Feature
Teach it your words, review every call
When the phone agent can't find an item or answer a question, the caller's words go on a review list. A person decides what they meant, and the agent uses it from the next call.
Everything here is from our test system. No restaurant uses Zogu yet, and only our team can sign in.
The review list
Two kinds of miss land on the Teach page: an item not found on the menu, and a question it couldn't answer. A test caller on 4 November 2026 asked for "the beef burger", which our test menu doesn't have. The agent named three burgers it does have, and "beef burger" went on the list.
Each row shows the phrase, how often it came up, a link to the last call, the agent's reply and what search makes of it today.
The Phrases to review table on the Teach page: each row has a caller's phrase, how often it came up, a link to the last call, the agent's reply, and controls to map, answer or dismiss it.
- Map it to an item, option or category, and it becomes an alias.
- Answer it, for questions like "Do you sell gift vouchers?".
- Dismiss it, like sushi at a kebab shop.
Aliases and store words
Both are saved for one restaurant only.
- Alias: a whole phrase for one thing. Once "cheeseburger" is linked to the Mega Burger, the next caller who asks gets a Mega Burger.
- Store word: swaps one word inside every item name. With "fries" saved for "chips", "medium fries" adds the Medium Chips. Item names only, for now.
- Built in:
1/4,12",3pcs, "cheesecake" for CHEESE CAKE and "a coke" for the cola need no teaching. See how the agent understands callers.
The Try it box runs any phrase through the same search a call uses, before a caller ever says it.
How dish names are said
The voice reads English spelling and takes no phonetic symbols, so names are respelled in plain syllables. Only the voice changes; the order keeps the menu's spelling.
| On the menu | What the voice is given |
|---|---|
| Cacik | jah-jik |
| Gyro | yee-roh |
| Baklava | bucklava |
Zogu comes with 84 respellings. On 5 November 2026 we checked 26 by machine round trip and dropped ones the voice already said well, like shish and lahmacun. Nobody has checked the rest by ear on a call, so your own respelling always wins. A Check pronunciation list picks out menu words that may need one.
The How to say it section: Written and Say as boxes, the restaurant's saved respellings with who added each, and a Check pronunciation list of menu words with the reason each was picked.
The console has no listen button. To hear it, ring the number and order the dish.
A transcript of every call
The call list shows the newest 200 calls: start time, length, caller number, how it ended, POS status and reply times. See how Zogu takes a phone order. Each call shows:
- every line with its time, the speech recognition's confidence and the reply time
- interruptions, a "yeah" mid-sentence, long silences
- the order exactly as sent to your POS, and its status
- how a call without an order ended: no order, abandoned, ended early (with the reason) or failed
No audio is recorded.
A call page: how the call ended, reply times by step, and the transcript line by line, with a confidence figure on each caller line and a reply time on each agent line.
The confidence figure showed us "and a doner kebab" heard as "And I'm Donica Buck" at 0.69; the agent now asks again (the whole story). Transcripts also show the offers as they happened (see the add-on offers), and problems that aren't a missing word, which we fix in code.
Reply times on test calls
Across 123 replies on 14 test calls by our team on 4 and 5 November 2026:
- median 1.3 seconds from the caller stopping to the reply starting
- two in three within 1.5 seconds; one in twenty over 3.4
- typically 0.47 s to tell the caller had finished, 0.6 s for Zogu and the AI model, 0.18 s for the voice
That's measured on our side, before the phone network. Our targets are a 0.8 second median and 1.5 seconds for the slowest one in twenty, and we're short of both.
Who can sign in
The console sits behind Cloudflare Access with Google sign-in, and checks the sign-in on every request. Only our team has access today. Store words, respellings and dismissed phrases record who made them. A transcript holds the caller's number, words and delivery address, so access stays narrow. See the UK privacy policy and the US privacy policy.
Questions and answers
Do I have to train the agent before it takes orders?
No. It works from your POS menu and already handles shorthand and spellings like hummus and houmous. Teach is for your callers' own words and hard dish names.
Does the agent learn from calls on its own?
No. It writes down what it missed and a person decides. Until then, the next caller gets the same reply.
Is the call audio recorded?
No, only a text transcript of what speech recognition heard, with its confidence.