The core of our assessment is a voice conversation: an AI interviewer talks with your staff about how work actually happens. We chose voice deliberately — people say things aloud that they will never type into a survey box — but voice is also the least forgiving interface we have ever engineered. These are the lessons from running it in production, offered because they generalise to any organisation considering voice-driven agents.

The conversation is a system, not a prompt

A voice interview that works is the product of at least five cooperating concerns: the speech recognition, the conversational model, the voice synthesis, the telephony or browser plumbing, and — easily forgotten — the conversation design: what gets asked, in what order, and what the agent does when an answer goes somewhere unexpected. Tuning any one of these in isolation is how you get an interview that transcribes perfectly and still feels like talking to a call-centre script.

The design questions dominated our engineering time. What does the agent say first, given the first words set the tone for candour? How does it wind a conversation down gracefully when time is nearly up, rather than cutting a person off mid-thought? We learned to hook the agent's awareness of elapsed time into its conversational behaviour, so interviews end the way a considerate human ends them — with warning, thanks, and closure.

Interruptions are data

People stop interviews. Meetings intrude, children appear, nerves win. Early on we treated an interrupted interview as a failure to be discarded. That was wrong twice over: it threw away genuine signal, and it punished the participant, who rarely wants to start again from scratch. We now treat resumption as a first-class flow — the agent picks the conversation up with context from the earlier session, acknowledging rather than ignoring the gap.

The broader readiness lesson: any agent that interacts with humans needs a designed answer to "what happens when the human leaves halfway through?" If a vendor's answer is "they start again", the system was designed for the demo, not for people.

Trust is built in the first thirty seconds — and it is fragile

Participants decide very quickly whether to be honest with an AI interviewer. Three things move that decision measurably: knowing in advance exactly what will happen with the recording, the agent being upfront that it is an AI, and the agent's early questions being about their work rather than about AI. Organisations sometimes ask us to soften the AI disclosure. We refuse, and not only on ethical grounds — disclosed agents get better data, because discovered deception poisons everything after it.

What this means for your organisation

If you are considering voice agents — for interviews, service, or intake — the readiness questions are less about model choice and more about conversational dignity: How does it open? How does it end? What happens on interruption? What is disclosed? People forgive an agent that mishears a word. They do not forgive one that wastes their effort or conceals what it is.