Draft ticket replies from your help center and past tickets
For each ticket the system drafts a reply grounded in your help center articles and similar solved tickets, cites its sources and leaves the agent to edit and send. Simple, well-covered questions can be answered automatically once you decide to allow it. For teams with a large help center and slow replies.
Function: Support · Industry: Any industry
Draft ticket replies from your help center and past tickets
6 steps
At a glance
- Trigger
- A ticket is assigned to a queue or an agent.
- Typical inputs
- Ticket and conversation history
- Help center and internal knowledge articles
- Solved tickets with accepted answers
- Tone guide and banned-claims list
- Systems it usually connects to
- Zendesk
- Freshdesk
- Intercom
- Confluence
- How the work splits
3 rules, 2 model, 1 person steps
How it runs, step by step
6 steps, each tagged by what does the work. Rules are fixed logic, a model reads or drafts, and a person decides.
- Rules Fixed logic in code
- Model AI reads or drafts, with a confidence score
- Person A named person decides
- 1
Find the sources
ModelA model retrieves the relevant articles and similar resolved tickets, and ranks them by how closely they match the question.
- 2
Draft the reply
ModelThe reply is written in your tone, using only the retrieved material, with links to the articles it used.
- 3
Run the checks
RulesRules block drafts that promise refunds, give legal or medical advice or use phrases on your banned list.
- 4
Edit and send
PersonThe agent reads the draft, edits it and sends. The edit is recorded as feedback.
- 5
Auto-send where allowed
RulesFor categories you approve, such as password reset instructions, the reply goes out directly with a visible escalation link.
- 6
Report on drafts
RulesAcceptance rate and edit distance are tracked by topic to show where the knowledge base or drafts need work.
Where people approve
The decisions that stay with your team, whatever the confidence score says.
- Agents review every draft until a category is approved for auto-send.
- Support leads approve each auto-send category.
- Knowledge owners approve new articles the assistant may use.
KPIs to track
What to measure, before and after. Targets come from your own baseline.
- 01Share of drafts sent with light or no edits
- 02Median first-response time
- 03Reopened tickets after a drafted answer
- 04Articles updated because of a gap the drafts exposed
Questions about this workflow
What happens when the help center has no answer?
The draft says what is missing and asks the agent to answer, rather than filling the gap. The missing topic is added to a list for the knowledge team.
Can it refund or change an order?
Not from a draft. Actions that change money or data are separate, permission-limited steps with their own approvals, and are never sent as a drafted reply.
Which steps use rules, a model and a person?
Rules: run the checks; auto-send where allowed; report on drafts. Model: find the sources; draft the reply. Person: edit and send. Rules are fixed logic in code. A model reads or drafts and reports its confidence. A named person decides.
What do we need to start?
Typically: ticket and conversation history; help center and internal knowledge articles; solved tickets with accepted answers; tone guide and banned-claims list. You also need read access to Zendesk, Freshdesk, Intercom and Confluence, or the equivalents you use. A workflow audit measures your volumes and past cases first, then gives a fixed price for the build, so you can decide with real numbers.