Writing · AI & Automation · Dec 11, 2025 · 5 min

Two Zapier + GPT flows for a one-person support desk

One support rep, one inbox, four departments' mail. I built two Zapier flows with a GPT-4o mini step in the middle: a review responder and a routing layer. They took the repetitive and misrouted work off her plate, with a human approving every reply.

6 stepsone afternoon to build
1 → 3 platformssame single rep

This one starts with a person, not a tool: a single support rep covering Trustpilot reviews, product questions, and a steady stream of tickets that actually belonged to other departments. Before I touched any software, that was the system: one inbox, one human, every category of work mixed together.

One inbox, four departments' mail

Before building anything, I spent a day sitting next to the rep watching the queue. The bottleneck was the mixing. Urgent work sat in the same queue as repetitive work. Every switch cost her a beat and bought nothing. A review reply, a product question, and an accounting ticket all demanded the same attention, and only one of them belonged to support in the first place. The rep was effectively triaging four departments' mail to answer one department's questions.

One inbox · one rep
Support tickets · reviewshers
Sales ops mailnot hers
Accounting · product mailnot hers
The starting system
1 of 4 streams belonged to support
The starting system: one rep triaging four departments' mail (support, sales ops, accounting, and product) through a single inbox, when only one of those streams was actually hers.

Watching her for a day ruled out the default answer. The job was to take the repetitive and misrouted work off one person's plate.

Build around the pain

I started from the pain the rep already had and wrapped the workflow around it. Nobody had to learn a new tool. I built two pieces on that principle. Both run on Zapier with a GPT-4o mini prompt step in the middle, no agent framework, no custom code. It is the same pattern behind the 200+ workflows I run in marketing ops.

The review responder

The build is unglamorous. Trustpilot's notification email hits a Zapier email parser, the review text gets posted into a dedicated Slack channel, and a GPT-4o mini step drafts two replies, one formal and one conversational. The rep picks one, edits if needed, posts it. Six steps, one afternoon to build.

trustpilotemail inzapierparsesgpt-4o mini2 draftsreppicks & postsa human still presses send
The whole review responder: notification email → Zapier parser → Slack → two GPT-4o mini drafts, and a human picks the winner. Six steps, and they later stretched from one review platform to three without adding headcount.

The first version was embarrassing. The drafts read like a hotel apologizing. “we truly value your feedback” twice in three sentences. Nobody would believe a human wrote them, which defeats the entire point of replying to a review. The fix was a better prompt. I pasted ten of the rep's real past replies in as examples and capped the length. After that, the drafts sounded like us on a good day.

Once Trustpilot held up for a few weeks, I pointed the same workflow at G2 and Capterra notification emails. Three platforms, one zap, still one rep.

The routing layer

The second piece I shipped was a light AI classification layer on inbound tickets. Mechanically: the helpdesk fires a trigger on every new ticket, a webhook hands the subject and body to GPT-4o mini, and the model returns exactly one label, support, sales ops, accounting, or product. Zapier writes the label back as a tag, and a plain helpdesk rule assigns the ticket by tag. “Light” is deliberate: the model's only job was to read a ticket and decide where it belonged. No orchestration. A classifier in front of an inbox.

I built three guardrails: doubt defaults to support, refunds count as support, and every drafted reply waits for a human. Anything the model wasn't confident about stayed in support, because a wrongly forwarded ticket costs more goodwill than one the rep forwards by hand. And that rule earned its keep early: the first week, the classifier kept sending refund requests to accounting, who wanted invoices and only invoices. One line I added to the prompt, “refunds are support,” fixed it. Classifiers are literal; you find the edge cases by shipping.

Every new ticketa classifier in front of an inboxHelpdesk triggerfires on every new ticketGPT-4o minigets the subject and bodyTag written backa plain rule assigns by tagexactly one label comes backFour departments, four labelsthe label list came from a day of tagging tickets by handSUPPORTSupport ticketsreviews and product questionsSALES OPSSales ops mailskips support entirelyACCOUNTINGAccounting ticketsinvoices, and only invoicesPRODUCTProduct mailassigned straight from the tagGuardrailsUncertain stays in supportRefunds are supportDrafted replies wait for a human
The model's only job is reading a ticket and deciding where it belongs, with no orchestration behind it. A wrongly forwarded ticket costs more goodwill than one the rep forwards by hand, so doubt defaults to support.
Every AI-generated review reply landed as a draft. A human chose and approved every one, so the voice stayed hers.

What actually improved

Honest caveat: I never wired a dashboard to this project, so there's no headline percentage. Here's what visibly changed:

  • Review replies stopped queueing, drafts were waiting in Slack before the rep even opened the review, so responses went out while the review was still fresh instead of whenever a quiet hour appeared
  • Coverage tripled without headcount, one platform became three (Trustpilot, G2, Capterra), answered by the same single rep
  • Misrouted work disappeared from the queue, sales ops, accounting, and product tickets skipped support entirely instead of taking a bounce through the inbox first
  • We finally saw the workload, the tags from my classifier gave us the first honest picture of what support actually spent time on, which no one had counted before because the counting itself was work

Do this before you open Zapier

If you're about to build something similar, don't start in the automation tool. Sit next to your support rep for a day and tag every ticket by hand: yours, sales ops', accounting's, product's. That tag list becomes your classifier's labels, the awkward edge cases (refunds!) become lines in your prompt, and you'll find out which part of the queue is even worth automating.

The zap took an afternoon. Knowing what it should do took the day of sitting there. Spend the day first.

Suggested posts