Xavi transparency statement (GDPR and AI Act)

What Xavi is, and who can use it

Xavi is the AI assistant in Xavio. You ask it a question about your company in ordinary words — “who owes me money?”, “which products are below minimum stock?” — and it answers from the company data, which it reads on demand through tools. It never receives the whole database: it asks only for what the question needs.

Xavi is a module, switched on separately for each company. Today only these people can use it:

• users who have the superadmin role in a company where the module is switched on — no other company role has the right;
• members of the Xavio team working in the company workspace: the platform administrators — on any company, including from the server terminal, even where the module is not switched on — and colleagues from support or technical, only in a company where they have an approved support session and the module is switched on.

Xavi always works on a single company: the one open on the screen. The model cannot choose or change the company, and every tool filters the data on that company.

Which data Xavi reads

Xavi reads the company data through 23 tools, and changes none of it. Each tool is a query written by us, not one composed by the model; the model only chooses the tool and its parameters — the period, the product, the partner. On the server, these tools read through a database account that has read rights only, so the database refuses any write coming down their path.

A single tool writes: tine_minte (“remember”). It writes only one short line into the memory of the person asking, on the company they are on — see section 5 — and only with that person's “yes”. It does not touch the company data.

• Customers and suppliers
  fisa_partener — the account statement of one customer or supplier for a period: opening balance, documents and payments one by one, closing balance, per currency
  solduri_clienti — how much each customer owes on a chosen date, per currency, with name and tax number
  solduri_furnizori — how much the company owes each supplier on a chosen date, per currency, with name and tax number
  restante_pe_vechime — what customers still have to pay, split by the number of days past the due date
• Invoices
  facturi_emise — the invoices issued in a period: number, date, customer, total, how much was collected and how much remains
  facturi_primite — the invoices received from suppliers, through SPV or from an import: number, date, supplier, tax number, total, how much was paid and how much remains
• Receipts and payments
  trezorerie_pe_perioada — how much came in and went out in a period — receipts on issued invoices and payments to suppliers —, by kind of money
• Sales
  vinzari_pe_luna — the company's sales per month, over a period
  vinzari_pe_client — how much each customer bought in a period
  vinzari_pe_produs — what was sold per product: quantity, value, average price
  comparatie_an_la_an — the sales of a period compared with the same period last year
  marja_pe_produs — sales revenue minus purchase cost, per product
• Stock and goods movements
  stoc_curent — the current stock: quantities and value at purchase price, by category
  stoc_sub_minim — the products below their minimum stock level, and how much to order
  fisa_de_magazie — the receipts and issues of one product over a period, with the document of each movement
  balanta_stoc — the stock balance over a period, for each product that moved
  stoc_fara_miscare — the products that are in stock but have not moved for a long time
  jurnal_miscari — the stock movements of a period: product, quantity, document, warehouse
  intrari_pe_furnizor — goods received, per supplier or per product
  iesiri_pe_client — goods sold, per customer or per product
  top_rulaj — the products that moved the most in a period
  evolutie_cost — how the purchase cost of one product changed, receipt by receipt
• PDF documents
  scoate_pdf — writes a table with the results of the tools above into a PDF on the Xavio server, only when you ask for it. It reads no other data and sends the file nowhere: you download it yourself, from the Xavi page
• Xavi's memory (the only tool that writes)
  tine_minte — writes a short line into your memory, on the company you are on — a preference, for example “wants amounts without VAT”. It is saved directly only when you tell it “remember…”; when Xavi notices a preference on its own, it only asks “Shall I remember this?” and saves nothing until you answer “yes”. It writes nowhere else: not into the company data, not into another person's memory, not into the company's shared memory. It refuses passwords, keys, personal identification numbers (CNP), card data and IBANs

The data read can include the names and tax numbers of customers and suppliers — some of whom may be natural persons —, document numbers and amounts, payments and stock. The tools do not read postal addresses, telephone numbers, email addresses or bank accounts.

What is sent to the model provider

To answer, Xavio sends to Anthropic, through their application programming interface (API):

• the text of the question, with today's date next to it;
• the results of the tools called for that question — rows from the company database, cut at a fixed number of rows per tool. They can contain the names of customers and suppliers, their tax numbers, document numbers and amounts;
• the history of the current conversation: the earlier questions and answers in the same conversation (not the tool results of earlier questions);
• Xavi's fixed instructions and the description of the tools — the same text on every question, with no company data in it;
• Xavi's memory, when there is any: the lines it remembers about the person asking, on that company, and the company's shared lines — at most 30 lines / 2,000 characters for the person and 20 lines / 1,500 characters for the company. Unconfirmed proposals are not sent.

We do not send the name, email address or any other identifier of the person asking, nor the name of the company as such; the company name can appear only if it is in the question or in the data read.

The provider and the model

The language model is claude-opus-5-5, made by Anthropic. For a customer in the EEA, Anthropic's Commercial Terms name the contracting party as Anthropic Ireland, Limited. Anthropic processes the data sent by Xavi as our subprocessor, under its Data Processing Addendum, which is part of those terms.

What Anthropic says in its public documents, read on 5 November 2026:

• Training. By default, Anthropic does not use inputs and outputs from the API to train its models, and its Commercial Terms say it may not train models on customer content. The exception is feedback that a customer sends explicitly; Xavi sends none.
• Retention. API inputs and outputs are deleted automatically within 30 days of receipt or generation. The exceptions, according to Anthropic: when a request is flagged by their automated systems as violating their Usage Policy (then up to 2 years), when the law requires retention, or when agreed otherwise. Xavi does not use any Anthropic service with longer retention, such as the Files API.
• Location. Data is stored in the US. By default, requests may be processed in the US, Europe, Asia and Australia; Xavi does not ask for a particular region today.
• Safeguard for the transfer. The standard contractual clauses of the European Commission (Decision (EU) 2021/914), incorporated into Anthropic's Data Processing Addendum.

