What this automation does
Agents open a ticket and find a draft reply already waiting, grounded in what this customer actually ordered and what your policies actually say.
- Trigger on a new customer message, either on ticket creation or on any new public reply from the customer, but not on internal notes or agent replies.
- Assemble context. Pull the full thread, the customer’s recent orders or subscription from Shopify or your billing system, any tags from triage, and the three to five most relevant macros and help articles.
- Draft. A model writes a reply in your house style that answers every question in the latest message, uses real order numbers and dates, and marks anything it could not verify with a bracketed note such as
[confirm tracking number]. - Post as an internal note at the top of the ticket, with a list of sources used and a confidence flag.
- Agent edits and sends. The agent copies, edits and sends, or ignores the draft and writes their own.
- Capture the difference. When the ticket is solved, the workflow compares the draft with the reply actually sent and logs how much changed.
When it is worth building
Drafting pays off when your agents spend most of their day writing, not investigating. Email-heavy teams in ecommerce, SaaS and travel are typical. If a large share of replies are variations on the same ten macros with customer-specific details filled in, drafting will save real time.
It pays off less when most tickets need investigation in other systems before anyone can answer. In that case, the time goes into the lookup, not the writing, and a support agent with tools or better internal tooling is the bigger win. Also check your helpdesk first. Some now include basic AI drafting, and if it already reaches the data you need, use it. Our Zendesk AI comparison covers when the built-in option is enough.
Architecture
| Layer | What we use | Why |
|---|---|---|
| Trigger | Helpdesk webhook on new public customer message | Drafts are ready before an agent opens the ticket |
| Orchestration | n8n | Built-in Zendesk, Freshdesk, Help Scout and Shopify nodes; HTTP Request for Gorgias and Front |
| Customer context | Shopify, Stripe or your own order API | Drafts with real order numbers and dates need far less editing |
| Reference content | Macros and help articles in a small vector index | Keeps drafts aligned with approved wording and current policy |
| Style examples | A curated set of well-rated past replies per ticket type | The fastest way to make drafts sound like your team |
| Drafting model | Claude Sonnet 5.5 or a comparable OpenAI GPT model | Long threads and careful tone need a capable mid-tier model |
| Feedback store | Postgres or Google Sheets table of draft versus sent reply | The data you need to improve drafts and decide what to automate next |
How to build it
- Export solved tickets with their final replies and the CSAT or QA score where you have it. Pick 20 to 40 strong replies across your main ticket types as style examples.
- Clean up macros. Delete or update outdated ones. Every bad macro becomes a bad draft.
- Index macros and articles with their titles and IDs so drafts can cite which ones they used.
- Build the context step. Look up the customer by email, fetch their last few orders or current plan, and pass only the fields that matter (order number, items, status, dates, tracking link).
- Write the drafting prompt with the tone guide, the bracket rule for anything unverified, and an instruction to address every question in the latest message.
- Post as an internal note and add a visible label such as “AI draft, check before sending.”
- Pilot with three or four agents for two weeks. Ask them to flag drafts that were wrong, not just unhelpful, and review those together.
- Add the comparison step at ticket solve, and build a weekly view of edit size by ticket type.
The downloadable starter template below covers the Zendesk trigger, Shopify order lookup, drafting and the internal note.
What breaks in production
- Rubber-stamping. After a few weeks of good drafts, agents stop reading carefully. Keep the bracket convention for unverified facts, spot-check sent replies, and watch for replies sent within seconds of opening the ticket.
- Race with fast agents. An agent replies before the draft arrives, then a stale draft appears. Check whether an agent has replied since the trigger and skip posting if so.
- Drafting on the wrong message. The trigger fires on an internal note or a satisfaction survey reply. Filter on public messages from the requester only.
- Very long threads. Twenty-message threads with forwarded chains overflow useful context. Summarize older messages and keep the latest few in full.
- Policy promises. A draft offers a refund or discount because a past reply did. Keep money-related wording tied to current policy articles, and have the draft defer to the agent for exceptions.
- Language mismatch. A customer writes in Spanish and the draft comes back in English because the examples were English. Detect language and instruct the model to reply in it.
Timeline and cost
One helpdesk with one commerce or billing lookup typically takes two to three weeks. Lumeor builds reply drafting in the range of $5,000 to $12,000. The low end drafts from the thread and macros alone. The high end adds order or account lookups from more than one system, multi-brand tone guides and the draft-versus-sent reporting view. Model costs scale with ticket volume and thread length, and are typically small compared with agent time saved.
How to measure success
- Edit distance between draft and sent reply, by ticket type, tracked weekly
- Draft usage rate: share of replies where the agent kept most of the draft
- Handle time per ticket, before and after, on the same ticket types
- QA scores and CSAT on drafted replies compared with the prior period, to confirm speed did not cost quality
- Wrong-draft reports from agents, each reviewed and traced to a missing macro, stale article or context gap