AI, explained

What is an AI voice agent?

A voice agent is three systems in a trench coat, and the one that decides whether customers hate it is the one nobody talks about.

Strip away the marketing and there is a loop. Speech to text converts the caller's audio into words. A language model, given your instructions and access to your tools, decides what to say or do. Text to speech turns the reply into audio. Around all of that sits turn detection, which works out when the caller has finished speaking and when the agent should stop talking because it has been interrupted.

Why latency is the whole product

Humans notice conversational delay at a few hundred milliseconds. Every stage adds some, and a lookup in your CRM in the middle adds more. This is why voice agents that sound superb in a demo feel wrong on a real call: the demo had nothing to look up. Practical mitigations are architectural. Stream partial transcripts rather than waiting for a full sentence, start generating before the caller finishes, use a filler acknowledgement while a slow tool call runs, and cache the data you know you will need before the call connects.

Where voice genuinely beats chat

• The customer's hands are busy, they are driving, on site, or in a workshop.
• The task is urgent and emotional, a cancellation, a fault, a missed delivery.
• Your customer base does not use your app and never will, which describes a large share of Portuguese trade and hospitality clientele.
• You are already paying for phone coverage out of hours and getting voicemail instead.

Where it does not

Anything long, detailed or requiring a document is better in text. Complex account changes, disputes, and anything with a legal edge belong with a person. Voice also suffers where accents, background noise and code-switching are common, which is a real constraint in Portugal where callers move between Portuguese and English mid-sentence.

The handoff is the design

The single most important decision is when the agent gives up. Good agents transfer early and cleanly: they detect frustration or repeated failure, tell the caller a human is coming, pass a written summary of the conversation to the agent picking up, and never make the customer repeat themselves. An agent that fights to keep a call it cannot resolve is worse than no agent at all.

What it actually costs to run

Charge structures usually combine per-minute telephony, speech processing, model tokens and text to speech. It lands well below a staffed phone line for high-volume repetitive calls, and above it for low-volume complex ones. Model the mix before committing.

If you want a concrete view of whether your call profile suits voice, an AI audit maps call types against resolution rates first, before any build.

Frequently asked questions

Related