Zabuno AI

The first evening is the hard one

The work that stops a restaurant from starting is typing. Ninety dishes, their categories and their prices, entered by hand on a phone, after service, by someone who has been on their feet since morning.

The menu already exists. It is printed, laminated and sitting on the table. What is missing is a way to get it out of the paper and into a screen.

A draft, never a publication

Photograph the menu and Zabuno turns it into a draft you can correct. It is a first pass, not a final answer, and the product is built around that distinction.

Nothing generated is written anywhere until a person applies it, and even then it lands in your working draft rather than on a guest's table. A model that misreads a price would otherwise put that price on a menu.

How it works

Photograph the menu Send one photo, up to ten in one go, or queue a stack of pages to be read in the background.
The reply is checked against a schema A reply that does not match the expected shape is rejected at the door. It is never stored and never shown to you.
It waits as a draft The result is filed as an unapplied draft with the model and prompt version that produced it, so you can tell later what read what.
You read it and apply it Applying is a separate, deliberate action by a person with permission to manage the menu, and it can only be done once.
You publish, or you do not Applying writes to your draft menu. Guests still see nothing until you publish.

What it does today

A photo becomes a menu draft: Categories, dish names, prices and currency are read out of a photograph into rows you can correct before applying.
A stack of pages at once: A multi-page menu can be queued and read page by page in the background, with a per-restaurant limit so one large import cannot crowd out everyone else.
A first draft of a description: A dish description is drafted for you to edit. The draft is applied only after you approve it, and it replaces the description alone.
The same dish, typed twice: Products with near-identical names are listed as candidates so you can see the duplicates a long menu hides.
Instructions and menu text are kept apart: Your menu text is sent as data, never as instructions, so a line inside a menu cannot tell the model what to do.
Allergens are never invented: Generated text is forbidden to make allergen claims, and an imported dish arrives with no allergens at all. You declare them yourself.
It says when it is off: Each capability reports whether it is available and why not, so a button that cannot work is not offered as though it could.
One restaurant's drafts stay its own: Every AI record is bound to the workspace that created it, and disappears with it.

What you need

AI has to be switched on	AI ships switched off. Until it is enabled and a provider is configured, every AI action reports that it is unavailable and refuses rather than pretending.
A configured provider	Provider keys are entered once in the encrypted platform vault. They are not read from environment files.
A photograph, not a document	The importer reads image files. A menu that exists only as a PDF has to be photographed or turned into images first.
Permission to manage the menu	Applying a draft is checked against the same permission as editing the menu by hand, because that is what it does.

What it does not do

It does not write your menu There is no dish invention. It reads a menu you already have; it does not propose one you do not.
It does not translate There is no translation assistant. Your menu stays in the language you wrote it in.
It does not touch photos, nutrition or reports No image generation or retouching, no calorie or nutrition estimates, no written summary of your analytics, no campaign copy.
It does not rename products Descriptions are drafted; names are left exactly as you typed them.
It finds duplicates but does not merge them The duplicate list is something to read. Merging two products is still done by hand.
We publish no accuracy figures The provider adapters have been tested against recorded responses, not against a live provider. Any percentage we printed here would be a number we had not measured.
Drafts are kept until you remove them Imported drafts and their records stay until the workspace is deleted. There is no automatic expiry.

Questions owners ask

Can it read my menu from a photo?

Yes. One photo, ten at once, or a queued stack of pages. The result is a draft you correct before applying.

Can it read a PDF?

No. The importer reads images. A PDF menu has to be photographed or converted to images first.

Can it put a wrong price in front of a guest?

Not on its own. Nothing is written until you apply it, and applying writes to your draft. Guests see it only after you publish.

Will it guess the allergens?

No, and that is deliberate. Generated text may not make allergen claims and imported dishes arrive with none. Allergens are yours to declare.

How accurate is it?

We do not publish a figure, because we have not measured one against a live provider. Treat every import as a draft to be checked.

Is it on by default?

No. AI ships switched off, and each capability reports whether it is available and why not.

Start with the menu, not the AI

Building and publishing a menu by hand works on the free plan, with or without AI.

See the plans