AI Agent Overview

AI Agent Overview #

AI Agents are intelligent assistants that can automatically handle customer conversations across all your communication channels. They use advanced natural language processing and retrieval-augmented generation (RAG) to provide accurate, context-aware responses based on your knowledge base.

What Are AI Agents? #

An AI Agent in Support Unicorn is a configurable virtual assistant that:

  • Understands customer questions in natural language
  • Searches your knowledge base for relevant information
  • Generates accurate responses using retrieved context
  • Handles conversations across live chat, SMS, and WhatsApp
  • Escalates to humans when needed
  • Learns from your content without requiring manual training

Think of AI Agents as your first line of support that works 24/7, never gets tired, and always responds consistently based on your approved knowledge base.

How AI Agents Work #

Support Unicorn uses a sophisticated RAG (Retrieval-Augmented Generation) architecture:

1. Knowledge Indexing #

When you add content to your knowledge base:

  • Documents are automatically chunked into semantic segments
  • Each chunk is converted to a vector embedding using OpenAI
  • Embeddings are stored in Turbopuffer (vector database)
  • Metadata is preserved for filtering and attribution

2. Query Processing #

When a customer asks a question:

  • The question is converted to a vector embedding
  • The system performs semantic search across your knowledge base
  • Most relevant content chunks are retrieved (not just keyword matching)
  • Context is ranked by relevance score

3. Response Generation #

With retrieved context:

  • OpenAI generates a response using the relevant chunks
  • The AI follows your custom system prompt (tone, style, guidelines)
  • Responses are constrained to your knowledge base (no hallucination)
  • Citations and sources can be included

4. Delivery & Learning #

After generating a response:

  • The message is delivered via the appropriate channel
  • Conversation history is tracked
  • Performance metrics are recorded
  • You can review and improve based on actual usage

AI Response Modes #

Support Unicorn offers three AI modes to fit different support strategies:

Auto Response Mode #

The AI automatically responds to all customer messages without human approval.

Best for:
- High-volume, repetitive questions (FAQs, hours, policies)
- Off-hours support (nights, weekends, holidays)
- Tier 1 support that can be fully automated
- Reducing response time to under 1 minute

Considerations:
- Requires a comprehensive, accurate knowledge base
- Should be monitored regularly for quality
- Consider starting with limited channels or topics

Hybrid Mode #

The AI generates suggested responses, but your team must approve before sending.

Best for:
- Teams new to AI automation
- Situations requiring human judgment
- Learning what questions the AI handles well
- Complex products or sensitive topics

Benefits:
- Human oversight ensures quality
- Faster response time (vs writing from scratch)
- Training opportunity for team members
- Gradual transition to full automation

Fallback Mode #

The AI only responds when no human agents are available or after a timeout period.

Best for:
- Teams who want to maintain a human-first approach
- Providing 24/7 coverage with a small team
- Handling overflow during peak times
- High-touch support with AI safety net

How it works:
- Human agents see the message first
- If no response within configured timeout (e.g., 5 minutes)
- AI automatically responds
- Agents can still take over anytime

When to Use AI Agents #

Ideal Use Cases #

AI Agents excel at:

  1. Frequently Asked Questions

    • Business hours and location
    • Shipping and return policies
    • Account management procedures
    • Product specifications

  2. Documentation Navigation

    • Helping customers find information
    • Explaining how features work
    • Troubleshooting common issues
    • Onboarding new users

  3. Information Lookup

    • Order status and tracking
    • Account balance and history
    • Appointment scheduling
    • Store locations and availability

  4. Routine Transactions

    • Password resets
    • Email updates
    • Subscription changes
    • Basic account modifications

When Human Support Is Better #

Consider keeping these conversations human-handled:

  • Complex technical troubleshooting
  • Sensitive account issues (fraud, disputes)
  • Angry or emotionally distressed customers
  • Edge cases not covered in knowledge base
  • Situations requiring judgment or empathy
  • Custom requests or exceptions to policy

Pro tip: You can disable AI for specific conversations using the toggle in Slack threads.