Sources:

• Anthropic — Is my data used for model training? (18 August 2026)
• Anthropic — How long do you store my organization's data? (1 July 2026)
• Anthropic — Where are your servers located? (15 June 2026)
• Anthropic — Commercial Terms of Service (effective 17 June 2025)
• Anthropic — Data Processing Addendum (effective 24 February 2025)

What we keep, and for how long

For every question, Xavio keeps a log in its own database, on the server in Germany:

• the conversation: the company, the person who asked, the model used, and when;
• the question and the answer, or the error if there was one;
• every tool called, with what it was asked and its result, cut to at most 4000 characters;
• how many tokens were used, how long it took and what it cost.

The log has no automatic deletion period. Each person can delete their own conversations, from the “Conversations” list: the deletion is real — the questions, the answers and the tools called go with it. What is not deleted stays until the company workspace is deleted, and is deleted with it. When a person's account is deleted, their conversations stay in the company log, no longer linked to the account. Backups containing the log roll off on their own schedule, within six months at the latest.

Xavi's memory. Xavi keeps two kinds of short notes, also in our database:

• about you, separately for each company: preferences such as “wants amounts without VAT”. A line is written only when you say “remember…”, or when you answer “yes” to Xavi's question “Shall I remember this?” — a proposal you do not confirm is never sent to the model. At most 30 lines / 2,000 characters; when the memory is full, Xavi says so and saves nothing more. What it remembers about you on one company does not appear on another;
• about the company, common to everyone who uses Xavi on it — for example “this company is an archive, no more invoices are issued”. Only someone with the right to manage the company settings writes it. At most 20 lines / 1,500 characters.

Everything is visible on the “What Xavi knows about you” page, where you can correct or delete any line about you; the company lines are corrected or deleted by whoever manages the company. The memory has no deletion period; your lines are deleted when you delete them, or together with your account.

What never goes into the memory: passwords, access keys, personal identification numbers (CNP), card data and IBANs. A filter refuses them when written — whether you say them in the conversation or type them on the memory page — and the refused text is not kept in the tool log either.

The PDF files asked of Xavi also stay on the server with no deletion period. Today they are not deleted automatically even when the workspace is deleted; until that is fixed, we delete them on request — write to contact@xavio.ro.

When the retention period changes, this page is updated first.

Who sees the conversations

Each person sees only their own conversations, in the “Conversations” list, and only on the company they are on: they can reopen them, continue them and delete them. A colleague from the same company cannot see, reopen or delete them — not even by knowing their number. The memory about you is likewise seen only by you; the company's shared memory is seen by everyone who uses Xavi on that company.

The full log — questions, answers, the tools called with their results, the cost — can also be read by the Xavio platform administrators, directly in the database, to fix errors and to keep an eye on cost. Nobody else in the company has a screen where they can see it.

Do not type personal data it does not need

Xavi answers questions about the company; it needs no personal data to work. Do not type personal identification numbers (CNP), identity card numbers, IBANs, card numbers, passwords or health information into a question. Everything you type in a question goes to Anthropic and stays in our log.

There is no filter on the question itself: what you type leaves exactly as you typed it. The filter for such data exists only for Xavi's memory — it stops them from being remembered, not from being sent.

What Xavi does not do

• It changes nothing in the ERP: it does not create, change or delete invoices, customers, payments or stock. The only things it writes are its own log, the PDFs you ask it for, and its memory about you — only with your “yes”.
• It sends nothing anywhere other than the request to Anthropic: nothing to ANAF, no emails, no text messages.
• It makes no decisions about people. It does not approve, refuse or rate anyone; it only answers the question asked.
• It does not remember anything about you without your consent. From one conversation to the next it carries over only the memory you confirmed and the company's memory — not the conversations themselves.
• It is not advice — accounting, legal or tax.

Transparency (AI Act, Article 50)

You are talking to an automated system, not to a person. Xavi's answers are generated by an artificial intelligence model and can contain errors — a wrong figure, a missed document, a period understood differently. Do not rely on them without checking them in Xavio, especially before a decision, a tax return or a payment. They do not constitute accounting, legal or tax advice.

Within the meaning of Regulation (EU) 2024/1689 on artificial intelligence, AUTO ȘULEA MOBILE SRL, which develops and operates Xavio, is the provider of the Xavi system, and Anthropic is the provider of the model it relies on. Article 50(1) requires that people be informed that they are interacting with an AI system; this page says so.

Your rights, and the contact

For the company data that Xavi reads, your company is the controller and we are the processor — just as for the rest of the ERP — and Anthropic is our subprocessor. The terms are on the Data processing page, and your rights in the Privacy Policy.

You may ask for access, correction, deletion, restriction or portability of the data, or object to the processing, and you may complain to the supervisory authority — in Romania, ANSPDCP. You can delete your Xavi conversations and the memory about you yourself, from Xavi's page; for anything else — for instance, a copy of a conversation, or the deletion of the company memory — write to contact@xavio.ro from the email address of the account. We never ask for your password.

Changes

When Xavi gets a new tool, another model, another provider or another retention period, this page is updated in the same version of the application, not after. The date at the top of the page shows the current version.