How to Use Claude for Customer Support

Customer support is one of the cleanest fits for Claude. It is text in, text out. It rewards long-context analysis over big document sets, accurate retrieval from a knowledge base, and tool-using agents that can look up an order or check a subscription. Claude does all three well. This walkthrough builds a working assistant step by step, with concrete settings you can copy.

Step 1: Pick the right model for your support tier

Support workloads are price-sensitive because volume is high and most tickets are routine. Match the model to the job rather than defaulting to the flagship. All three models read images, so they can handle screenshots customers attach.

• Haiku 4.5: fastest and cheapest, $1 in / $5 out per million tokens. Use it for the front line: FAQ deflection, classification, routing, short answers. 200K context.
• Sonnet 5: the current default on Claude Free and Pro and available in the API. Introductory pricing is $2 in / $10 out through August 31, 2026, then $3/$15. Use it as the starting point for real conversations and reasoning over your help center.
• Opus 4.8: flagship, $5 in / $25 out, 1M-token context. Reserve it for the hardest escalations, complex multi-step agent runs, and long autonomous investigations.
• Fable 5 was restored globally on July 1, 2026. Mythos 5 is limited to approved Project Glasswing organizations, so do not build support operations around it without confirmed access.

Step 2: Write a system prompt that sets the rules

The system prompt is where you encode tone, scope, and guardrails. Be explicit about what the assistant must not do, when to escalate, and that it should never invent policy. Keep it tight and concrete.

Example system prompt: "You are a support assistant for Acme. Answer only from the provided knowledge base context. If the context does not contain the answer, say so and offer to escalate to a human. Never invent prices, policies, or refund terms. Keep replies under 120 words unless the customer asks for detail. Confirm the customer's intent before taking any account-changing action."

Step 3: Ground answers in your docs with RAG

The single biggest driver of support quality is grounding. Retrieval-augmented generation (RAG) pulls the relevant help-center articles, policy pages, and past tickets, then passes them to Claude as context so it answers from your truth instead of guessing. Claude is strong at long-context analysis, so you can pass generous context without losing the thread.

• Index your help center, policy docs, and resolved tickets into a vector store.
• On each query, retrieve the top relevant chunks and inject them above the customer message.
• Tell the model in the system prompt to answer only from retrieved context and to cite which article it used.
• On Sonnet 4.6 or Opus 4.8 you have 1M tokens at a flat rate, so you can pass long policy documents in full when retrieval is ambiguous.

Step 4: Connect your tools with MCP

Real support needs actions, not just answers: look up an order, check subscription status, create a ticket. Anthropic created the Model Context Protocol (MCP) and Claude has first-class support for it, so you connect your systems as MCP servers and the model calls them as tools. Keep write actions gated behind a confirmation step so the assistant proposes the change and a human or the customer approves it.

Common support tools to expose

• Order lookup (read-only): pull status, tracking, and line items.
• Account/subscription lookup (read-only): plan, billing date, entitlements.
• Knowledge search: query your indexed docs on demand.
• Ticket create/update: log the conversation and route escalations.
• Web search: Claude can search the web for public, non-account information when your docs fall short.

Step 5: Cut cost with caching and batching

Support traffic repeats. The same system prompt, the same policy docs, and the same tool definitions go out on every request. Prompt caching reuses that stable prefix for up to 90% off cached input, which is a large saving when your context is mostly fixed. For work that does not need a live reply, like overnight ticket triage or backfilling tags, batch processing runs 50% cheaper.

• Prompt caching: cache your system prompt, policy context, and tool schemas. Up to 90% off cached input tokens.
• Batch processing: 50% cheaper for non-real-time jobs like sentiment tagging, summaries, and bulk classification.
• Model tiering: start on Sonnet 5, then route volume and hard cases separately only after support-quality evaluations pass.
• Watch output tokens: they cost 5x input on every model, so cap reply length in the system prompt.

Step 6: Test, set the human handoff, and ship

Before going live, run a set of real past tickets through the assistant and check that it answers from context, escalates when it should, and never fabricates policy. Define a clear handoff: any low-confidence answer, any account change, and any frustrated customer goes to a human. Start the assistant in a draft or suggested-reply mode for agents before letting it reply to customers directly.

• Replay 50+ historical tickets and score accuracy and escalation behavior.
• Confirm the assistant says "I don't know" instead of guessing when context is missing.
• Verify all write tools are gated behind confirmation.
• Set thresholds for automatic human handoff.
• Start in agent-assist (suggested replies) before full automation.

Note on scope: Claude is text and code focused. It does not generate images, so it will not produce diagrams or marketing visuals for your support content. Voice exists in the mobile apps but is not a core support differentiator. Lean on what Claude does best: grounded answers, long-context reasoning, and reliable tool use.

THE CLAUDE INSIDER

Get the Claude playbook in your inbox.

One weekly email for Claude and Claude Code users. Real workflows, no hype. Subscribe and we send you The Claude Power-User Cheatsheet.