The short answer
Using AI for customer service works best in stages: first let agents draft replies in a business AI plan such as ChatGPT Business or Claude Team, then turn repeat answers into help articles, automate triage with approval steps, and only then add a customer-facing agent such as Fin, which works across chat, email, voice and social channels and hands over to people. From August 2026, EU transparency rules mean customers should know they are talking to AI.
What you need before you start
- Your help content in one place: help-centre articles, policies (refunds, delivery, cancellations), macros and your 20 best past replies. AI answers are only as good as this material.
- Your ticket numbers: volume per week, the top ten reasons people contact you and how long replies take. You need a baseline to know whether AI helped.
- A business AI plan for anything that touches customer data, such as ChatGPT Business or Claude Team, both of which do not use your content for training by default. Never paste customer details into a personal account.
- A rule for handover: which questions must always reach a person, such as complaints, legal threats, vulnerable customers and anything about money you are unsure of.
AI in customer service does two jobs. Agent assist helps your team write replies, summarise threads and keep help articles current. Customer-facing agents answer customers directly and hand over to people. Start with the first: it is cheaper, lower risk and teaches you what the second needs.
Step by step: from agent assist to an AI agent
- Write down your tone and policies. One page: how you greet, how you apologise, what you can offer, what you never promise. This becomes the instruction you give every AI tool.
- Set up a shared workspace. In ChatGPT Business or Claude Team, create a support project and add the tone page, the policies and your best replies as files (use a shared project if your plan offers one, or give each agent the same files). Every reply drafted there starts from the same rules.
- Draft replies, never send them blind. Agents paste the customer's message with personal details removed, use the reply prompt below and edit before sending. Measure time per reply for two weeks.
- Turn repeat answers into help articles. Ask the AI to group a week of anonymised tickets into themes and draft an article for the top three. Customers who find answers do not need to write in.
- Automate triage inside your helpdesk. Zapier's AI by Zapier step can sort messages inside an automation and return structured fields such as a category or sentiment score, its ready-made Support Email Agent template drafts customer replies from your knowledge base, a Zendesk Ticket Agent template handles Zendesk tickets, and there is a per-tool Require approval before running toggle so nothing is sent without a person. AI Guardrails by Zapier can detect personal data in workflows. Every AI step costs tasks, so check your plan.
- Only then add a customer-facing agent. A dedicated product such as Fin, the AI customer agent that is natively integrated with Intercom, works across voice, chat, email, Slack and social channels, and Fin says it works with Salesforce, Zendesk, HubSpot, Freshdesk and other helpdesks. It hands conversations to your human team with the full context, and its testing suite lets you run simulations, regression tests and manual checks before a change goes live. Fin reports an average resolution rate of 76% across more than 12,000 customers; that is the maker's own figure, and your result depends on your help content.
- Tell customers it is AI. The European Commission says people using chatbots should be made aware they are interacting with a machine, and the EU AI Act's transparency rules take effect in August 2026. Say it in the first message and make "talk to a person" easy to find.
Worked example: a late-delivery complaint
A customer writes: their order is ten days late, they have emailed twice, and they want a refund. Your agent removes the name and order number and pastes the message into the support project in Claude Team, which already holds the delivery and refund policies.
The prompt asks for a reply that apologises once, explains the next step, offers what the policy allows and stays under 120 words, plus a note on anything the agent must check. The draft offers a refund or a replacement and flags: "Policy allows a refund after 14 days late; this order is 10 days late, so check whether you can make an exception." The agent checks the order in the helpdesk, decides to offer a replacement with express delivery, edits one sentence, adds the order details and sends it. Two minutes instead of eight, and the policy decision stayed with a person.
What to keep human
- Complaints that may become legal claims, and anything mentioning a lawyer or regulator.
- Vulnerable customers, bereavement and health situations.
- Exceptions to policy, goodwill gestures and refunds above a set amount.
- Final approval of new help articles and automated replies.
Common mistakes and how to fix them
- Launching a bot on thin help content. It will make answers up or fail. Fix: write and fix the top ten articles first.
- No clear way to reach a person. Customers get angry and your reviews suffer. Fix: offer handover in every conversation and route it with the full history.
- Pasting customer data into personal accounts. Fix: business plans only, and remove names and order numbers when you do not need them.
- Letting AI promise things. Fix: put "never promise refunds, dates or compensation the policy does not allow" in the instructions, and test it.
- Judging success by deflection alone. Fix: track resolution, satisfaction and repeat contacts, not just how many tickets the bot closed.
- Not testing changes. Fix: run a set of 20 real past questions every time you change the bot's instructions or content.
Which tool to use
For agent assist, use the business plan you already have: ChatGPT Business, Claude Team or Copilot with work accounts. Claude is our pick for drafting careful replies and help articles; see the Claude guide and the ChatGPT guide. To add AI to your current helpdesk without switching, look at Zapier. For a customer-facing agent across channels, Fin is our pick in the best AI for customer support. For more prompts, see prompts for customer support, and for a ready-made support kit, see Yuzoki kits.
Prompts to copy
Draft a reply from your policies
Using our tone guide and policies in this project, draft a reply to the customer message below. Apologise once, explain the next step, offer only what the policy allows, stay under 120 words. After the draft, list anything I must check before sending. [paste message with personal details removed]
You get: A ready-to-edit reply plus the checks that keep you within policy
Calm an angry customer
This customer is upset. Rewrite my draft reply so it acknowledges their frustration in one sentence, takes responsibility where we were at fault, and gives a clear next step with a time. Keep it human and under 100 words. Customer: [paste] My draft: [paste]
You get: A warmer reply that still keeps your promises realistic
Find the themes in a week of tickets
Here are [number] anonymised support tickets from last week. Group them into themes, count how many fall into each, show how many you could not place, and for the top three themes suggest the help article title that would prevent most of them. [paste tickets]
You get: A short list of articles that will reduce volume
Write a help article
Write a help-centre article answering "[question]" for [product]. Use our policy: [paste]. Start with the answer in two sentences, then numbered steps, then what to do if it does not work. Under 300 words, no marketing language.
You get: An article a customer can follow without writing in
Test the bot before customers do
Act as 15 different customers contacting [company] support. Mix easy questions, edge cases, angry customers and people trying to get a refund the policy does not allow. Give each as one realistic message I can paste into our bot to test it.
You get: A test set to run every time you change the bot