What is voice AI?

Why it matters for enterprise CX leaders

• Voice AI is the first automation category that can plausibly resolve calls end-to-end, not just route or deflect them.
• For enterprise CX leaders, the relevant question is no longer whether voice AI works, but which intents it should handle and what operating model is required to keep it useful.
• UK and ANZ readers will see the same category called "contact centre voice AI".

What voice AI is not

Three adjacent categories get conflated with voice AI in vendor decks. An IVR follows scripted menus — it routes, it does not converse. A voicebot in the 2019 sense mapped speech to intents with a fixed classifier — it could converse within narrow rails but could not act flexibly. Agent assist whispers to a human agent — valuable, but the human still carries the call. Voice AI carries the call itself: it listens, decides, acts on systems of record, and speaks back, in real time.

The stack, in order of what breaks

• Telephony and audio — the least glamorous layer and the source of most production incidents
• Speech-to-text — accuracy on your real audio, not the demo set; accents and line noise move accuracy 10–30 points between providers
• The language model — where intent is understood and responses are planned
• Tool use and integrations — the difference between answering questions and resolving calls; usually the largest source of latency and failure
• Text-to-speech — naturalness, latency, and barge-in behaviour the caller actually feels

What to test before you buy

Evaluate on your own recorded calls, including your worst audio. Test barge-in, background noise, and a caller who changes their mind mid-sentence. Break an integration mid-demo and watch the recovery. Ask for cost per resolved call on your intent mix, not cost per minute. And instrument escalation reasons from day one — that data is what makes the system improve instead of drift.