Draft First Responses Grounded in the Ticket and the KB

Tier: Foundation · Workflow area: Ticket handling & triage · ~20 minutes

This playbook covers drafting personalized, tone-correct first responses with Freddy Copilot, Zendesk AI, or a plain assistant — with grounding rules that keep the draft tied to your actual knowledge base and a 30-second checklist that catches invented ETAs and hallucinated steps.

When to use

• First touch on common categoriespassword, VPN, printer, access requests: high volume, well-documented, where a KB article already holds the right steps.
• Tone-sensitive repliesfrustrated user, VIP, or a reply you have rewritten twice already. AI is genuinely good at "firm, warm, no corporate filler."
• Language supportreplying in a language you read but do not write fluently; draft in your language, have AI translate, verify key terms.

When NOT to use

• Incidents with no known fixif the KB has nothing, the model has nothing to ground on and will improvise steps. Send a short human acknowledgment instead.
• Anything contractualSLA disputes, refunds, security disclosures. These carry legal weight; templates approved by a human owner, not generated text.
• When the draft takes longer than typinga two-line "your account is unlocked, please try now" needs no AI.

The recipe

1. 1.

   Gather the grounding material first. The draft is only as honest as its inputs: the ticket description, the verified summary if one exists (see pb-02), and the one or two KB articles that actually apply. In Freshservice, Freddy Copilot's reply suggester pulls linked articles automatically; in Zendesk, the agent copilot works from linked help-center content. In a plain assistant, you paste them — which has the advantage that you chose them.

2. 2.

   Prompt with grounding and tone rules.

   Copy

   Draft a first response to this IT support ticket.
   
   Ticket:
   [paste subject + description + user name and role]
   
   Knowledge base article(s) — the ONLY source of technical steps:
   [paste article body or the exact relevant section]
   
   Rules:
   - Technical steps must come from the KB text above. If the KB does not cover
     the user's situation, say we are investigating and ask the ONE most useful
     clarifying question instead of giving steps.
   - Do not state or imply any resolution time, ETA, or "this will fix it."
   - Do not promise callbacks, escalations, or refunds.
   - Tone: warm, direct, no filler ("I hope this email finds you well" is banned).
   - Address the user by name. Reference their specific symptom in the first line
     so it does not read like a template.
   - Under 150 words. End with what happens next and what we need from them.
   

3. 3.

   Run the no-promises checklist. Thirty seconds, every draft, no exceptions:

   • •
     No invented ETA — search the draft for "within", "by", "shortly", "soon". Delete or replace with what the SLA actually commits to.
   • •
     Every step exists in the KB — each instruction traces to the pasted article. A plausible extra step is a hallucination even when it is technically correct advice.
   • •
     No policy promises — nothing about refunds, exceptions, approvals or "we'll make sure this never happens again."
   • •
     Right person, right system — the name matches the requester (not someone quoted in the thread), and product names match your environment (the model will happily say "Google Workspace" to your M365 shop).
   • •
     Links are real — click every URL. Assistants fabricate plausible KB links.
4. 4.

   Compare a good and a bad draft once, as calibration. A passing draft looks like:

   Copy

   Hi Meera — thanks for flagging that the VPN drops every 20–30 minutes on the
   Pune office Wi-Fi.
   
   Two quick checks from our side first:
   1. In the GlobalProtect app, click the gear icon > Settings > General and
      confirm the portal address reads vpn.yourco.com (not vpn-old.yourco.com).
   2. If it does, please send us the time of the next disconnect — we'll match
      it against the gateway logs.
   
   Once we have that timestamp we can see whether this is your client or our
   gateway. Reply here and it comes straight to me.
   

   Grounded steps, one specific ask, zero promises. The failing version of this draft ends with "this should resolve your issue within 24 hours" — which is the exact sentence you will be apologizing for tomorrow.

5. 5.

   Personalize the one line that matters, then send. Add or adjust the opening line so it reflects the user's actual words. That single edit is what separates "helpful engineer" from "bot with my name on it."

Make it yours

• Freddy Copilot / Zendesk agent copilotuse the built-in suggester for speed, but keep step 3 manual: vendor copilots ground on linked articles, yet none of them enforce a no-ETA rule. The checklist is yours.
• Response macros + AIfor your top five categories, save the prompt (with the KB section embedded) as a snippet, not the response. You get consistency and per-user personalization instead of the canned-reply smell.
• MSP / multi-client desksadd a rule line per client: "Client context: refer to IT as 'the Contoso IT team', never mention our company name, sign as service desk." Wrong-client branding in a first response is a churn-level error.

Quality check

• Checklist compliance is auditableduring the pb-01 daily sample, grade first responses too: any sent draft containing an invented step or ETA counts as a defect, same as a misroute.
• Measure reopens, not speedfirst-response time will improve on its own; the number to watch is replies that bounce back with "that didn't work" because a hallucinated step shipped.
• Test tone changes in a sandboxtrying a new prompt style? Run it against ten closed tickets and read the outputs before pointing it at live users.
• Never auto-sendevery AI draft passes through human hands. The day drafting becomes sending, you have deployed an unsupervised chatbot with your signature.