The short answer: an AI attendant can genuinely earn its place — answering questions, booking appointments, routing inquiries around the clock — if it's built on three non-negotiables: it's grounded (speaks only from your approved information), guarded (escalates to a human instead of guessing), and honestly labeled (never pretends to be a person). Skip any of the three and you've installed a liability that speaks for your brand.
What it can actually do (done well)
- Answer from your real information — hours, policies, services, pricing rules, order status — instantly, at 2 a.m., in a customer's own words.
- Take actions — book and reschedule appointments, start a return, open a support ticket, collect the details a human will need.
- Qualify and route — figure out what the visitor needs and hand off to the right person with full context, instead of a cold "someone will contact you."
- Absorb the repetitive 80% so your team's hours go to the conversations that need judgment. (If most of your repetition is status/documents/rescheduling, compare a customer portal — sometimes self-service beats conversation.)
What it can't do — and the failure everyone's seen
An ungrounded assistant will answer confidently whether or not it knows — inventing return policies, promising discounts, misquoting your own prices. Those aren't rare quirks; they're the default behavior of AI without grounding. There have been widely-reported cases of companies held to commitments their chatbot invented. The engineering answer is strict: the assistant may only speak from your approved sources, must say "I don't know — let me connect you" past its limits, and every action it can take is explicitly designed, not improvised.
The three build paths
Fast, cheap, fine for simple Q&A
Upload your documents, embed the widget, get generic-but-useful answers. Right when the job is answering common questions. Limits: generic behavior and voice, weak or no ability to take actions, and grounding quality varies — test it adversarially before trusting it.
The middle path
Assistant platforms that connect to your booking/support/commerce tools with configuration rather than code. More capable than a widget; still shaped by the vendor's model of your business.
When the attendant is part of the product
Built on your live systems, in your voice, with designed escalation, permissioned actions, and — the part usually skipped — an evaluation harness: a test suite of real customer questions (including hostile ones) the assistant must pass before and after every change. This is AI systems engineering, and it's the difference between an attendant and a hazard.
The honesty rules (non-negotiable, and good business)
- Label it. "AI assistant" in the header, always. Customers don't mind AI; they mind being tricked — and disclosure requirements are spreading.
- Ground it. Approved sources only; "I don't know" is a feature.
- Escalate visibly. A human path that's always one message away, with the conversation's context handed over.
- Log and review. Read transcripts weekly at first — they're a goldmine of what customers actually ask, and your early-warning system.
- Test before trust. Try to make it invent a policy before your customers do.
Common questions
What can an AI assistant on my website actually do?
Grounded and guarded: answer from your real information, book and reschedule, qualify and route, hand off to humans with context — around the clock. Ungrounded: invent policies. The build discipline is the whole game.
Should it pretend to be human?
No. Label it clearly — customers accept AI and resent deception, and disclosure requirements are spreading. Honest labeling is free trust.
Widget or custom build?
Widget for simple Q&A from documents. Custom when the assistant must take actions on your live systems and speak in your voice — that's engineering, with an evaluation harness before launch.
Want one built grounded, guarded, and tested?
AI systems with evaluation harnesses are exactly what the Studio builds — assistants that survive contact with real customers.
AI strategy & execution Which AI assistant for you?