What should an AI receptionist do when it cannot answer?

Say so, then offer the best available next step. A warm transfer with context, a scheduled callback, or, as a last resort, a detailed message. Never a guess, never a dead end.

An admin approves every new account by hand. Nothing is created until then. We reply by email; no newsletter, no sequence.

app.salescrew.io/inbox
The unified reply inbox with classified threads

The short answer

  • The first requirement is honesty. The agent needs to recognize and admit it cannot fully help, rather than produce a confident but unverified guess. A caller has no way to tell a real answer from an invented one by listening.
  • A warm transfer to an available human, with context passed along so the caller does not have to repeat themselves, is the strongest fallback when it is available.
  • When no human is immediately available, a scheduled callback with a specific promised window beats an open-ended 'someone will get back to you'. A specific commitment is easier to hold the business to and easier for the caller to plan around.
  • Taking a detailed message is the weakest but still necessary fallback, for when neither transfer nor callback scheduling works. It only helps if the message reaches someone who acts on it.

Why the failure path deserves as much design as the success path

Most evaluation of a voice agent focuses on how well it handles the calls it is built for. Booking an appointment. Answering a common question. Less attention goes to what happens when a call falls outside that design, which is inevitable however well the agent is scoped. A caller with an unusual question, a complaint, or a request the agent was never given information about will happen regularly. How the agent handles that moment often matters more to the caller's overall impression than how well it handled the easy calls.

The wrong response is continuing to answer as if it knows, producing something that sounds plausible but is not grounded in real information. The better response has a clear order of preference. Recognize the limit. Attempt a warm transfer to a human, with context passed along so the caller is not starting over. If no one is available, offer a specific callback window rather than a vague promise. Only fall back to plain message-taking if neither of the stronger options is available.

Fallback options, ranked

OptionCaller experienceWhen to use
Warm transfer with contextFeels like one continuous conversationA human is available now
Scheduled callback with a specific windowFeels like a concrete commitmentNo one is available immediately but can call back
Detailed message takenFeels uncertain, depends on follow-throughLast resort, no transfer or callback path exists
Continued guessingFeels fine until the guess is wrongNever the right choice

What to check in an agent's design

Ask what triggers the agent's recognition that it cannot help. A defined scope boundary, a confidence threshold, an explicit caller request for a human. Then check what happens next in each of the three fallback scenarios: transfer, callback, message. Confirm none of them silently drop the caller with no next step at all.

Disclosure: SalesCrew is our product, and AI receptionist capability is on our roadmap and not shipped today. This page describes the general design standard rather than a SalesCrew feature available now. The fallback hierarchy described here applies to evaluating any vendor's voice agent.

A dead end is worse than an honest limitation

An agent that admits it cannot help and offers a real next step preserves trust. One that guesses, or ends the call with nothing resolved, loses the caller either way, only less honestly.

Questions

Is taking a message ever the best option?
It is the fallback of last resort. It is useful when no human is available and no callback slot exists. It is the weakest option, because a message requires someone to notice it and follow up, unlike a warm transfer or a scheduled callback.
How does the agent know it cannot answer?
Common signals include the topic falling outside its defined scope, its own confidence dropping below a threshold, or the caller saying the response was not what they needed or asking for a person directly.
What is the worst thing it can do in this situation?
Continue guessing as if it has an answer. Transfer the caller with no context, so a human has to start the conversation over. Or simply end the call with no resolution offered at all.