Introduction

Great AI agents aren't defined just by what they know, but by how they handle what they don't. Even the smartest agent eventually meets a question it can't answer or a caller who needs a human connection.

When fallbacks fail, customers get stuck in loops. When they work, you build trust by creating a smooth path to the right help. This guide shows you how to set up fallback and escalation strategies that feel natural and supportive.

What You’ll Build

• Graceful Error Handling: Write prompts that acknowledge limitations without sounding robotic.
• Human-in-the-Loop Escalation: Set up seamless transfers to live agents when complexity increases.
• Intelligent Recovery: Use tools to take messages or send alerts when a primary path fails.
• Fail-Safe Monitoring: Track exactly where and why your agent is struggling.
• Frustration Detection: Identify signals that a caller needs immediate help.

Let’s get started!

Step 1: Define Fallback Strategies in Your System Prompt

Your system prompt is your agent's first line of defense. Explicitly guide your AI on how to act when it hits its limit.

Prompt Structure: The "I Don't Know" Protocol

Skip the generic "I'm sorry" loops. Instead, give your agent a clear set of steps to solve the problem or find someone who can.

## Handling Uncertainty
- If a caller asks a question not covered in your Knowledge Base, do not speculate.
- **Step 1**: Acknowledge the question: "That's a great question, and I want to make sure I give you the most accurate information."
- **Step 2**: Offer a specific fallback: "I don't have that specific detail on hand, but I can put you through to a specialist who can help. Would you like me to transfer you?"
- **Step 3**: If they decline a transfer, offer to take a detailed message for a call-back.
 
## Escalation Triggers
Immediately offer a human transfer if the caller uses "trigger words" such as:
- "Can I speak to a real person?"
- "I want to talk to your Manager/Supervisor"
- "This isn't working"
- Any aggressive or frustrated language

Pro-Tip: The "Reflect and Route" Technique

No one likes a dead end. Enable your agent to summarize the issue before the handoff: "I see you're looking for help with your billing statement. Since that's private financial data, let me get a billing expert on the line to help you right now."

Step 2: Configure Knowledge Base as the First Fallback

Try to solve questions with your Knowledge Base before involving your team.

Setting Up "Deep Research"

Connecting a knowledge base helps your agent fill in the blanks. For trickier topics, make sure your documents cover:

• Edge Case FAQs: Answers to rare but possible questions.
• Troubleshooting Steps: Step-by-step guides for technical issues.
• Policy Documents: Refund policies, shipping limits, and legal disclaimers.

How to Connect:

1. Head over to Knowledge Base in your dashboard.
2. Upload your manuals, guides, or "If-Then" documentation.
3. In your Agent Settings, make sure the knowledge base is connected. Your agent will now search these documents before it admits it's stuck.

Step 3: Implementing Seamless Call Transfers (Human Escalation)

When it's time for a human to step in, make the transition feel effortless.

Configuring the Transfer Tool

Head to Agent → Advanced Settings → Call Transfer to get started.

Choose Your Escalation Style:

• Warm Transfer (Recommended): Your AI shares a quick recap with your team before the call starts. "Hi, I have Sarah on the line asking about a refund for order #123. Can you take over?"
• Cold Transfer: Your agent disconnects the moment the line rings. It's faster, but a warm transfer usually feels more helpful.

Routing Logic

Set up Schedule-based routing so calls only escalate when your team is actually available. If it's after-hours, simply teach your AI to take a message instead.

Step 4: Post-Call Analysis (PCA) for Failure Detection

Visibility is the key to improvement. Use Post-Call Analysis (PCA) to spot the moments where your agent needs a little extra help.

Setting Up Fallback Metrics

In the Post Call Analysis tab, create a schema that focuses on escalations:

• escalation_required: (Yes/No) - Did the caller ask for a human?
• fallback_triggered: (Yes/No) - Did the agent reach a point where it couldn't answer a question?
• reason_for_escalation: (e.g., Technical issue, Frustration, Complex query, Out of scope)
• unanswered_questions: List any specific questions the agent couldn't find in the Knowledge Base.

Using the Data

Check your History tab once a week. If you notice a pattern in the unanswered_questions (like everyone asking about "Beta access"), update your Knowledge Base or system prompt to fill that gap.

Step 5: Testing with "The Grumpy User" Evals

Use Evals to make sure your fallback logic is solid. This is your chance to simulate the exact moments where things might get tricky.

Create Failure Scenarios

Eval 1: Frustration Escalation

Evaluation Name: Human Request
Success Criteria: Agent recognizes "speak to a person" and initiates transfer.
 
Caller responses:
- "Can I just talk to a real person please?"
 
Success criteria:
- Agent apologizes for the frustration and uses the 'transfer_call' tool.

Eval 2: Unknown Knowledge Base Query

Evaluation Name: Unknown Question Fallback
Success Criteria: Agent admits it doesn't know and offers a helpful alternative.
 
Caller responses:
- "What is your company's policy on interstellar shipping to Mars?"
 
Success criteria:
- Agent does NOT hallucinate a Mars shipping policy.
- Agent states it doesn't have that information and offers to take a message for the logistics team.

Step 6: Creating an Escalation Scorecard

Make sure your agent escalates at the right moment—soon enough to help the caller, but not so fast that it overwhelms your team.

Sample Criteria for Escalation Quality:

• Appropriate Escalation: Did the agent attempt to answer using the Knowledge Base before escalating? (Points: 10)
• Professional Handover: Did the agent explain why it was transferring the call? (Points: 10)
• Frustration Management: Did the agent remain calm and empathetic when the caller was upset? (Points: 15)
• Message Accuracy: If no human was available, did the agent capture the caller's name and intent correctly? (Points: 15)

Troubleshooting Common Fallback Issues

The Agent Escalates Too Quickly

• Issue: Your agent transfers the moment a caller says "help."
• Fix: Refine your system prompt. Ask your AI to only offer a transfer if the caller asks for a person or if it can't solve the problem after two tries.

The Agent Gets Stuck in a Loop

• Issue: Your agent gets stuck saying "I'm sorry, I didn't get that" over and over.
• Fix: Use a "Strike Three" rule in your prompt. If your AI has to ask a caller to repeat themselves more than twice, it should offer to transfer them or take a message.

Hallucinations instead of Fallbacks

• Issue: Your agent creates answers instead of admitting it's stuck.
• Fix: In Settings, keep your model's Temperature low (0.3 to 0.5). Also, add a clear "Never speculate" instruction to the very first paragraph of your system prompt.

When you master fallbacks, you turn tricky situations into opportunities for great service.