When you phone your bank or your energy supplier these days, there is a growing chance the voice answering is not a person but an artificial intelligence agent. The interesting question is not whether such systems can understand you, since most can manage that in a quiet demonstration. The hard part, the part that decides whether the experience is bearable or infuriating, is making them work on a real phone call, in a real language, without talking over you or leaving you hanging. A recent decision by a company called Moveo, which builds these voice agents for banks, telecoms firms, utilities and insurers, offers an unusually clear window into that hidden difficulty.
The gap between the demo and the phone call
The core of any voice agent is speech-to-text, the software that turns the sounds you make into words the machine can act on. It is easy to make this look impressive. Feed a system a clean recording in American English and it will transcribe it almost flawlessly.
The trouble is that real calls are nothing like that clean recording.
A live phone call arrives compressed and degraded, carried over the ageing infrastructure of the telephone network, full of the hiss, clipping and dropouts that engineers politely call artifacts.
The caller might be speaking Brazilian Portuguese or Greek, languages that the big speech systems, tuned mostly for English, handle far less reliably.
And if the company is a bank or an insurer, the whole exchange has to satisfy strict rules about where data is processed and how it is protected.
A system that aces the demo can fall apart the moment it meets these conditions, which is why Moveo set out to judge potential suppliers on how they behave in production rather than how they score on a leaderboard.
Measuring the pauses you feel but cannot name
What Moveo really cares about is timing, because timing is what makes a conversation feel human or robotic.
Think about the small, almost unconscious rhythms of a phone call, the tiny gap after you stop speaking before the other person replies.
If that gap is too long, the silence feels dead and you start wondering if the line has dropped.
If it is too short, the system barges in and cuts you off mid-sentence, which is worse.
To pin this down, Moveo measures three separate delays on every call: how long the software takes to turn speech into text, how quickly it works out that you have actually finished talking rather than just pausing, and how long the agent then takes to respond.
It records these delays for every call and looks not just at the typical case but at the bad cases, the slowest one in twenty, because it is the occasional long, awkward silence that ruins the experience and drives callers to demand a human.
Trying before buying, in three stages
Rather than take any supplier's word for it, Moveo runs candidates through a widening series of tests that move progressively closer to the real thing.
First it feeds them pre-recorded and artificial audio in a simple online tool, just to weed out the obvious failures.
Next it builds working prototype agents and runs them over both internet and genuine telephone connections, logging how they cope with the messiness of an actual call.
Finally it plugs the survivors into its full system and subjects them to conditions resembling live traffic, watching for errors and instability under load.
Only a supplier that holds up through all three stages is worth deploying to real customers.
What Moveo actually wanted
The qualities Moveo was hunting for explain what "production-ready" really means in this business.
It wanted suppliers whose systems process audio continuously as it streams in, rather than waiting for the caller to finish, so replies come faster.
It wanted genuinely strong support for less common languages, built-in noise cleaning to cope with bad lines, and the option to process data on servers inside the European Union to meet the region's tough privacy rules.
The company it settled on as a core provider, Deepgram, met that bar, offering the streaming approach, coverage of Brazilian Portuguese and Greek, European hosting and noise reduction.
Moveo says the result was measurably faster responses and, importantly, more calls resolved by the agent without being handed off to a person.
The wider point
None of this is visible to the person on the line, and that is rather the point.
The difference between a voice agent you tolerate and one you fight to escape comes down to engineering choices most callers will never see or name.
The next time an automated voice answers your call and, against expectation, actually handles the whole thing without cutting you off or leaving you in silence, it is worth remembering that a surprising amount of invisible work went into making those few hundred milliseconds feel effortless.
