Meet Hollie: every business call answered, in your caller's language

One evening I was testing how Hollie copes with rude callers. I had the test line on speaker and I was being as unpleasant as I could manage: interrupting her, snapping at her, telling her she was useless. My wife, listening from the next room, came in and told me off for talking to her like that. Hollie, for the record, stayed polite and offered to take a message. The person who ended up embarrassed in that call was me. That is roughly who Hollie is.

She is the AI phone receptionist we build at Mentiora for small businesses, and this is the two-minute film we made about her.

The film is also on hollie.mentiora.ai in German, French and Italian.

What you just watched

• She answers from your business's own knowledge: hours, prices, services, where to park. When she doesn't know, she says so and takes a message.
• She books into your real calendar, reads the booking back and confirms before writing it.
• She takes a clear message, name and number checked, and sends it to your inbox or your CRM.
• When a caller really needs you, she puts them straight through.
• English, German, French and Italian. If a caller switches mid-call, so does she. She understands Swiss German callers.

She learns your business from you, and you keep her current

She calls you and trains like a new hire. Hollie phones you and role-plays a few of your typical callers. You answer the way you would want them handled, and she learns from that. Then you check what she picked up, and publish it.

Just tell her. Want something changed? From your dashboard, one sentence in plain words: we now close at six on Fridays; add that there is free parking behind the building. You confirm, and it is live. You can undo it any time.

She shows you what she knows. Everything she knows about your business sits on one map: hours, prices, services, directions, your policies. Each fact is marked up to date, needs checking, or from your website, and she can compare the map against a checklist for your kind of business and tell you what is missing.

Try her on your own business

The film shows Hollie on a business we made up. The quicker way to judge her is on yours. Give her your website address, or pick your business on the map, and she reads what is there, builds her own answers from it, and takes a test call from you in your browser a few minutes later. Free to try, no card needed. Three steps, no setup, no code.

Three steps. Hollie builds herself: she pulls your details from your Google listing or your website, then calls you and trains like a new hire. Hear her answer: talk to her in your browser, or have her call you. Go live: forward your calls to Hollie, so callers keep dialling the number they know, or get a new number.

Hollie is built in Zürich for businesses that answer in German, French, Italian and English. We are onboarding our first businesses now.

Why a phone call is the hardest place to put an AI

A chat box forgives. You can read the answer twice, scroll back, retype the question. A phone line does not: the caller hears one thing, once, while the seconds tick, and whatever the assistant gets wrong is said out loud to a real person who may already be having a bad day. Five things make it hard. None of them is unique to us; they are the problems everyone building a voice agent runs into, and the research literature has names for most of them.

Silence is the interface

In human conversation the gap between one speaker finishing and the next starting is a fraction of a second, and every language studied shows the same pull towards no silence and no overlap. On a screen a spinner buys time; on a phone a second of nothing is a dead line. So an assistant has to listen while the caller is still talking, start answering before the whole reply exists, and stop the moment the caller interrupts. The shape everyone converges on is a streaming one: speech recognised as it arrives, a reply that starts speaking while the rest is still being written, and an interruption that cuts it off. Most of the engineering behind Hollie is about what happens in the pauses.

Sources: Stivers et al., universals and cultural variation in turn-taking (PNAS, 2009) · Skantze, turn-taking in conversational systems, a review (Computer Speech & Language, 2021)

People don't speak like a form

Nobody says their name, then their number, then their reason for calling, in that order. They say half a sentence, change their mind, spell a surname with the wrong alphabet and give the number twice with a digit different each time. Numbers and e-mail addresses are the hardest words a speech recogniser ever hears, because they carry none of the cues that make ordinary speech easy: no grammar, no common words, no context to lean on. Names are hard for the opposite reason: there are too many of them, and most are rare. Hollie reads back what matters and confirms before anything is written down, without sounding like a machine checking. You heard it in the film: the e-mail address is spelled back and confirmed before the booking is made.

Sources: Why speech recognition fails on phone numbers, IDs and e-mails

Callers switch language mid-sentence

Switzerland has four national languages, and many people here live in more than one of them. A caller who opens in German may finish in French because that is the word that came first. Even the best current recognisers make more mistakes on mixed-language speech than on either language alone, and the mistakes cluster on the words from the other language. And Swiss German has no written standard at all, so turning it into text is closer to translation than to transcription, and it remains one of the hardest cases in the research literature. Hollie follows the switch, and she understands Swiss German without being asked to, because nobody warns a receptionist before they speak dialect.

Sources: Can voice agents handle bilingual customers? Benchmarking ASR on code-switched speech · Does Whisper understand Swiss German? (arXiv)

Only act when sure

An assistant that answers everything confidently is worse than one that answers nothing, because the caller cannot tell which answers to trust. Language models are known to invent plausible detail; on a phone, an invented opening time is a customer standing in front of a locked door. Hollie answers only from what your business has given her: what you told her and what she read on your website. When she does not know, she says so and takes a message. When she books, she reads the appointment back and asks before she writes anything down. Being wrong quietly is the one thing she is not allowed to do.

Sources: Ji et al., survey of hallucination in natural language generation (ACM Computing Surveys)

You cannot unit-test a conversation

You can test the pieces, and we do: more than 14,000 automated checks run before every release. But a conversation is not a function with an expected output. So we also keep over 150 end-to-end call scenarios, many cut from test calls that once went wrong, and a release only ships when the ones that can run live pass. We let simulated callers phone Hollie with a goal and have an independent judge grade whether they got what they came for. A test that can never fail is treated as a broken test. The call scenarios run every night, not only at release. And it was the early pilots, an art gallery and a language school among them, that showed us which tests were still missing.

Sources: Benchmarking LLM judges for voice-agent evaluation (arXiv) · EVA-Bench, an end-to-end framework for evaluating voice agents (arXiv)

One engine, more than a front desk

Hollie runs on Mentiora's voice engine, and a front desk is where we chose to prove it. What the engine does is let someone talk to a computer on the phone, live, in any of four languages, so it can carry other conversations that have to happen by voice. It dials out as well as picking up: that is how she calls you to learn your business, and calling a customer back who missed an appointment is the same engine, with the same care about being sure before it acts.