I'm building an AI agent for customer support, what should I use?

It depends on your constraints, but for a customer support AI agent, the best choice is usually:

Recommended default

Use a hosted LLM with tool/function calling, retrieval (RAG), and strong guardrails.

A solid stack is:

• Model: GPT-4.1 / GPT-4o, Claude, or Gemini
• Orchestration: your app server + simple agent logic
• Knowledge: RAG over help docs, policies, tickets, and product data
• Actions: tool/function calls for order lookup, refunds, account changes, ticket creation
• Safety: policy checks, escalation rules, human handoff

When to choose what

1) If you want fastest time-to-market

Use a managed API model:

• Minimal ops
• Good quality
• Easy tool calling
• Best for support automation

2) If you need high accuracy on company knowledge

Use RAG + citations:

• Index docs, FAQs, macros, SOPs, product manuals
• Retrieve relevant passages before answering
• Keep answers grounded in source material

3) If the agent must take real actions

Add function calling/tools:

• Check order status
• Reset password
• Cancel subscription
• Create/route tickets
• Update CRM

4) If you handle sensitive data / strict compliance

Consider:

• Data redaction before sending to model
• Restricted tool permissions
• Audit logs
• Role-based access control
• Private deployment or enterprise plan if needed

5) If you need lower cost at scale

• Use a smaller model for routing/classification
• Use a stronger model only for complex cases
• Cache frequent answers
• Deflect simple FAQs with search/article suggestions

Good architecture for support agents

1. Classify intent
2. Retrieve relevant knowledge
3. Answer or take action
4. Check confidence/policy
5. Escalate to human if needed

My practical recommendation

For most teams:

• Model: GPT-4.1 or GPT-4o
• Framework: lightweight custom logic, or LangGraph / Semantic Kernel if you want agent workflows
• Knowledge: vector database + RAG
• Integrations: tool calling to your support systems
• Fallback: human handoff via Zendesk/Intercom/Freshdesk

If you want a simpler answer

If you’re just starting:

• Use GPT-4.1/4o
• Add RAG
• Add tool calling
• Escalate uncertain cases to humans

If you tell me:

• your support volume,
• where your docs live,
• what systems you need to connect to,
• and your compliance requirements,

I can recommend a specific stack and architecture.