The Tagging Problem

Your support tool has a tagging system. Nobody uses it consistently.

Agent A tags a billing question as "billing." Agent B tags the same type of question as "payment." Agent C doesn't tag at all because they're busy. By the end of the month, your analytics are useless because the data is messy.

Manual tagging has two problems: it takes time (even 10 seconds per ticket adds up), and it's inconsistent (different people use different labels for the same thing).

How AI Tagging Works

AI classification reads the customer's message and assigns an intent tag automatically. Before any human sees the ticket, it already has a label:

• "I was charged twice" → billing_dispute
• "How do I export my data?" → feature_how_to
• "Your app crashes when I click Settings" → bug_report
• "Can you add dark mode?" → feature_request
• "I want to cancel my subscription" → subscription_cancellation

This happens in 100-200 milliseconds. Every ticket gets tagged. Every tag uses the same taxonomy. No manual effort.

The Taxonomy

A good classification model uses a fixed taxonomy — a defined list of possible intents. This eliminates the "billing vs payment vs charges" inconsistency problem. There's one label for billing disputes, one for payment method questions, and one for pricing inquiries. Always.

With 315 pre-built intents across 13 categories, the taxonomy covers support questions out of the box. You don't need to define your own labels or train the model on your data.

Setting It Up

Option 1: Replace your tagging system entirely.

Route all incoming messages through classification before they hit your help desk. Each ticket arrives pre-tagged with intent, category, and confidence score. Your agents never manually tag again.

Option 2: Augment your existing system.

Keep your current workflow but add classification as a first pass. The AI adds its tag; agents can override if needed. Over time, you track how often agents override and adjust your routing rules.

Option 3: API-only for analytics.

Don't change your support workflow at all. Run historical tickets through the classification API in batch mode (50% discount for batch processing) to get retroactive tagging. Use the structured data for analytics: what are your top ticket types? Which intents are growing? Where should you invest in automation?

What Good Tagging Gets You

Analytics that actually work. "Bug reports increased 40% after the last deploy." You can say this with confidence because every ticket is tagged consistently. No more guessing.

Routing rules that fire correctly. Tag = billing_dispute? Route to the billing team. Tag = bug_report? Create a GitHub issue. Without consistent tagging, these rules don't work.

Staffing decisions based on data. If 35% of your tickets are technical support and 10% are billing, you know where to allocate people. Without tagging, it's vibes.

Automation opportunities. Once you see that 25% of your tickets are password_reset, you know exactly what to automate first. The data tells you where the biggest wins are.

Accuracy and Edge Cases

Classification hits 92% accuracy out of the box. That means roughly 8 out of 100 tickets get tagged incorrectly or with low confidence.

Low-confidence messages (the system isn't sure) get flagged for human review. This is by design — you'd rather catch a misclassification than let it route wrong. Over time, as you add routing rules for more intents, the effective accuracy of your system improves because even imperfect tags + human review beats manual-only tagging.

Multi-topic messages ("I want a refund AND my account is locked") are the hardest cases. The classifier picks the primary intent (usually the first one mentioned). If multi-topic messages are common for your product, you can set up rules that route to a general queue when confidence is split between intents.

The Migration

If you're switching from manual tagging:

1. Export your existing ticket data with current tags
2. Run a batch classification on the same data
3. Compare: how do your manual tags map to the AI's intent taxonomy?
4. Create a mapping: "billing" → billing_dispute + billing_inquiry + payment_method. This helps you translate historical reporting.
5. Switch to AI tagging for new tickets. Keep your old tags in a custom field for reference.

Total migration effort: 1-2 hours. Ongoing maintenance: zero.