Ship your first support agent in a weekend (without breaking anything)

I'm going to be upfront: before this project my technical high point was building a pivot table that made my boss say "ooh". I run a four-person support team at a B2B software company. About 320 tickets a week, most of them variations of eleven questions.

On a Friday I decided I'd build an agent over the weekend. On Sunday evening I nearly sent 400 customers an unfinished draft email. On Monday morning it went live properly, and it now handles first-touch on 38% of our tickets.

This is the whole thing, in order, including the near-disaster.

Saturday morning: don't build anything yet

Open that spreadsheet of 200 tickets. Read them. Actually read them — I know you think you already know what's in there. I did too. I was wrong about the ordering.

Sort them into three piles:

1. Pure information requests — "how do I reset my password", "where's my invoice", "does your product do X". Mine: 41%.
2. Needs a lookup — "what's the status of my order", "why was I charged twice". Requires touching a real system. Mine: 34%.
3. Needs a human — angry, complex, contract-related, or emotionally delicate. Mine: 25%.

Your first agent handles pile one only. Not pile two. Definitely not pile three. I know pile two is where the impressive demo lives. Resist.

This sorting exercise took me two hours and was the single most valuable part of the weekend. It's also the part everyone skips.

Saturday afternoon: write the knowledge, not the prompt

For each of your top ten pile-one questions, write the ideal answer as a human would send it. Short. In your voice. With the actual links.

Put them in one document. That's your knowledge base. It doesn't need to be fancy — mine is a Google Doc with ten headings. Ten good answers beats a 300-page manual dumped into a vector database, and I'll die on that hill.

Then write the instructions

Here's mine, more or less verbatim. Steal it:

You are a first-response assistant for [Company] support. You draft replies for human agents to review — you never send anything yourself.

Rules:
1. Only answer using the knowledge document provided. If the answer isn't there, say "ESCALATE" and one sentence explaining why.
2. Maximum 120 words. No preamble, no "I hope this email finds you well".
3. Match the tone in the examples: warm, direct, slightly informal. Contractions are fine.
4. Never mention prices, refunds, contracts, legal matters, or outages. Those always get ESCALATE.
5. If the customer sounds upset, say ESCALATE regardless of the topic.
6. Ask at most one clarifying question. If you'd need two, escalate instead.

Rule 5 is the one I'd add first if I were starting over. It's caught things a keyword filter never would.

Saturday evening: wire it up

In n8n, the whole workflow is six nodes:

1. Trigger: new ticket created in the helpdesk (webhook or polling — polling every 2 minutes is fine).
2. Filter: only tickets from the general support inbox, only in business hours to start with.
3. AI Agent node: your instructions above, your knowledge doc as context, model set to Claude Sonnet or GPT-4-class.
4. IF node: does the output contain "ESCALATE"?
5. Yes → add an internal note flagging it for a human, tag the ticket, stop.
6. No → save the draft as an internal note on the ticket. Not a reply. An internal note.

The near-disaster

At about 9pm on Sunday I changed step 6 from "internal note" to "reply" because it felt like the obvious next step, and then I clicked "execute workflow" to test it. On the live inbox. With 400 open tickets in the queue.

The only reason nothing sent is that my API key had hit the $20 cap I'd set on Friday. That cap saved my job. Set the cap.

(I now test everything against a duplicate sandbox inbox with three fake tickets in it. Took twenty minutes to set up. Should have been step zero.)

Sunday: run it in shadow mode

For a full week, the agent only writes internal notes. Your human agents read the draft, then write their own reply as normal. At the end of each day, log two numbers in a spreadsheet:

• Would you have sent that draft as-is? (yes / with small edits / no)
• If no — why not, in one line?

Week one for us: 44% sendable as-is, 31% small edits, 25% no. The "no" reasons were nearly all the same three things, which I fixed by adding three sentences to the instructions. Week two: 71% sendable as-is.

Week three: let it go live, carefully

Our rollout order, which I'd repeat:

1. One-click send. Agent drafts, human reads and hits send. Still 100% human-approved, but the typing is gone. Biggest single time saving of the whole project.
2. Auto-send for three question types only. Password resets, documentation links, opening hours. Low risk, high volume.
3. Everything else stays human-approved. Still true today, eighteen months on. I have no plans to change it.

What it's done for us, honestly

The team-size row matters to me. This wasn't a headcount project. My team stopped writing the same eleven answers and started doing the harder work they're actually good at. Two of them have told me it's the best change we've made in years.

Useful links

• n8n AI tutorial — the official walkthrough for the Agent node
• Zendesk Tickets API
• Anthropic prompt engineering guide
• Our n8n teardown and the guardrails thread

---

Replies (44) — highlights