MenuStart a projectStart

Customer service

AI support inbox triage and reply drafts

AI inbox triage reads each incoming support email, labels it by topic and urgency and prepares a draft reply from your approved help content. The draft waits in your helpdesk for a team member, who checks it, edits it and decides whether to send it.

An example of how the studio would build it. No live services are connected.

From a support email to a checked reply.

  1. Step 01

    Email received

    A customer email arrives in Gmail or Outlook and becomes a ticket in Zendesk.

  2. Step 02

    Topic sorted

    Claude labels the ticket by topic and urgency, and flags complaints or anything sensitive for a person.

  3. Step 03

    Answer found

    Look up the relevant articles and policies in your help content, so the draft rests on what you have actually written.

  4. Step 04

    Reply drafted

    Claude writes a suggested reply in your tone, notes which help content it used and saves it as an internal draft.

  5. Step 05

    Person sends it

    A team member checks the draft, edits anything that needs it and sends it, or writes their own reply instead.

The result

Faster replies that still sound like you.

Drafts are never sent without a person’s approval. When your help content doesn’t cover a question, the draft says so and leaves the answer to the team.

Before we connect it.

Where this is useful

Useful when a small team answers similar questions every day and spends time searching for the right policy or article. A sensible first version only sorts and labels emails, then adds drafts for one or two well-documented topics.

What we need to connect

  1. 01Access to the support mailbox in Gmail or Outlook, or to the Zendesk account it feeds.
  2. 02Your help articles, policies and example replies, kept current and approved for use.
  3. 03The topics, urgency levels and routing rules your team already works with.

What needs careful handling

Support emails contain personal information, so data handling, retention and where the AI processing happens need agreeing up front. The model should answer only from your content and say when it can’t, rather than filling gaps. Complaints, refunds, legal matters and customers who may be vulnerable go straight to a person without a draft.

Questions people ask.

Will the AI reply to customers on its own?

No. Every draft waits for a person, who can edit, send or discard it. Automatic sending isn’t part of this design, because support replies often need judgement that only your team has.

What stops it making things up?

Nothing removes that risk entirely, which is why a person checks every draft. Drafts are based on your own help content and show which articles they used, and when nothing relevant is found the draft says so instead of guessing.

Question received → Conversation kept
Automation 66

WhatsApp AI answers with human handover

Review posted → Reply published
Automation 65

AI-drafted Google review replies

All automation examples