It refuses anything not on your published site, anything a code-level safety gate blocks, and anything that calls for judgment instead of a fact.
The assistant answers strictly from the business's own published content: services, pricing, hours, policies, booking pages. If a question falls outside that content, or the assistant can't verify the answer against it, it says "I don't know" and offers a person instead of guessing. It never invents a price, a policy, or an answer that isn't actually on your site.
Some refusals happen before a language model even sees the message. A deterministic crisis gate runs in code first, catching emergencies and routing them appropriately, rather than leaving that judgment call to a model's interpretation of a sentence. On top of that, safety guardrails are tuned per field. A med spa's assistant carries a no-medical-advice line with a handoff to a clinician. A gym's assistant declines an injury question under its standing safety rules, offering to pass the visitor's contact to staff only with their permission.
Beyond the built-in gates, the business owner decides what sources the assistant can draw from and what boundaries it holds. Those boundaries can be changed from the dashboard at any time. If there's a topic you don't want the assistant touching at all, that's a setting the owner controls, not something baked in and unchangeable.
Every answer passes through a second model that checks it against the business's published content before it ships to the visitor. If that check fails, the answer doesn't go out. This is the same mechanism behind the "I don't know" behavior: refusal isn't a mood, it's what happens when verification comes back empty.
A visitor to a med spa's site asks: "What dose of filler should I get for my lips?"
The assistant declines to answer the medical question directly and hands off to a clinician: "That's a question for one of our providers to answer in person. Source: Services page (filler overview). Would you like me to have someone reach out?" If the visitor says yes, their name and contact go to the owner's inbox with their permission.
No. If it can't verify an answer against the business's published content, it says "I don't know" and offers a person. That's true whether the gap is a missing policy page or a question outside what the business has published anywhere.
Yes. The owner sets the sources, boundaries, and routing from the dashboard, and can change any of it at any time. Refusal boundaries aren't fixed once at setup.
No. Safety guardrails are tuned per field. A med spa's assistant handles medical questions differently than a gym's assistant handles an injury report, because the risks for each business aren't the same, and the assistant is built around that difference rather than one generic rule for everyone. Try it free: a private preview, built in under 24 hours, no card required, at askmy.business.
Give us your website and we build your private assistant — free, usually within a day, no card. Nothing goes live without your approval.
Build my free previewUpdated 2026-09-01 · AskMyBusiness by Vision Genesis · Knoxville, TN