This is spot on. For my niche tiny SaaS that gets hundreds of support requests a month, I wanted to build an AI support agent. In other words, I needed to build something to work on my own email client, helpdesk knowledge base, and back office admin interface. I couldn’t afford an Intercom/Front/HelpScout/Zendesk, and those don’t really have fully functional agents anyway. I quickly realized I could build my own help center because I needed an interface for the operator who would approve/edit the AI-written responses and actions taken. So I built my own minimal version of Zendesk and was able to get a functional customer support platform deployed in a few days (thank you Claude Code). The key thing is, now the AI (and the operator) have the information in one place to actually make decisions on incoming support requests. It turns out: the operator interface is key to building a “perfect” AI system, now I can incrementally improve the percentage of support requests that can be handled with AI as I gradually reduce the amount of human intervention required. Having my own operator interface is key here because I will bring in exact contextual information required to make decisions and restrict the types of actions that can be performed by the agent (human or AI). I can’t just handover my business to AI but I can have it gradually take over.
We buried the submission as the URL returns 404. Submitter, please email us at hn@ycombinator.com when you've fixed it and we can restore it.
The link is working for me now.
Henry, could you add some screenshots of mocks to the article?
Absolutely. I’ll need to anonymize what I’m working on before sharing publicly, but I’ll follow up soon with a link and more context. Appreciate the interest!
Where I landed on this for a current project is - don’t build any interface at all and just use Linear - meets most of the needs listed here perfectly, has a well documented API, and they make it easy https://linear.app/agents