Summarize a Messy Ticket Thread on Pickup

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

Use the native summarizer where you have one (Now Assist, Freddy Copilot), and a disciplined prompt over the exported thread where you do not. Either way, the verification step is the same: every claim in the summary must be traceable to a line in the thread.

When to use

• Reassigned or escalated ticketsanything with 10+ updates, multiple agents, or a reopen. This is the highest-value single use of AI at the service desk.
• Returning from leave or shift changeyour own tickets, three days later, are someone else's tickets.
• Long-running problem recordsweeks of intermittent updates where the original symptom statement no longer matches reality.

When NOT to use

• Short threadsunder five updates, just read it. The summary costs more than it saves and adds a paraphrase layer where none is needed.
• Legal, HR or disciplinary ticketsthreads that may become evidence should be read verbatim, and often should not leave the ITSM tool at all.
• As a substitute for reading before a customer callsummarize to orient, but skim the raw thread before you promise anything to the user; summaries drop tone and frustration signals.

The recipe

1. 1.

   Try the native button first. ServiceNow Agent Workspace exposes a Now Assist Summarize action on the incident; Freshservice has Freddy Copilot's ticket summary in the right-hand pane. Native summarizers see the full journal including work notes, which an export can miss. If the output already answers "what's broken, what's been tried, what's next", verify (step 4) and stop here.

2. 2.

   Otherwise, export the full conversation. From Freshservice:

   Copy

   curl -s -u "$FS_API_KEY:X" \
     "https://yourdomain.freshservice.com/api/v2/tickets/12345/conversations?per_page=100" \
     | jq -r '.conversations[] | "\(.created_at) | \(.from_email // "agent") | \(.body_text)"' \
     > ticket-12345-thread.txt
   

   From ServiceNow, pull the journal (comments and work notes):

   Copy

   curl -s -u "$SN_USER:$SN_PASS" \
     "https://yourinstance.service-now.com/api/now/table/sys_journal_field?sysparm_query=element_id=<incident_sys_id>^ORDERBYsys_created_on&sysparm_fields=sys_created_on,element,value" \
     | jq -r '.result[] | "\(.sys_created_on) | \(.element) | \(.value)"' > thread.txt
   

   Copy-pasting the activity stream from the UI is a legitimate fallback — completeness matters more than elegance.

3. 3.

   Prompt for the structured summary. The format is fixed; the rules are the important part:

   Copy

   Summarize this IT support ticket thread for the engineer picking it up.
   
   Format exactly:
   SITUATION: who is affected, what is broken, since when, business impact (2–3 lines)
   TRIED: everything attempted so far — by whom, when, and the RESULT of each attempt (bullets)
   ENVIRONMENT: device, OS, app versions, network/location details stated anywhere in the thread
   OPEN QUESTIONS: what has been asked but never answered
   NEXT: the 1–3 most promising next steps implied by the thread
   
   Rules:
   - Use only facts present in the thread. If something is unknown, write "not stated" — never infer.
   - Quote error messages verbatim, in quotes.
   - If two updates contradict each other, surface the contradiction explicitly.
   
   Thread:
   [paste ticket-12345-thread.txt]
   

4. 4.

   Verify before you trust. Take 2–3 minutes: for each TRIED bullet, find the update it came from; confirm the verbatim error quotes actually appear in the thread (search the raw text); confirm nothing in NEXT was already attempted and buried mid-thread. A summary that fails any of these gets corrected by hand, not re-rolled.

   A good verified output looks like:

   Copy

   SITUATION: A.Rao (Finance, VIP) cannot open shared mailbox "AP-Invoices" in
   Outlook desktop since Mon 09:10 IST. Month-end payment runs blocked. OWA works.
   TRIED:
   - L1 (Mon 10:02): recreated Outlook profile — no change
   - L1 (Mon 11:40): confirmed FullAccess via management portal — permission present
   - User (Tue 08:15): error persists after reboot: "Cannot expand the folder.
     The set of folders cannot be opened."
   ENVIRONMENT: Win 11 23H2, Outlook M365 build 16.0.17531, cached mode, Hyderabad office.
   OPEN QUESTIONS: whether automapping was disabled when access was granted (asked Mon, unanswered).
   NEXT: check automapping; test with mailbox added as secondary account; review recent moves between databases.
   

5. 5.

   Post it where the next person looks. Paste the verified summary as the newest work note, labeled [AI summary — verified <your initials> <date>]. Unlabeled AI text in a ticket erodes trust the first time someone finds an error in it; labeled and verified text builds it.

Make it yours

• Jira Service Managementexport via GET /rest/api/2/issue/{key}/comment, and add "distinguish customer-visible comments from internal comments" to the prompt rules, since the two often tell different stories.
• Email-heavy shopsforward the mail chain to yourself, save as .eml/text, and add a rule: "collapse quoted duplicate text; each message counts once." Quoted-reply pyramids are the classic token-waster and confuser.
• Sensitive threadsrun the identical prompt against a local model (ollama run llama3.1:8b) so the thread never leaves the machine; expect a slightly worse summary and budget an extra verification minute.

Quality check

• Traceability is the barevery factual claim in the summary maps to a specific update. "Sounds right" is not verification.
• "Not stated" appears where it shoulda summary with no unknowns on a messy ticket means the model filled gaps. Treat that as a failed output.
• Contradictions surfaced, not smoothedthreads where the user says "still broken" after L1 says "resolved" must show that conflict; it is usually the whole story.
• Timebox itexport + prompt + verify + post should be under 10 minutes. Longer means the thread was short enough to just read.