Security & Trust

Avrosh is multi-tenant on PostgreSQL with row-level security enforced at the database layer. Every query runs under the tenant that issued it, so one business can never read another business's rows. Isolation is a property of the database, not just the application.

The QR chat carries no account, no phone number and no login, so a customer can begin without identifying themselves. That is where it stops being true: if they book, the agent asks for a name and a number before it will hold anything, and what they type is kept in the conversation record. This page used to print the first half of that without the second. The privacy policy has the full list of what is kept.

Avrosh runs in two regions today, and your business is assigned to one of them: your application servers and database live there. A European region is available on request, provisioned for the account that asks for it, and it is not listed as running because these two are the machines running now.

• United States, Northern Virginia. For customers in the Americas.
• Singapore, Singapore. For customers across Asia Pacific.

A region decides where your servers and your database live. It does not yet decide where the AI runs, and we would rather you heard that here.

No model certified to serve an account today is hosted in the European Union. EU-served models are in the catalogue but are not certified, so they cannot be assigned to a business. Where each certified model actually runs is in the table below, read from the same registry the router uses. If in-region AI processing is a requirement for you, tell us before you sign rather than after, and we will tell you where certification stands.

Traffic is encrypted in transit with TLS. Data at rest is protected by the hosting platform.

Avrosh answers only from each business's own data, and it escalates to a human when it is unsure. Treating a wrong answer as worse than no answer is a security property: it protects customers and operators from misrepresentation.

The mechanical part is narrower than the slogan, so here is the real scope. Prices, clock times, durations, arrival estimates and calendar dates in a reply have to appear in what a tool actually returned on that turn, or the reply does not go out. That check is real and it runs on every answer.

Names are checked too, but only on the accommodation path, where a room type invented out of nothing is caught. The equivalent check for a salon service, a dental procedure or a menu item is not built yet, and neither is name checking outside English. We would rather you knew which half is enforced by code than take the sentence on faith.

We hold no SOC 2 report, no ISO 27001 certificate, no HIPAA status and no PCI attestation. Competitors print those logos; we do not have them, and a procurement team finds that out in the first call.

What we can give you instead is a written answer to your security questionnaire, the per-model provenance record above, and the reasoning behind every claim on this page. Card data never reaches us at all: subscriptions run through a payments provider acting as merchant of record, so we hold a subscription status rather than a card number.

No Meta pixel, no ad networks and no third-party ad trackers, on this site or in any customer chat.

For a customer who holds an account with a business, export and erasure are built into the product and are a click rather than a support ticket. An export returns their contact record, bookings, waitlist entries and consent history. An erasure strips every identifier in place while leaving the booking rows the business needs for its own records, at every branch where branches are grouped.

Where that tool does not reach: a customer with no account has nothing to erase from, and not every kind of business in the product has customer accounts at all. The tool also strips fields rather than rewriting the conversation, so a name spoken or typed inside a transcript is removed on request instead of by button. Both routes are real and both reach us at the address below.

An operator-side export of your own business content is not self-serve yet. Ask us and we provide a copy. This page said otherwise until we checked, which is the kind of claim a buyer tests.

These are the models that can actually be assigned to a business today. The table is generated from the same configuration the running system picks from, so it changes when the product changes rather than when someone remembers to edit this page.

Two facts are tracked separately on purpose. Served from is where the request is actually answered. Vendor jurisdiction is where that company is incorporated. They are often not the same place, and a page that prints only one of them is not answering the question your counsel is asking. Where we have not verified an origin ourselves, the last column says so instead of leaving a blank.

Avrosh is built to be multi-provider, and these are the vendors we hold relationships with but have not certified. A model is only offered once we have measured it and built an adapter for it, so nothing here touches anyone's data. It is published so that a question like "can you serve my language model from the European Union" gets an honest answer with a reason rather than a flat no.

Beyond the AI layer, a small set of vendors runs the service. These are listed by role; the named vendor for each is confirmed on request at hello@avrosh.com, because publishing a guess would be worse than publishing nothing.

Found a security issue? Write to hello@avrosh.com with the details. We investigate every report and will acknowledge you.