AI Agent Capabilities #

Multi-Channel Support #

AI Agents work seamlessly across all channels:

  • Live Chat Widget: Embedded on your website
  • SMS: Via Twilio integration
  • WhatsApp Business: Via Twilio integration
  • Future: Email, social media, voice

Conversations maintain context across channels if the customer is identified.

Contextual Understanding #

AI Agents remember conversation history:

  • Reference previous messages in the thread
  • Understand follow-up questions
  • Maintain context across multiple exchanges
  • Provide coherent, connected responses

Custom Actions (Advanced) #

Beyond just answering questions, AI Agents can:

  • Look up order information via API calls
  • Create support tickets in Zendesk
  • Process refunds through Stripe
  • Update customer records
  • Schedule callbacks
  • Trigger workflows

See Agent Actions for more details.

Getting Started #

Ready to create your first AI Agent?

  1. Train Your Agent - Add knowledge sources
  2. Test Your Agent - Verify responses
  3. Integrate Channels - Connect communication channels
  4. Analyze Performance - Monitor and improve

Best Practices #

Start Small #

  • Begin with a focused knowledge base (10-20 documents)
  • Cover your top 20 most common questions
  • Test thoroughly before enabling auto-response
  • Expand based on actual customer questions

Iterate Continuously #

  • Review AI conversations weekly
  • Add content for questions it couldn't answer
  • Update outdated information promptly
  • Refine system prompts based on performance

Set Clear Boundaries #

  • Configure when the AI should escalate
  • Define topics that require human handling
  • Set expectations with customers (AI vs human)
  • Provide easy escalation path

Monitor Quality #

  • Track customer satisfaction scores
  • Review escalation rates and reasons
  • Check response accuracy regularly
  • Solicit feedback from your team

Technical Details #

  • AI Model: OpenAI GPT-4o-mini (default, configurable)
  • Embeddings: OpenAI text-embedding-3-small
  • Vector Database: Turbopuffer
  • Context Window: Up to 128K tokens
  • Response Time: Typically 2-5 seconds
  • Supported Languages: 50+ languages via OpenAI

Pricing #

AI responses consume credits from your organization's balance:

  • 1 credit per AI message (configurable)
  • 50 free credits per month included
  • Purchase additional credits as needed
  • Subscriptions include monthly credit allocations

See Billing for more details.

Need Help? #

FAQ #

Can AI Agents make mistakes? #

Yes, like any AI system, agents can occasionally provide incorrect or incomplete answers. That's why we recommend:
- Starting in Hybrid mode for quality control
- Maintaining an accurate, up-to-date knowledge base
- Monitoring conversations regularly
- Using Fallback mode for sensitive topics

How accurate are AI responses? #

Accuracy depends on your knowledge base quality. When properly trained with comprehensive, accurate content, AI Agents typically achieve 90%+ accuracy on questions within their knowledge domain. They're designed to say "I don't know" rather than guess.

Can customers tell they're talking to AI? #

By default, yes. We recommend transparency - customers appreciate knowing they're getting instant AI assistance with the option to speak with a human. You can customize the greeting message in your agent's configuration.

What happens if the AI doesn't know the answer? #

The AI will:
1. Acknowledge it doesn't have enough information
2. Offer to connect the customer with a human agent
3. Provide related information if available
4. Log the question for you to add to the knowledge base

Can I have multiple AI Agents? #

Yes! Create different agents for different purposes:
- Product support agent
- Sales assistant
- Billing and accounts agent
- Technical support specialist

Each agent can have its own knowledge base, system prompt, and configuration.

How do I improve my AI Agent over time? #

  1. Review conversations in the Analyze section
  2. Identify common questions the AI struggled with
  3. Add that content to your knowledge base
  4. Update outdated information
  5. Refine your system prompt for better tone/style
  6. Adjust escalation rules based on performance

Next: Train Your AI Agent to build a powerful knowledge base.