Zero Data Retention

Zero data retention is a commitment by which an AI model provider guarantees that it does not store, log or retain the information you send it. Your prompt and the model's response are processed in memory just long enough to resolve the request and are then discarded: they do not stay in logs, they are not saved to a database and, above all, they are not used to train future models.

It is one of the core guarantees when a company brings sensitive data (customers, contracts, health, finance) to a language model. Without zero retention, what you write could remain stored in the provider's infrastructure, exposing you to leaks, unauthorized access or unapproved secondary uses. That is why it is part of the security and confidentiality standard Vantegrate applies when implementing AI, described in Security.

In practice, zero data retention takes two forms worth telling apart: a contractual guarantee (what the provider commits in writing not to do) and a technical control (how the architecture actually prevents data from persisting). A serious implementation combines both.

Why it matters for a company

When a company starts using generative AI or AI agents in its processes, prompts stop being innocent questions: they start to contain real business data. A collections agent sees amounts, invoice numbers and customer names; a sales assistant sees opportunities and prices; a legal copilot sees contract clauses. If the model provider keeps all of that, the company loses control over where its most sensitive information lives and is exposed to major regulatory and reputational risk.

Zero data retention cuts that risk at the root: it guarantees that this data does not outlive the request. It does not stay in a provider repository, it cannot be read by the provider's staff and it does not end up, through training, improving a model your competitor uses tomorrow. For regulated industries in Argentina and the region (financial services, healthcare, pharmaceuticals), it is often the condition without which the AI project simply does not get off the ground.

How it works in practice

Zero retention rests on several layers that work together:

• Ephemeral processing: the prompt is loaded into memory, the model generates the response and everything is released. The content is never written to persistent disk.
• No content logging: operational metadata can be recorded (how many requests, latency, errors) but not the text of the prompt or the response.
• Exclusion from training: the data you send is kept out of any dataset used to retrain or fine-tune the model (the opposite of fine-tuning on your data).
• Contractual and compliance guarantee: it is written into the contract or terms of service, ideally backed by independent certifications and audits.

In the Salesforce ecosystem, this principle is exactly what the Einstein Trust Layer provides: a layer that sits between the company's data and the model, guaranteeing zero retention with the LLM provider, in addition to masking sensitive data and keeping an audit trail. It is the difference between calling a public API with your raw data and operating inside a trust perimeter designed for enterprises.

A concrete example

Think of an Argentine lender that wants an agent to answer credit scoring queries. Each query includes the applicant's tax ID (CUIT), declared income and history: data protected by personal data regulations. Without zero retention, that information could remain stored on the model provider's servers, outside the country and outside the company's control. With zero data retention, the query is processed, returns the result and disappears: the data never persisted outside, and the company can show the auditor that its sensitive information was neither stored nor reused.

How it differs from similar concepts

It is easy to confuse zero retention with other privacy guarantees. The table clarifies the boundaries:

The practical takeaway: zero data retention answers the question what happens to my data after the model processes it. A secure architecture combines it with data residency, encryption and access control, rather than using it as the only line of defense.

Common mistakes when evaluating it

• Assuming zero retention is the default: many public AI APIs, on the contrary, retain data for days for abuse/security purposes unless you turn on a specific mode or plan. You have to verify it, not assume it.
• Confusing it with anonymization: zero retention does not anonymize anything; it means the data (anonymized or not) is not stored. They are complementary controls.
• Settling for the contractual promise alone: a clause without a verifiable technical control and, better still, third-party audits is a weak guarantee.
• Ignoring the full chain: if your own integration or an intermediary logs the prompts, the model provider's zero retention will not do you much good. The guarantee has to cover the data's entire journey.