Customer support
Tag every ticket against your taxonomy, draft replies grounded in your help center, score conversations against a rubric, escalate without leaking customer data, close knowledge gaps weekly, and design the handoff from a frontline AI agent.
For Support agents, team leads, QA specialists and support operations managers
Answer from your knowledge base, escalate cleanly, and close the gaps weekly
- Every new ticket tagged and prioritized against your taxonomy, with the deciding phrase quoted
- Reply drafts that cite a help-center article, or say plainly that no article covers the question
- QA scores on far more conversations than a manual sample, calibrated against your graders
- Escalations with repro steps and impact, and no customer PII in Slack, Jira or Linear
- A weekly list of questions your knowledge base cannot answer, with draft articles
On this page
- Where AI helps in this job
- Start in 15 minutes
- The tools people in this role use
- Members only: Workflow: triage and tagging you can trust
- Members only: Workflow: grounded reply drafts from the knowledge base
- Members only: Workflow: escalation summaries
- Members only: Workflow: CSAT analysis
- Members only: Frontline AI agents and the human handoff
- Members only: A skill for customer support
- Members only: Guardrails and hooks
- Members only: Routines that run without you
- Members only: Prompt ladder for customer support
- Members only: Measure it
- Members only: Risks and what not to automate
Support work is a stream: new tickets, replies, escalations, quality reviews, and the help center that should have answered half of them. AI reads fast and drafts well. It also states policy it made up, with total confidence. Every customer-facing answer has to trace to an article, a macro or a person.
Where AI helps in this job#
| Task | What AI does well | What you still own |
|---|---|---|
| Triage and tagging | Sets topic, intent, priority and language on every new ticket against your taxonomy | The taxonomy, routing rules and edge cases |
| Reply drafts | Drafts a reply from the matching help-center article and the customer's history | Accuracy, tone, and any promise about refunds or dates |
| Escalation summaries | Turns a long thread into repro steps, impact and what was already tried | Deciding to escalate, and to whom |
| QA scoring | Scores conversations against your rubric, quoting the evidence | Calibrating the rubric and coaching people |
| Knowledge gaps | Finds questions your articles do not answer and drafts new articles | Approving and publishing |
| CSAT analysis | Groups survey comments into drivers and ties them to tags, queues and handle times | What to change in product, policy or staffing |
| Frontline AI agent | Answers routine questions in chat, email or voice and hands off the rest | Which topics it may handle, and the handoff rules |
| Macros and saved replies | Rewrites old macros for accuracy, length and tone | Policy wording and legal review |
AI is weak where support is hardest: angry or vulnerable customers, policy exceptions, and anything the knowledge base does not cover. When an article is missing, a model fills the hole with a plausible policy, and a wrong refund promise costs more than a slow reply. Ground every customer-facing draft in a source. Keep a person on anything that commits money, changes an account or touches a complaint.
Start in 15 minutes#
These work in Claude, ChatGPT, Gemini or Microsoft Copilot (formerly Microsoft 365 Copilot). Mask names and emails first unless your workspace is approved for customer data.
Prompt
Tag a batch of tickets
Tag each ticket below. For each one give: ticket ID, topic, intent (question, problem, request, complaint), priority, language, and the exact phrase that decided the topic. Topics: [LIST YOUR TOPICS]. Priority: P1 = service down or money lost for many customers, P2 = one customer blocked, P3 = degraded or workaround exists, P4 = question or cosmetic. If a ticket fits no topic, write "unclear" and say what is missing. Tickets: [PASTE UP TO 30 TICKETS: ID, SUBJECT, FIRST MESSAGE]
Compare ten of the tags with what your team chose. Where they differ, the reason is usually a topic definition that needs one more sentence.
Prompt
Reply draft from a help-center article
Draft a reply to this customer using only the help-center article below. Customer message: [PASTE MESSAGE] Article: [PASTE ARTICLE TEXT AND LINK] Answer their question first, then give the steps, then link the article. Under 120 words, plain words, one apology at most. If the article does not answer the question, do not guess: write a short reply asking for the detail you need, and tell me what the article is missing.
Every factual sentence in the draft should appear in the article. If one does not, delete it or find the source.
Prompt
Escalation summary for engineering
Summarize this support thread for an engineer who has not seen it: [PASTE THREAD]. Give the issue in one line, steps to reproduce, expected and actual behavior, environment (browser, app version, plan), exact error messages, what support already tried, and how many customers are affected. Do not include customer names, emails or phone numbers. Refer to the customer by ticket number.
The test is whether an engineer could reproduce the problem without opening the thread.
Warning
Tickets contain names, emails, addresses and sometimes card numbers. Use the workspace your company approved for customer data, and never paste a full card number anywhere: PCI DSS requires stored card numbers to be protected even when they are masked on screen, and card security codes must not be stored at all after authorization. If you run an AI agent that talks to customers in the EU, Article 50 of the EU AI Act, which has applied since 2 August 2026, requires that people know they are dealing with an AI system, at the latest at their first interaction, unless it is obvious.
The tools people in this role use#
- Your general assistant. Claude, ChatGPT, Gemini or Microsoft Copilot for drafts, summaries, analysis and macro rewrites.
