Architecture
How it works
Written for whoever on your side will ask the hard questions. This is the actual architecture, including the parts that are limits rather than wins.
The pipeline
What happens during a conversation
- 1
The visitor opens the Concierge
One script tag on your page, locked to your domain. The visitor can talk or type. The audio is carried over a real-time media server that we run, built for the network problems that make voice hard: packet loss, jitter, and visitors on bad mobile connections.
- 2
Speech becomes text, while they are still talking
Transcription is streaming. The Concierge starts forming a response before the visitor has finished the sentence, which is most of the difference between a natural reply and an awkward pause.
- 3
It decides, and looks things up
A language model works out what the visitor wants. General knowledge of your field comes from the model; anything particular to your business comes from your systems, queried mid-conversation: the catalogue, opening hours, appointment slots, and the knowledge base we load and you approve.
- 4
It replies, shows and acts
The response is spoken in the voice chosen for that language and written at the same time. Where the conversation calls for it, the Concierge shows product cards and tappable choices, opens the product page, books the slot, and posts the lead to your CRM.
Languages
Speaks their language, switches when they do
Five languages, one Concierge
English, Arabic, Russian, German and Dutch, each with its own voice and its own translated prompt. It opens in your site's language and moves the moment the visitor does. It is not a translation layer.
Tested in every one of them
A prompt that scores well in English can fail translated. Every change to the Concierge's wording is measured in each language it speaks before it goes live.
Trust
Trusted with your brand
Stays on subject
It answers about your business and its field, and politely declines the rest, in whatever language the conversation is in. A clinic's concierge books consultations; it does not give medical advice.
Says what it is
It has a name your customers will use. Asked whether it is a person, it answers plainly that it is an AI.
Confirms before it acts
No booking or lead is written without a spoken yes. Every capture is read back first.
Books only real slots
A closed location or a past hour is refused with a reason it can read aloud, so it cannot confirm a Tuesday at a showroom that shuts on Tuesdays.
Locked to your site
Only pages on your domain can start a conversation. No one can embed your Concierge elsewhere.
Reads your systems, writes one endpoint
Bookings and leads go through a single endpoint you control. Nothing on the public side can change your configuration.
Your data
Your conversations stay yours
Isolated per client
Every conversation, text and voice, is stored in an isolated environment in our database and cloud storage. Your conversations are never mixed with another client's.
Kept as long as you want
Nothing is deleted automatically. Your conversations stay for as long as you want them, and they are deleted when you ask.
Or in your own storage
On request, your conversations are stored in your own database and cloud storage instead. The Concierge works the same.
Limits
On limits, honestly
The Concierge answers from your data and from general knowledge of your field. When your catalogue has no answer, it says so and takes a number rather than guessing. That is a limit as well as a rule: a question your systems cannot answer becomes a callback, not an answer.
It books only what it can check. A slot is confirmed against real opening hours, and where a business publishes none, the Concierge registers interest for an advisor to call rather than promising a time it has no right to.
It is good at the same handful of questions your staff answer every day, and bad at long, unusual or emotionally difficult conversations. We will tell you which of yours are which before you spend anything.
After the conversation
You can see what happened
Every conversation on the record
Full transcript, a written summary of what the visitor wanted and what happened, and the audio, in the Intentport app or in your CRM.
The lead it produced
A booked visit, a callback or a registered interest, posted the moment the visitor confirms, with the conversation behind it.
Costed per conversation
Each conversation is costed by component (transcription, language model, speech) against a rate table. You can see what a conversation cost, not just what the monthly bill was.
Latency you can inspect
Turn-by-turn timings, so when a conversation felt slow there is a number showing where the time went.