I built a voice bot that answers the phone in Hebrew. Not a demo reel and not a slide - a system that picks up real inbound calls, works out what the caller wants, answers in a natural voice and then does something about it. After dozens of live calls, all recorded and transcribed, I know exactly what works, what breaks and when it pays for itself. Most of what follows never makes it into a sales presentation.
What actually works
It answers, at any hour. A customer calls at nine at night and the phone is picked up on the second ring. For a business that loses its after-hours calls, that alone pays for the project. A call nobody answers is usually a customer who then calls your competitor.
The questions you get a hundred times. Opening hours, prices, order status, how to find the address. Between 40 and 60 percent of the calls into a typical contact center are variations on the same ten questions. The bot closes those outright, which leaves your people for the calls that genuinely need a person.
Warm handoffs. Even when the bot cannot close the case, it reaches the agent with a summary: who is calling, what they want, what has already been said. The human conversation starts in the middle instead of at zero.
Everything on the record. Every call is recorded, transcribed and tagged by speaker. For the first time, management sees what customers actually ask, what they complain about and where the process breaks.
Consistency. A bot never has a bad day. It does not forget to offer the upgrade, does not cut the script short at the end of a shift, does not answer curtly after a hard call. Call number one hundred sounds exactly like call number one, and a change to the script takes effect on every call at once - in a human contact center the same change takes a full round of retraining.
What breaks - the list nobody gives you
Names and numbers in Hebrew. Hebrew speech recognition handles flowing sentences well and falls apart exactly where it costs you: an unusual surname, a phone number rattled off at speed, an address with a house number. What worked for me is having the bot repeat every critical detail back and ask the caller to confirm it before acting. That adds ten seconds to the call and prevents errors that cost real money.
Silence. Transcription engines invent text when the line goes quiet or noisy. Without silence filtering (VAD) you get transcripts containing sentences nobody said. It is a known failure with a known engineering fix, and most vendors leave it alone until it embarrasses them in front of a customer.
Interruptions. People talk over the bot, change their mind mid-sentence, answer with "yes, no, hang on". A bot you cannot interrupt sounds like a phone menu from 2005. Handling interruptions is the difference between a conversation and a machine you want to hang up on.
Response time. Past a second and a half between the end of the customer's sentence and the start of the answer, the call is dead. The customer says "hello?" and the trust is gone. That is an engineering constraint you design around from day one.
Knowing when to hand off. The single biggest mistake is a bot that insists. The moment the caller is frustrated, angry, or simply asks for a representative, you transfer. Immediately. A good bot knows what it does not know.
And privacy. Customer calls are sensitive data, and in some industries they are extremely sensitive. In the systems I build, transcription can run on a local server, so the recordings never leave the organization at all. Worth asking any vendor before you sign: where exactly does my customer's voice travel, and who keeps a copy.
What it saves
International surveys of AI applications in customer service put the saving in contact centers at around 30% of operating costs, with return on investment measured in weeks rather than years - the fastest measured ROI of any application area.
In practice that looks like this: a center of 10 agents where the bot closes 40% of the simple calls performs like a center of 13 to 14 agents, without a single hire. And the bot's own running cost - call minutes, transcription, the language model - comes to a few shekels per call.
Look at the shape of the cost, not only its size. The money goes into setup and tuning, and the ongoing cost barely moves as call volume grows. In a human contact center every rise in demand means hiring. Here it means a few more shekels on the monthly bill.
The part sellers leave out: the bot does not replace the contact center. It lifts the repetitive, grinding layer off it and leaves people the calls that need judgment and empathy. Anyone promising you "zero agents" is selling you a disappointment.
How to start properly
You do not buy "a voice bot for the contact center". You pick one narrow scenario - after-hours calls, say, or appointment scheduling and nothing else - and you set the success metric before you start: share of calls closed without a human, response time, satisfaction. Run it for a month, measure, and only then widen the scope. A pilot like that, from decision to live calls, is a matter of weeks.
And one rule I learned the hard way: every irreversible action - an outbound call, a cancelled appointment, a change to an order - needs the customer's explicit confirmation before it happens. A bot that acts without checking is not a saving, it is a lawsuit waiting to happen.
Want to hear how it sounds in Hebrew, on a real call? Send me a message and I will be glad to show you a live demo.
