Six places to search. One draft to review.
An AI support copilot inside Freshdesk. It researches each ticket across the client's knowledge, drafts the reply and scores its own confidence. Support engineers review the draft, and only replies that clear every safety gate go out on their own.
- Researches: past tickets, product docs, Jira records, attachments and earlier fixes, in one pass.
- Drafts: a grounded, customer-ready response with the evidence behind it.
- Routes safely: human review by default, direct reply only through five gates.
Thanks, but I am still seeing the same error on another package.
Meet our client
- Client
- PE-backed DevOps software company
- Ownership
- PE-backed
- Industry
- Enterprise Software, DevOps & Configuration Management
- Market
- Global Enterprise
In a nutshell
The client, a PE-backed enterprise software company, builds platforms for DevOps automation, application deployment, configuration management and release orchestration. We built a support copilot that reads each Freshdesk ticket, retrieves the relevant knowledge from across the company, drafts a grounded response and routes it through review and approval, so support engineers spend less time searching and more time resolving.
Every answer started with a search
To answer one ticket, an engineer often checked six places by hand, then wrote the reply from scratch. The quality of the answer depended on who picked the ticket up.
Hi, thanks for the update. I'm looking into the error on the second package and will
Past tickets, product documentation, Jira, internal knowledge, attachments and earlier fixes, then a reply from a blank page.
- Research time: engineers spent significant time investigating each issue across systems.
- Uneven answers: response quality varied with each engineer's experience.
- Scattered knowledge: documentation, tickets and engineering records lived in separate tools.
- Unused history: past resolutions were hard to find and reuse at scale.
- Repetitive drafting: similar responses were written again and again.
- No quality signal: there was no view of how reliable a response was before it went out.
Speed was easy. Trust was the hard part.
An AI that answers customers on its own was never the goal. Eight requirements shaped the build. Tap each one to see how it was met.
A copilot that lives inside Freshdesk
Ten n8n workflows built around one core orchestration flow, with Claude for drafting, Qdrant for retrieval and Postgres for memory. Tap any part to see its job, or trace a ticket through the system.
Freshdesk
Orchestration · n8n
AI & knowledge
Human review
Memory
- One core flow: a 58-node n8n workflow carries every ticket event from webhook to reply or review.
- Modular APIs: drafting, attachments, context, editing and approval each run as their own workflow, so each can change safely.
- Retrieval-augmented drafting: Claude writes from knowledge retrieved from Qdrant, indexed with OpenAI embeddings.
- State you can audit: sessions, context versions, drafts, confidence and routing decisions are kept in Postgres.
Follow a customer message through the copilot
Pick one of the four messages from the production test plan and send it. Watch which of the thirteen stages run. Some messages need no AI at all.
Draft ready for human review
A real follow-up gets a fresh, researched draft. This one scored 54, below the bar for a direct reply, so it went to a support engineer as a private note.
Try the five gates yourself →- 01Webhook received
- 02Ticket and conversation loaded
- 03Mailbox and assignment checked
- 04Latest message classified
- 05Skip, close or resolve
- 06Duplicate check (context hash)
- 07Existing-context check
- 08AI draft generated
- 09Session saved, note built
- 10Older AI notes tidied (dry run)
- 11Final routing decision
- 12Direct public reply
- 13Human review note
Every draft is built from the client's own knowledge
The Draft and Edit API gives Claude the ticket, its history and the company's knowledge, and gets back a structured draft with a confidence score. Tap an input or a part of the draft.
Six kinds of input go in. A draft in five parts comes out, with a confidence score that decides where it goes next.
Attachment intelligence
Supported attachments are downloaded, then read or analyzed so the draft can reason over screenshots, logs and documents. Results are cached by file fingerprint, so the same file is never processed twice, and count and size limits keep processing predictable.
Confidence score
Claude scores its own confidence in each draft, from 0 to 100, and returns the score with the draft and the context it used. It is the copilot's signal of how reliable a draft is before anything goes out. The score is saved with the draft in Postgres, and at routing a score of 60 or below sends the draft to a support engineer.
Five gates stand between a draft and a customer
A public reply goes out only when all five pass. Change the conditions and watch where the ticket goes.
Private note for human review
The draft and a review link are posted as a private Freshdesk note. A support engineer takes it from here.
The support engineer stays in charge
The private note links to a Review Interface. There the engineer can edit by hand, ask the AI to revise, add context and approve. Try each action.
- Before: one ticket often meant checking six places by hand, then writing the reply from a blank page.
- Routine messages: thank-yous, requests to close and reports that the issue is fixed are handled without an AI draft, and the action is logged.
- Real follow-ups: by default, the ticket already carries a private note with a researched draft, the context it used and a review link.
- The engineer's part: open the review link, check the evidence and references behind the draft, edit it or ask the AI to revise it, then approve it and send it to the customer.
- 01Private note with link
- 02Ticket and draft loaded
- 03Edit, revise, add context
- 04Approve
- 05Approved note in Freshdesk
- Open from Freshdesk: the private note carries the draft and a link to the Review Interface.
- Edit by hand: changes are saved to the ticket session and approval is enabled.
- Ask the AI to revise: a new draft is generated from the engineer's instruction and context.
- Add context: notes and files are stored and used in the next draft.
- Approve: the final draft is converted to HTML, marked as approved and posted as a private note.
- Everything recorded: drafts, edits and decisions stay in Postgres for audit and debugging.
A copilot for support engineers, not a chatbot for customers
FAQ bots answer from static content and reply on their own. This copilot researches like an engineer and hands the decision to one.
Faster answers, with human judgment intact
Fragmented support knowledge became a retrieval and response system built into the support workflow.
Automated triage and context
Each ticket is classified and its context gathered before an engineer opens it.
Faster access to history
Past tickets and resolutions surface on their own.
Less manual research
Engineers start from a researched draft instead of a blank reply.
AI-assisted drafting
Customer-ready drafts with recommended steps and supporting evidence.
One view of knowledge
Documentation, tickets and engineering records retrieved together.
More consistent replies
Every response starts from the same knowledge and structure.
Human-controlled approval
People approve, and automation runs only through five gates.
Built to scale
Modular workflows and stored state grow with ticket volume.
Faster answers. People still decide.
A copilot that does the research and the drafting inside the tools your support team already uses.
- 01Your business objective
- 02The technical context
- 03The right next step