AI Safety at CallPilot

This is the part most AI receptionists do not publish. We think it is the most important thing about the product.

What the AI is forbidden to do

• Booking confirmations — CallPilot captures requests, humans confirm them
• Prices, policies, or hours that are not in your approved knowledge
• Medical, legal, or financial advice
• Pretending to be a human — automation is always disclosed
• Answers when no approved knowledge covers the question

Booking requests are not confirmations

When a visitor asks for an appointment, CallPilot captures the request — service, requested time, contact details — and records it as requested. A human confirms it. The AI never marks a booking confirmed, because it cannot verify availability the way your front desk can.

How uncertainty escalates

If a question is not covered by approved knowledge, or the visitor's intent is ambiguous, the conversation is escalated. Your team receives the full exchange and the reason, rather than discovering a confident wrong answer later.

Knowledge approval

Nothing the receptionist says comes from the open web. It comes from articles and documents a business has reviewed. Drafts stay drafts until approved, and conflicts between an article and configured hours are surfaced rather than guessed at.

Automation is disclosed

Visitors are told they are interacting with an automated receptionist. We do not aim for indistinguishability from a human — we aim for usefulness with honesty.

Test before you publish

Every workspace includes a simulator so a business can test its receptionist against real questions before the widget is installed publicly. The resulting conversation lands in the inbox for review.