When you ask about an invoice, knowing where its information goes is reasonable. We do not train on your data answers part of the question while leaving provider storage open. Tario's integrated AI policy requires model providers with zero data retention. Understanding that condition also means distinguishing it from the accounting records the business needs to keep.
Training and retention are separate questions
A service can exclude conversations from training while retaining them under its terms. Tario's policy addresses both: providers used for model requests must not retain submitted conversations, documents, or data and must not train on them. If no provider meets the required conditions, the request is not sent. The restriction accompanies the model request.
This does not remove Tario's operating responsibilities for company records. Invoices, XML, Hacienda responses, and attachments remain material the business consults. Backups and conversations stored inside the product follow their own retention criteria in the privacy policy. Zero retention at a model provider does not mean deleting accounting support after every answer.
Follow one query
Suppose you ask about a difference between a payment and an invoice. The assistant uses tools available within the company and user permissions. It may need amounts, currency, dates, and relevant references. Its answer should remain connected to evidence. Query context is processed under the AI conditions while the original business document remains available for review.
That does not mean every stored attachment belongs in a model request. The question and tools determine necessary context. An account summary may use totals, while a specific discrepancy needs document detail. Ask a bounded question and inspect returned sources. More information rarely fixes an unclear request.
Reading permissions and approval
Privacy also depends on who can ask. A sales user may need customers and invoices without every financial report. Each entity in an accounting firm's portfolio must retain its scope. Confirm active company and user permissions before a sensitive question. A correct answer about the wrong company is still an access problem.
Preparing and approving a draft are separate steps. A data policy does not authorize issuance, payment, or stock changes. Keep the review required by the action. Check counterparty, date, currency, and effect before confirming a proposal. Confidence in processing conditions cannot replace validation of the transaction.
External assistants introduce another recipient
MCP lets an authorized external client read Tario tools. That client also receives results and applies its own terms. Zero retention required for Tario's integrated AI does not automatically extend to it. Before connecting, inspect the account the team will use, readable information, and how to revoke access.
An export emailed or pasted into another chat does not inherit Tario's policy either. Sharing creates another recipient. When a summarized report is sufficient, avoid unnecessary attachments and individual data. Apply the same rule to screenshots: identify the recipient's actual question and limit the material to that scope.
Questions for a demonstration
- Which policy governs integrated-assistant requests, and where is it published?
- What context does this query need, and what supports its answer?
- Which company information can this user access?
- What does Tario retain as business evidence, and what does the model process?
- Which terms change when an external MCP client is authorized?
Usage analytics has a separate purpose from an accounting query. The public policy explains providers and preferences. Visit statistics, documents sent for AI processing, and the company's fiscal archive are different information paths. Each should be explainable with an example rather than a general statement about security.
Read the policy and test a question using demonstration records. Follow its supporting source and the account's permissions. A useful review identifies what was queried, which company it concerned, and which terms apply. That clarity helps bring AI into accounting work with recognizable responsibility for the information.