Skip to content

Create a brand-tuned support reply drafting API

Generate empathetic first-draft replies for customer agents using ticket context, macro rules, and strict policy boundaries.

Copy the prompt and paste it into the AI assistant you already use.

Steps to ship a safe reply-drafting API

  1. 01

    Provide structured ticket state

    Pass verified account tier, shipping state, or bug status fields alongside the raw message.

  2. 02

    Codify non-negotiable guardrails

    State explicit rules against inventing timelines, undocumented features, or policy exceptions.

  3. 03

    Include curated few-shot examples

    Embed three to five real solved ticket replies in the builder to anchor greeting and closing style.

  4. 04

    Output agent review metadata

    Return confidence and missing-info flags so agents know when manual verification is required.

How it works in practice

POST /v1/draft-reply

{ "customer_msg": "My export stuck at 90% for 20 mins.", "account_plan": "Pro", "known_incident": "CSV worker lag resolved 14:10 UTC" }

Agent-Ready Draft Payload

{ "draft_reply": "Hi! Thanks for flagging this. We resolved a brief CSV worker slowdown at 14:10 UTC—could you refresh and retry your export? Reply here if it stalls again and we'll pull it manually.", "needs_human_check": false }

Agents receive accurate, grounded drafts ready for one-click approval instead of hallucinated promises.

Common questions

How do you stop the API from promising unauthorized refunds?

Explicit negative constraints in the system instruction plus grounded account-status inputs ensure the endpoint only cites approved facts.

Is this meant to auto-send emails to customers?

Most teams route drafts into agent review queues first, using confidence and missing_context flags to highlight tricky tickets.

Can we match formal enterprise and casual startup tones?

Pass a tone or customer_segment field in the request JSON so the endpoint adapts phrasing per brand or tier.