- Fin AI Agent. Made by Fin, the company formerly called Intercom, which Salesforce acquired in September 2026. It runs on Intercom's helpdesk or on Zendesk, Salesforce, HubSpot, Freshdesk and others. The US list price is $0.99 per outcome. Procedures handle multi-step workflows, Simulations test them before launch, and Guidance sets style and escalation rules in plain language. The Insights add-on adds a CX Score and content-gap recommendations.
- Zendesk AI agents. Since mid-2026 every Suite and Support customer gets the capabilities that used to sit in the Advanced tier, and billing counts automated resolutions that an LLM verifies, not assisted escalations. The Copilot add-on adds intelligent triage (topic, sentiment, language), auto assist that follows admin-written procedures, suggested replies and summaries. Zendesk QA, an add-on, autoscores every solved or closed ticket.
- Other help desks. Agentforce for Service (Salesforce), HubSpot's Customer Agent (credits used only on resolutions), Freshdesk's Freddy AI Agent, Gorgias AI Agent and Help Scout AI Answers all price AI handling per resolution, session or interaction. Read how each defines a resolution before you compare prices.
- Connectors for your own assistant. Intercom's MCP server reads conversations, contacts and articles and can create or update articles and add internal notes, but it cannot send customer-visible replies. Claude's directory lists it for US workspaces. HubSpot's connector reads, creates and updates tickets. Freshdesk has had an official MCP server since September 2026. Zendesk had announced an MCP server, but it was not generally available on 2026-10-13. Microsoft Copilot has connectors for Zendesk, Intercom and HubSpot.
- Claude Cowork with Anthropic's customer-support plugin. Open-source skills for ticket triage, response drafting, escalation, customer research and knowledge-base articles.
Workflow: triage and tagging you can trust
Members only below this line
Get the full Customer support playbook free
What you get on this page:
- 4 step-by-step workflows
- 13 prompts
- 1 skill
- 1 hook
- 2 routines
Also in the members part
- Workflow: grounded reply drafts from the knowledge base
- Workflow: escalation summaries
- Workflow: CSAT analysis
- Frontline AI agents and the human handoff
- A skill for customer support
- Guardrails and hooks
- Routines that run without you
- Prompt ladder for customer support
- Measure it
- Risks and what not to automate
Sources
Checked against these on 13 Oct 2026.
- Salesforce: Salesforce completes acquisition of Fin (10 September 2026) (opens in a new tab)salesforce.com
- Intercom: Pricing (Fin AI Agent and helpdesk) (opens in a new tab)intercom.com
- Intercom Help: Fin AI Agent outcomes (opens in a new tab)intercom.com
- Intercom Help: Fin Procedures explained (opens in a new tab)intercom.com
- Intercom Help: Run simulations for Fin Procedures (opens in a new tab)intercom.com
- Intercom Help: Manage Fin AI Agent's escalation guidance and rules (opens in a new tab)intercom.com
- Intercom Help: Trigger AI conversation summaries using Workflows (opens in a new tab)intercom.com
- Intercom Help: AI-powered content recommendations (opens in a new tab)intercom.com
- Intercom developer docs: MCP server (opens in a new tab)developers.intercom.com
- Zendesk: Expanded access to AI agent capabilities for all customers (opens in a new tab)support.zendesk.com
- Zendesk: About automated resolution tiers (opens in a new tab)support.zendesk.com
- Zendesk: About Zendesk Copilot (opens in a new tab)support.zendesk.com
- Zendesk: About intelligent triage (opens in a new tab)support.zendesk.com
- Zendesk: Understanding autoscoring categories (Zendesk QA) (opens in a new tab)support.zendesk.com
- Zendesk: Workflow recipe, AI agent collects info then escalates to a human (opens in a new tab)support.zendesk.com
- HubSpot: Set up the customer agent's handoff process (opens in a new tab)knowledge.hubspot.com
- Freshdesk: Model Context Protocol integration (opens in a new tab)support.freshdesk.com
- Microsoft Learn: Copilot connectors gallery (opens in a new tab)learn.microsoft.com
- EU AI Act Service Desk: Article 50 (opens in a new tab)ai-act-service-desk.ec.europa.eu
- European Commission: AI Omnibus enters into force (27 July 2026) (opens in a new tab)digital-strategy.ec.europa.eu
- PCI Security Standards Council FAQ 1146: masked PAN still needs protection when stored (opens in a new tab)pcisecuritystandards.org
- PCI Security Standards Council FAQ 1533: sensitive authentication data after authorization (opens in a new tab)pcisecuritystandards.org
- Presidio (data-privacy-stack) (opens in a new tab)github.com
- Intercom Help: Surveyed CSAT reporting (opens in a new tab)intercom.com
- Anthropic knowledge-work-plugins: customer-support plugin (opens in a new tab)github.com
- Claude Code docs: Hooks reference (opens in a new tab)code.claude.com
- Claude Code docs: Routines (opens in a new tab)code.claude.com
- Claude Help Center: Schedule recurring tasks in Claude Cowork (opens in a new tab)support.claude.com
- Claude Help Center: Use skills in Claude (opens in a new tab)support.claude.com
- Claude Help Center: Use connectors (tool permissions) (opens in a new tab)support.claude.com