: visitor is still talking, avatar starts replying anyway.

Likely cause: turn detection thinks the visitor is done. This usually self-corrects after a few sentences — visitors learn to pause cleanly.

Fix: add a note to your greeting or opening explanation: "When you're done speaking, just pause for a second and I'll reply." If this is happening constantly to multiple visitors, contact us — there may be an audio-quality issue or a turn-detection mismatch.

The avatar takes too long to reply

Symptom: 3+ second pause before the avatar starts speaking after the visitor finishes.

Likely cause: the intelligence model is reasoning over too much context — a very long System Prompt + a large Knowledge Base + a long conversation history.

Fix:

• Trim the System Prompt to 150–500 words.
• Split a giant Knowledge Base into one persona's most-relevant docs only.
• Try the faster intelligence model option.

A Share Link is "no longer available"

Symptom: visitor opens the link and sees a "no longer available" message.

Likely cause: one of three things — the link was revoked, the usage cap was hit, or the expiry date has passed. Check the Share Links tab — the link's status tag tells you which.

Fix: create a fresh link (with a higher cap or no expiry, if those were the issue). Old links don't un-revoke or refill — start a new one.

Still stuck?

If you've checked the above and your avatar is still misbehaving, the fastest path is usually:

1. Note the exact thing the avatar said wrong.
2. Open the persona in the builder.
3. Re-read your System Prompt and Knowledge Base out loud as if you were the avatar.
4. Usually, the wrong thing is right there in plain sight — a guardrail too broad, a doc too long, a prompt that contradicts itself.

Still stuck? Reach out via the support page with the specific persona name and what went wrong.

Related articles in AI Avatars

• Recording a training video that the engine will accept
• What is an AI Avatar?
• Writing a great System Prompt
• Anatomy of a great persona
• Choosing a Replica (the face + voice)
• Choosing the Intelligence Model
• Adding a Knowledge Base
• Building Objectives

Help Center · Learn · Privacy Policy · Terms of Service · Contact