There is a hidden tax on every support team's day. Before an agent can help a single customer, they spend two or three minutes on the same cognitive work every single time: read the ticket, figure out what type of problem it is, decide which queue it belongs in, apply the right label, and then start actually helping. Multiply that by 80 tickets a day and you have burned roughly two and a half hours of agent capacity on sorting, not on solving.
I set up a two-tool stack about a year ago that takes that sorting work off the desk entirely. Zapier catches the inbound ticket, classifies its intent, and writes a tag back to Help Scout. Help Scout Workflows read that tag and route the ticket to the right inbox or team automatically. By the time an agent opens a conversation, triage is already done and the right person is already looking at it.
Here is exactly how to build it.
Step 1: Design a taxonomy that is actually finite
The failure mode in most triage systems is too many categories. If you have 14 ticket types, your classifier breaks on every edge case, agents spend time correcting bad tags, and you've added process without removing friction.
My rule: six types maximum, with a single fallback bucket. For a typical SaaS or service business, these six hold up well:
- billing (payment issues, plan changes, invoices, refund requests)
- bug (something is broken, the product is not behaving as expected)
- feature-question (how do I do X, where is the setting for Y)
- account (login, access, password, settings)
- cancellation (intent to cancel, downgrade requests, churn signals)
- general (anything that doesn't map cleanly to the above five)
Keep the labels lowercase and hyphenated. Help Scout's Workflow conditions match on tag values exactly, and inconsistent casing causes silent routing failures that are genuinely annoying to debug.
One more decision to make before you build: where does general route? For me it stays in the main queue with no special routing. That means everything in the main queue that an agent opens manually is either genuinely ambiguous or something the classifier couldn't parse. If general runs above roughly 20 percent of your volume, the taxonomy is too narrow.
Step 2: Build the Zapier classifier
Zapier sits in the middle. The trigger is either "New conversation in Help Scout" (available on Help Scout's Standard plan and above) or "New email in Gmail / Outlook" if you forward to a shared inbox first. Either works; the Help Scout trigger is cleaner because the Zap fires only on real customer messages, not internal notes.
The action that does the classification is AI by Zapier (the built-in AI step, no extra API key required). I pass the first 500 characters of the message body along with the subject line and a short classification prompt:
Classify this customer support message into exactly one of these categories: billing, bug, feature-question, account, cancellation, general. Return only the category label, no other text.
Five hundred characters catches the intent on roughly 90 percent of real tickets. The issue is almost always stated in the first paragraph. If you want more control over the model or the prompt, swap the AI by Zapier step for a Webhooks step to the Claude or OpenAI API, but for most teams the built-in step is enough to start.
The final action in the Zap is Update Conversation in Help Scout, using the tag field to apply whatever label the AI step returned. That single tag is the handshake between Zapier and Help Scout Workflows.
Step 3: Set up Help Scout Workflows for routing
Once the tag is on the conversation, Help Scout Workflows take over. This part requires zero code. It lives entirely inside Help Scout's Automation section and takes about 20 minutes to configure the first time.
Create one Workflow per category, using the condition "Conversation tag is [tag name]":
- billing: Assign to the Billing mailbox, set priority to Urgent, send an internal note tagging the billing lead
- bug: Assign to the Support Engineering team, add an internal note that says "needs reproduction steps"
- feature-question: Assign to any available agent in the main queue, optionally suggest a saved reply folder
- account: Assign to the Account Access mailbox
- cancellation: Assign to a specific named senior agent, set priority to Urgent (this one must not go to a general queue)
- general: Leave in the main queue, no special routing
The cancellation rule is the most important one to get right. A cancellation intent needs a person who knows the retention conversation, and they need to see it fast. Routing a churn signal to a general queue and letting it sit loses deals that a 10-minute response would have saved.
The human override rules
Not every ticket should run through the Zapier classifier. I bypass it for three categories, and both of the first two require a manual tag applied in Help Scout before the Zap fires.
VIP accounts. Customers on enterprise plans or with a dedicated account owner get a VIP tag applied when their account is created. If a conversation is from a VIP-tagged customer, a separate Workflow routes it directly to their named contact, regardless of what the classifier would say. Mis-routing a VIP and sending them to a general queue is worse than no automation.
Repeat contacts. If a customer has opened more than three tickets in the past 30 days, a senior agent manually applies a "needs-senior-review" tag that triggers a separate Workflow routing all their messages directly to a named person. Repeat contact is almost always a signal of something unresolved upstream, and it deserves a human who has the full account history in front of them.
Active incidents. When we're working an outage, I pause the Zap. What looks like a billing or account ticket during an incident is often actually a bug report, and mis-routing floods the wrong queue. Pausing the Zap takes about ten seconds in Zapier's dashboard.
What to watch in the first month
Run a weekly spot-check for the first four weeks: sample 20 conversations from each tagged category and verify the tags are correct. The classifier makes most of its mistakes on very short tickets (single sentences), on messages written in informal or non-English language, and on tickets that contain two separate issues in one message.
When you see patterns in the misclassifications, add example sentences per category to the prompt rather than adding more categories. Better examples consistently outperform a bigger taxonomy. Also check on day one that your prompt's output matches your Workflow conditions exactly: a Zap returning "Billing" (capital B) routes nothing when the condition watches for "billing".
What the setup actually changed
After running this for roughly six months, average first-response time dropped by something close to 25 percent. The bigger change was not the speed. It was the quality of the first response itself.
Agents stopped spending their opening message asking "can you describe the issue?" They opened a billing ticket and immediately pulled up the account status. They opened a bug ticket and started asking for the specific steps to reproduce. The first message had a job to do from the first line, because the agent already knew what kind of conversation they were walking into.
The "general" bucket ran at about 14 percent of volume, which felt right. If general hits above 20 percent, I know the taxonomy needs a new category. If it drops below 10 percent, I know the classifier is force-fitting tickets into categories that don't quite fit.
The automation doesn't draft replies, it doesn't resolve anything, and it doesn't touch anything a customer sees. It just makes sure the right person is looking at the right ticket with the right context already in front of them.
Two tools, one job
Zapier and Help Scout divide the responsibility cleanly. Zapier connects external services, calls the AI, and writes the tag. Help Scout Workflows act on what's already inside the inbox: routing, assigning, prioritising. Neither tool tries to do what the other does better.
Agents still do all the actual support. What changes is that they start every conversation already knowing what kind of problem they're solving. The sorting tax disappears and that time goes to the customer instead. If you're on Help Scout and want to compare notes on the Zap structure or Workflow conditions for your specific ticket mix, drop me a line.
Comments