A help desk is a queue with rules: who picks up what, how fast, and what counts as urgent. The slowest part is usually the first touch: reading a vague ticket, working out what it's about, routing it. That's exactly the part an AI does well.
What you're building
- A request form for whoever raises tickets: customers, or employees for an internal IT or HR desk
- Tickets with status, priority, category and assignee
- AI triage on every new ticket: category, priority, suggested assignee
- A drafted first reply from your help docs, which an agent approves or edits
- A queue sorted by priority, plus urgent alerts to Slack
Should you build one or buy one?
| Option | Good for | The catch |
|---|---|---|
| Shared inbox | Very low volume | Nothing is tracked; things get dropped |
| Help desk SaaS (Zendesk, Freshdesk) | Large customer-support teams | Per-agent pricing; AI features often an add-on |
| Build your own in Something | Internal desks, small teams, custom workflows | You define the categories and rules |
Already on Zendesk or Freshdesk? You don't need to switch. Connect it and build just the AI triage layer on top.
Mistakes to avoid
- Letting the AI close tickets. Keep a person on the approve step until the drafts are consistently right for a category.
- Vague priority rules. "Urgent if important" means nothing to a model or a person. Write rules like "blocks a whole team" or "affects payroll".
- Thin help docs. The drafted replies are only as good as the runbooks behind them. When a draft is wrong, fix the doc, not just the reply.





