From client enquiry to AI-drafted reply, in under five minutes.
An n8n + Claude pipeline that classifies, drafts, and briefs every inbound enquiry — leaving the Client Manager two minutes of review and a click to send. Same workflow we deploy for client operations. Built and live in production on USource's own intake.
Every new enquiry costs a Client Manager 20 to 30 minutes. The work isn't the reply, it's the writing.
Reading the enquiry, classifying the service fit, looking up the right tier, finding similar past replies, drafting something personalised, checking the tone. Most of the time goes into producing one good first reply. Multiply by every enquiry, every week. The drafting eats the day.
Open ticket, scan body, scan attached company info
Match request to USource service tier; assess lead quality
Pricing notes, similar past replies, internal docs
Tone, structure, brand voice, links, sign-off
Final read, format check, send
Schedule Trigger polls Freshdesk every 5 minutes for new tickets.
Only "Digital Staff Enquiry" tickets pass; full description fetched.
One API call: service classification, lead score, draft reply, client brief.
Code node converts JSON to Freshdesk-ready HTML with brand voice locked.
Draft lands as private note. Review, tweak, send. Done.
The 10–15 minute drafting block is not a discipline problem — it's the cost of writing a personalised reply from scratch every time. Claude collapses it to thirty seconds.
Five stages. Four automated, one human. Brand voice locked, judgement preserved.
The pipeline does the slog: read, classify, look up, draft. The Client Manager does what only a human should — review, edit, send. Auto-send was on the table. We said no, on purpose.
Schedule polls Freshdesk every five minutes.
n8n runs a Schedule Trigger every five minutes, pulling any new tickets from Freshdesk. Polling instead of webhooks keeps the setup simple. No Freshdesk admin config, no firewall rules. The 2–3 minute average lag is well below the human reply window, so latency is a non-issue.
Only enquiries proceed. Full ticket body fetched.
A Filter node keeps only tickets with subject "Digital Staff Enquiry" — support tickets, internal notes, and unrelated correspondence drop out. Freshdesk's "Get many tickets" omits the description body, so a second "Get a ticket" node fetches the full text for the matched ticket. Clean payload ready for Claude.
- →Tickets pulled in window: 7
- →After filter (Client Enquiry ABC): 1
- →Full description fetched: ✓
- →Subject + body + contact: passed forward
- →Ticket ID + metadata: preserved
One API call. Four jobs done.
A single Anthropic Claude call handles classification, lead scoring, reply drafting, and client briefing. The system prompt is locked: USource brand voice, the no-em-dash rule, the no-contractions rule, the no-pricing rule, the under-150-words rule. Output is structured JSON — no free-text parsing required downstream.
JSON to ticket-ready HTML.
A small Code node parses Claude's JSON response into flat fields, converts newlines to
tags for Freshdesk's HTML rendering, and sets the private flag as a proper boolean. Brittle string parsing avoided. Anything Claude misses gets a structured fallback — the pipeline never crashes mid-run.
- →Fields extracted: 5
- →HTML formatting applied: ✓
- →Private flag: {{ true }} (boolean, not string)
- →Fallback values set for missing fields
- →Ready for HTTP Request → Freshdesk
Drafts go to humans. Sends stay with humans.
The draft posts as a private internal note on the original Freshdesk ticket — visible only to the team. The Client Manager opens the ticket, reads the brief and the draft, makes any tweaks, and clicks send. Brand voice protected. Edge cases caught. Two minutes of skilled review beats twenty minutes of cold drafting.
Raw enquiry to AI-drafted reply, sitting in the ticket.
One real enquiry tracked through the pipeline. Inbound ticket, AI processing, drafted reply waiting in the same ticket as a private note. Click any image to enlarge.
The raw enquiry as it lands. Subject locked. Filter passes it through to the pipeline.
Lands as a private note in the same ticket. Five minutes after the enquiry arrived.
Drafts go to AI. Sends stay with humans.
Auto-send is technically trivial, one HTTP request away. We chose not to. Every AI-drafted reply lands as a private note for the Client Manager to read, edit, and click send. Brand voice protected. Pricing claims checked. Edge cases caught — the test enquiry that came from a USource email got correctly flagged as an internal test, not replied to as a real lead.
Two minutes of skilled review per enquiry is the difference between AI-Enhanced and AI-Replaced. We chose AI-Enhanced, and every Client Manager edit becomes a training signal for the next draft.
One enquiry is a demo. Every enquiry is a system.
Drops from 20–30 minutes of manual work to ~3 minutes of structured review. Multiplied across every inbound, that's days back per month.
From the moment a ticket lands in Freshdesk to the AI-drafted private note appearing in the same ticket. The CM walks in to a queue that's already done the thinking.
Schedule fires every five minutes around the clock. Weekend enquiries are drafted by Monday morning standup. No hire required.
This is the workflow we run for client work.
Inbound lead processing, support triage, internal request routing, document summarisation. Same AI-Enhanced delivery, ported to your tools and your team. Days, not months.