AI Customer Support: Agents, Add-ons, Prompts

For customer support, use AI to sort tickets and draft replies from your own help articles, and keep a person sending every answer. A small script on a model API, built with a coding agent such as Claude Code, covers most teams. In our worked example the model cost is about one cent per ticket, and the real costs are reviewers and a current knowledge base. Checked 31 October 2026.

The job in one table

What the job is

Customer support is a stream of messages from people with a problem. Most are repeats: how do I reset my password, where is my invoice, why was I charged twice. A smaller share are hard: an angry customer, a bug nobody has seen, a legal threat, a refund that needs judgement. A good support setup answers the repeats quickly and gives people more time for the hard ones.

AI helps with the repeats. A model can read a ticket, find the matching help article and write a draft in your tone. It cannot, and should not, decide what your refund policy is, forgive a charge or promise a delivery date. The most useful framing is that the agent is a fast, tireless junior who drafts and a senior who reviews.

Which agent and tools fit

Support is not a coding job, so the "agent" is usually a small pipeline you build once. Pick by who will maintain it:

Our own build guide uses Claude Haiku for triage and Claude Sonnet for drafts. See Claude Haiku and Claude Sonnet, and the Anthropic provider page for access.

Add-ons that help

Add-ons connect the agent to the places your support work lives. Start read-only.

Each page shows the install line for several agents, the auth it needs and the risks. Browse the whole catalogue at MCP servers, and read what MCP servers are if the term is new.

Prompts and kit items

Our kit has two items that fit this task directly:

• Support reply drafting skill: a reusable procedure for reading a ticket and writing a grounded reply. See also what an agent skill is.
• Bug triage prompt: turns a messy report into a short, reproducible bug.

More prompts sit in the prompt library, filterable by task. For the standing rules your agent should follow, copy the CLAUDE.md starter template and add your support rules.

A safe workflow, step by step

The build guide gives the full version. In short:

1. Write the rules before the prompt. What the agent may and may not do, in plain sentences.
2. Set up a knowledge folder. Help articles as files, each with a title and date.
3. Triage with a forced, structured answer. Category, urgency, sentiment, escalate yes or no, in a fixed format.
4. Draft from the docs only. The reply must name the article it used. No article, no draft.
5. Hand off as a private note. The draft appears for an agent to review, not for the customer to see.
6. Test on past tickets before it sees a live one.
7. Go live in stages. A few queues first, then more.

The rules from the build guide are worth copying whole:

What it costs

Two bills: the model and the help desk.

Model. Prices from Anthropic's pricing page, checked 31 October 2026: Claude Haiku 5.5 is $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Claude Sonnet 5.5 is $2 input and $10 output. See Claude API pricing.

Worked example (our token counts are assumptions). A triage call of 500 input and 100 output tokens on Haiku costs about $0.0001. A draft call of 3,000 input and 400 output tokens on Sonnet costs about $0.01. A ticket with both costs about $0.0101.

Context size and retries move these numbers more than the price list does. A long thread can triple the input, and a rejected draft that is regenerated is paid for twice.

Building the agent. A terminal agent plan to write it: Claude Code Pro is $20 a month or $17 on annual billing, per our Claude Code pricing page. Many teams need one month of it.

Help desk and people. The help desk is a separate subscription, and the agent adds API calls, not seats. Check your vendor's pricing page for current plan prices. The larger cost is the reviewer's time and the work of keeping articles current.

Risks and how to review them

Prompt injection deserves care. The MCP spec says tools represent arbitrary code execution, and Claude Code's docs say to verify you trust each server, especially ones that fetch external content. A ticket is external content. Never give a drafting agent tools that move money or change accounts.

Measure it

Pick four numbers and track them for a month:

If edit time is higher than writing from scratch, the agent is not helping yet. Tighten the articles or narrow the queues it handles.

What a triage result should look like

Force the model to answer in a fixed shape so code can read it. A typical example, with made-up content:

Notice the last two rows. When no article matches, the right output is "I need a person," not a guess. A short escalation beats a confident invention every time, and your reviewers will learn to trust the agent because it admits gaps.

A staged rollout

Do not let the agent send anything to customers during this period. If after a month the drafts are good and the audit is clean, you can talk about letting it answer a narrow, low-risk category on its own, with a link to a human. Many teams never need to.

Common mistakes

• Starting with the customer-facing chatbot. Drafts for staff teach you where the gaps are at no risk to customers.
• Feeding the whole help centre into every call. It costs tokens and distracts the model. Search first, then send the two or three best articles.
• No owner for the knowledge base. The agent is only as current as your articles.
• Hiding the AI. Check your local rules on disclosure with your regulator.
• Measuring only speed. Fast and wrong costs more than slow and right.
• Skipping the privacy check. See below.

Privacy and data

Tickets contain names, emails, order details and sometimes health or payment information. Decide before you build what may leave your systems. Options are to redact personal data before the call, to use a provider under a contract that covers your data, or to run a local model so tickets never leave your network, which is the point of self-hosted AI. Check the provider's data retention and training terms on its data policy page. Keep logs of what was sent, and keep them as protected as the tickets themselves.

Who should do what

For the wider picture of business use, see AI tools for business, and for the sibling job on the revenue side, AI sales. A pattern for reading and sorting reports is also in build a support triage agent.

Our take

Support is one of the safest places to start with AI because the output is a draft a person reads. Keep the rules short and strict, ground every reply in a named article and give the agent no power over money. The cost of the model is small. The investment that decides success is a knowledge base that stays true. We built this from our own guide and vendor pages and have not run a support desk with it, and any help-desk product claims are marked for you to check at the vendor.