AI customer support in Help Scout works best as a draft-and-review loop. An agent reads the inbound conversation through the Mailbox API, pulls customer history, and posts a draft reply as an internal note. A human approves, edits, or rejects. Saved replies get suggested, not auto-sent. Help Scout Beacon traffic flows the same way. The setup suits teams of two to fifteen.
The workflow
1. Capture inbound. The Mailbox API exposes conversations, customers, and threads as JSON. Register a webhook for convo.created and convo.customer.reply.created. Every new conversation POSTs to your agent endpoint. Zapier handles this for teams that don't want to host a listener. For real volume, run a worker. The response-drafting workflow applies here.
2. Summarize the customer. The agent calls GET /v2/customers/{id} and pulls the last ten conversations. ChatGPT or Claude condenses the history into a five-line brief: plan, last ticket, sentiment trend, open issues. That brief posts as an internal note tagged ai-context, agent-only.
3. Draft the reply. The agent matches the inbound intent against your saved-reply library, then suggests an existing macro or drafts fresh. Drafts post as internal notes. Tag them proposed-by-mei so the reviewer knows the source, matching the agent identity practice of attributing AI work to a named agent.
4. Human review and send. The reviewer copies the draft, edits, sends through Help Scout's UI. Send actions stay human. The agent audit and compliance log captures the proposal, the edit diff, and the final body.
5. Learn from the edit. After send, the agent compares draft to sent. Large deltas flag the saved reply for revision. Small deltas mark it well-tuned. Suggestion quality climbs without a training job.
Worked example: a refund request
Customer emails: "I was charged twice for my March renewal." The webhook fires. The agent pulls the customer record, finds the plan, the last invoice, three prior billing tickets. It posts an internal note: "Likely duplicate from the May 12 retry. Refund eligible under standard policy. Suggested saved reply: billing/duplicate-charge-refund." Reviewer reads, clicks the saved reply, edits the dollar amount, sends. Ticket closed in ninety seconds.
The persistent-state pain
The interpretation work vanishes when the ticket closes. The customer brief, the policy reasoning, the match score, the edit diff: Help Scout has nowhere to put that. Tags help a little. Custom fields help a little more. Neither gives you a queryable history of how the agent thought about each ticket.
One way to solve this is a workspace like Dock that holds the interpretation layer next to the helpdesk. Help Scout stays the system of record for the conversation. Dock holds rows for each AI proposal, with a helpscout_conversation_id pointer linking back. Six weeks in, you can query which saved replies the agent suggests most, which get edited heavily, and which customers generate the most rework.
Why it matters
Small teams lose hours to context reconstruction. Every ticket starts from zero. An AI draft loop with persistent interpretation cuts that cost. Help Scout reports that 90% of customers rate an immediate response as essential, with 60% defining immediate as ten minutes or less. Drafts in the inbox shrink that window without expanding headcount.
Read the pillar guide on running customer support with AI.
FAQ
Does Help Scout have native AI features? Help Scout ships AI Assist for summarization and rewriting inside the reply editor. The workflow above wraps that with external drafting via the Mailbox API, giving you control over the prompt and the audit trail.
How does this compare to Zendesk's AI? Zendesk's AI suite assumes larger teams and deeper automation budgets. The AI Zendesk workflow covers that pattern. Help Scout suits teams that want lighter setup and human-in-the-loop sending.
Can Help Scout Beacon traffic use the same agent? Yes. Beacon chats land as conversations in the mailbox. The webhook fires the same way. Treat Beacon and email as one queue.
What if the agent suggests the wrong saved reply? The reviewer rejects or edits. The edit-diff loop downweights that reply for similar future intents. Accuracy stabilizes in six weeks.