The AI guardrails I use
The rules I've applied to AI in front of customers.
From the chatbot at Workable:
- Roll out in phases, so you can see the impact without disrupting customers or the team.
- Never block the ability to talk to a human.
- Audit the logs every month.
- Use fixed flows for mission-critical moments, like cancellations, purchases, or asking for a person.
- Judge the bot by what happened next: did the customer contact us anyway?
From building Hey Vitae:
- Never invent. It works only with what you've actually done, and it never applies to jobs on your behalf.
- Be honest about gaps. A fit score that hides the gap a hiring manager will notice doesn't help anyone.
- Trace every AI call, so I can see where answers hold up and where they quietly fail, then fix the prompt or the flow.
On this page: the AI is told to answer only from these articles, and any quote that doesn't match the source word for word gets dropped. Questions about salary and references get no AI answer, just a route to me. The AI can draft a reply for me, but only I can send one. Turn on X-ray to watch it happen, or see the test results.