If your AI Agent isn't behaving the way you expect, this guide covers the most common issues and how to fix them.
​

Check if AI responses are enabled. Go to Settings → Messaging and confirm AI Responses is set to Enabled. If it's disabled, no automated responses will be sent.
​

Check the flow-level override. If the AI isn't responding for a specific flow, open that flow in the Flow Builder and check the AI response mode under Configure Agent. It may be set to Disabled or inheriting from a global setting that's disabled.
​

Check if the conversation was overridden. If someone on your team previously overrode the conversation, the AI will no longer respond to it. Once overridden, AI responses are permanently off for that conversation.
​

Check if the subscriber is unsubscribed. If a customer has opted out of SMS, the AI cannot respond. The Inbox will show a message indicating the customer is unsubscribed.
​

Check for conflicting content. If the same topic is covered by a Knowledge Source, a Fact, and a Verified Answer, the AI has multiple sources to draw from and they may contradict each other. For example, if your Shopify policy says "14-day returns" but a Fact says "30-day returns," the AI may give either answer.
​

Fix: Pick one authoritative source per topic. Archive or update the others so they agree.
​

Check for stale content. A Knowledge Source or Fact may contain outdated information. Expand the Content Health section in the Knowledge Base tab to check for sources that may need attention.
​

Fix: Update or archive the outdated content. If the correct information has changed recently, add a Fact with the current details for an immediate fix while you update the full document.
​

Check the source badges. Open the conversation in your Inbox and click the source badge on the AI's response. It shows exactly which Knowledge Source, Fact, or Verified Answer the AI used. If it cited the wrong source, you know which content needs fixing.
​

The topic isn't in your knowledge base. If the AI escalates with a Knowledge Gap reason, it didn't have any relevant content to answer the question. The escalation banner includes a link to your Knowledge Base.
​

Fix: Add a Verified Answer, Fact, or Knowledge Source covering that topic. Then test it in the Playground.
​

The information exists but isn't being found. Your content may be buried in a long document, phrased differently than how customers ask, or categorized in a way that makes it hard to retrieve.
​

Fix: Add a Fact with the key information stated directly. Facts are short and specific, which makes them easier for the AI to match. If the question comes up frequently, add a Verified Answer for guaranteed matching.
​

A Knowledge Source failed to ingest. Check the Content Health section for sources with a Failed status. If a source failed during processing, the AI can't use it.
​

Fix: Check that the link is still accessible (for External Pages and Google Docs) or that the file is valid (for uploads). Re-upload or re-link the source.
​

The tone sliders need adjusting. If responses feel too formal, too casual, too enthusiastic, or too flat, adjust the tone sliders in AI Agent → Persona.
​

Fix: Make small adjustments (10-15 points at a time) and test in the Playground. Large swings can overcorrect.
​

Brand phrases feel forced. If the AI is using your brand phrases in awkward spots, you may have too many or they may be too long.
​

Fix: Remove phrases that don't work naturally in SMS. Keep them short and conversational. Fewer high-quality phrases work better than many mediocre ones.

Emoji usage is wrong. Too many, too few, or the wrong emoji.
​

Fix: Adjust the emoji frequency slider in AI Agent → Persona. Review your emoji collection and blocked emoji lists. Remove emoji that don't fit your brand.
​

The match confidence is too strict. If your Verified Answer uses Exact confidence, it only matches when the customer's question is very close to yours.
​

Fix: Try loosening to Confident (recommended for most cases). Add more variants that reflect how customers actually phrase the question.
​

The question variants don't cover how customers ask. Customers use shorthand, typos, and informal language. "Whats ur return policy" and "What is your return policy?" are the same question but phrased very differently.
​

Fix: Add variants based on real customer messages you've seen in your Inbox. Think about how customers text, not how they write emails.
​

Test it in the Playground. Send the exact question a customer asked and see if the Verified Answer triggers. If it doesn't, add that phrasing as a variant and test again.
​

The match confidence is too loose. If your Verified Answer uses Broad confidence, it may match questions you didn't intend.
​

Fix: Tighten to Confident or Exact. Review the question and variants to make sure they're specific enough.
​

The guideline may be too vague. "Be helpful with returns" doesn't give the AI a clear action. "If a customer asks about returning a final sale item, explain that final sale items cannot be returned and offer store credit as an alternative" does.
​

Fix: Rewrite the guideline to be specific and actionable. Include what the agent should do, not just what the topic is.
​

The guideline is still pending review. Auto-extracted guidelines start as Pending Review and aren't active until you approve them.
​

Fix: Check the Guidelines tab for pending items and approve the ones you want active.
​

Test in the Playground. Simulate the exact scenario your guideline should address and see if the agent follows it.
​

The skill may not be enabled. Go to AI Agent → Skills and confirm the skill is toggled on.
​

The skill needs context that isn't available.

The conversation didn't call for it. The AI uses skills when they're relevant to the conversation. If you ask a general question, the agent won't invoke a skill just because it's enabled.
​

Fix: Test in the Playground with a message that clearly calls for the skill: "Where's my order #1234?" for Track Order, "Do you have this in blue?" for Check Inventory.
​

If you tried to save a guideline and got an error message, the content didn't pass validation. Guidelines are checked for content that could compromise the agent's safety or behavior.
​

Common reasons for rejection:

Fix: Rephrase the guideline to focus on the customer interaction behavior you want. Instead of trying to change how the agent operates, describe what you want it to do in conversations.
​

If you've worked through the relevant sections above and the issue persists:

Contact TxtCart support for help.
​
​
​

Still have questions?
​
Please feel free to reach out to our wonderful Support team at support@txtcart.ai or via Live Chat.
​