What does zero data retention (ZDR) mean?

Zero data retention means an AI provider has committed, contractually and technically, not to store, log, or otherwise retain a customer's prompts and the model's outputs beyond the processing needed to generate that specific response. Once the response is returned, the data is not kept in the provider's systems for later use, whether for debugging, abuse review, or model training. It is typically offered as a negotiated enterprise capability rather than a default setting, it applies to specific named products and access methods rather than a vendor's entire platform, and credible versions of it still name a small set of exceptions, such as a short-lived safety classification result or data kept briefly to meet a legal obligation. The point of ZDR is to remove the standing risk that comes with stored data existing anywhere outside the customer's control.

How is zero data retention different from a standard AI API's terms of service?

Under standard commercial terms, most major providers already exclude API inputs and outputs from model training by default, but that is a separate commitment from retention. Providers commonly keep API requests and responses for a limited window, often described as roughly 30 days, specifically to investigate abuse, enforce usage policies, and support debugging. That window applies automatically and is not something a customer can disable from a self-serve account. A zero data retention agreement replaces that default window with a documented commitment that data is not persisted at all beyond the immediate request, and it typically requires an enterprise agreement, provider approval, and a scope limited to the products and endpoints named in the contract.

Why do enterprises require zero data retention agreements?

The reasons cluster around four concerns. Competitive confidentiality: prompts can carry product roadmaps, source code, or deal terms that an organization does not want existing anywhere outside its own systems. Regulated data: healthcare workflows touching protected health information, financial workflows touching material non-public information, and legal workflows touching privileged material each carry handling obligations that un-bounded third-party retention puts at risk, and in healthcare a business-associate relationship under HIPAA often makes ZDR a condition of the deal. Data sovereignty: government and defense-adjacent organizations frequently treat any indefinite retention by a commercial vendor as unacceptable regardless of the vendor's intentions. And training-data leakage: even a provider that promises not to train on customer data still holds a standing risk while that data is stored, since a policy can change, an incident can expose a data store, or a future product can be built on historical logs. Zero retention removes the data before any of those risks can materialize.

Is zero data retention the same as a vendor not training on my data?

No, and treating them as the same thing is a common and consequential mistake. Not training on customer data is a commitment about how stored or in-flight data is used. Zero data retention is a commitment about whether the data is stored at all. A vendor can offer one without the other: a standard commercial plan frequently excludes customer data from training by default while still retaining it for a period for abuse monitoring and debugging. When evaluating a vendor, confirm both commitments explicitly and separately, and read the specific language in the data processing agreement rather than assuming a training exclusion implies zero retention or the reverse.

Is ZDR available on every AI plan, or only enterprise tiers?

In the well-established pattern across major LLM API providers, zero data retention is not a self-serve toggle available to every account. It is typically offered only to enterprise customers through a negotiated agreement, subject to the provider's approval, and scoped to specific eligible products or endpoints named in that agreement rather than the vendor's entire platform. An organization using a personal or standard commercial account should not assume ZDR applies just because the vendor offers it somewhere in its enterprise lineup; the commitment has to be confirmed for the specific access method actually in use.

What should we verify before trusting a vendor's zero data retention claim?

Verify six things rather than accepting a trust-center statement at face value. First, that the commitment is written into the data processing agreement or order form, not only described in marketing material. Second, exactly which products, endpoints, and access method it covers, since a commitment for one surface does not extend automatically to another. Third, that it covers both prompts and completions, not just one direction. Fourth, the specific, narrow exceptions the vendor names, such as short-lived safety classification results or legally mandated retention, since a credible claim names its exceptions rather than asserting an absolute zero. Fifth, whether any sub-processor or underlying model provider in the data path is bound by the same commitment, not just the vendor's own front end. Sixth, that retention and training-data use are confirmed as two separate, explicit commitments rather than assumed to travel together.

How does zero data retention fit into a broader AI governance program?

Zero data retention answers one question, what happens to the data, inside a much larger vendor data-handling assessment that also covers encryption, identity and access controls, network isolation, audit logging, and applicable compliance attestations. Deciding which AI uses require a ZDR agreement, tracking which vendors and products actually carry one, and re-checking the commitment as agreements and vendor products change are exactly the kind of ongoing controls an AI governance framework assigns an owner for. A stated ZDR commitment is also a claim an AI audit can test: whether a signed agreement exists, whether it covers the products actually in use, and whether any exception has crept in since it was signed.

Does Lightbridge Automation hold a zero data retention agreement with an AI vendor on our behalf?

No. Lightbridge Automation is an independent AI-governance advisor. This page is buyer education: it explains what zero data retention is, why organizations require it, and what to verify, so a team can evaluate its own vendor agreements with a clear checklist. Lightbridge Automation does not hold, broker, or represent a specific vendor's zero data retention agreement on a client's behalf, and any such agreement is a contract between the organization and its AI provider. Where Lightbridge Automation helps is in scoping which of your AI uses need a ZDR agreement, reviewing the contractual language against what your program actually requires, and folding that evaluation into a broader AI governance framework and AI audit practice.

RL Written by Robert LabardeeFounder and CEO

Last updated August 2026.

Zero data retention (ZDR) for enterprise AI, explained

Zero data retention (ZDR) is a contractual and technical commitment from an AI provider that it will not store, log, or use an enterprise customer's prompts and outputs beyond what is strictly needed to serve the immediate request. It typically requires a negotiated enterprise agreement, and it is distinct from a separate commitment not to train on customer data.

Evaluate your AI vendor terms See Claude's data-handling paths

Zero data retention replaces a default retention window with a contractual commitment.

Every AI API has a data-handling default, and it is easy to assume that default is already "safe" once a vendor confirms it does not train on customer data. Retention is the separate question, and it is the one ZDR addresses.

Standard commercial API terms

On a self-serve or pay-as-you-go plan, most major providers already exclude API inputs and outputs from model training by default. Retention is the separate question: providers commonly keep API requests and responses for a limited window, often described as roughly 30 days, so they can investigate abuse, enforce usage policies, and debug incidents. That window is a default, set by policy, not something a customer can turn off from a dashboard.

A zero data retention agreement

ZDR replaces that default retention window with a commitment, documented in a data processing agreement or order form, that prompts and completions are processed only long enough to generate the response and are not persisted afterward. It is typically gated behind enterprise sales, requires the provider's approval, and applies only to the specific products, endpoints, and access method named in the agreement, not to every way an organization might touch the vendor's AI.

For how this plays out across the three commercial ways to consume Claude specifically, see the Lightbridge Automation Claude security and compliance guide, which maps retention defaults and contract instruments to SOC 2, ISO 27001, SOX, and GDPR.

Enterprises require zero data retention for four recurring reasons.

The demand for ZDR is not generic caution. It concentrates wherever the data in a prompt is worth protecting beyond the life of that single request.

Competitive and trade-secret confidentiality

Prompts routed through an AI API can contain product roadmaps, pricing strategy, source code, or deal terms. An organization that cannot rule out retention has to assume that data exists somewhere outside its control indefinitely, which is a risk many will not accept for their most sensitive work.

Regulated data: health, financial, and legal

A healthcare workflow that touches protected health information, a finance workflow that touches material non-public information, or a legal workflow that touches privileged material each carries its own handling obligations. In healthcare, a business-associate relationship under HIPAA is a common reason ZDR becomes a condition of the deal rather than a nice-to-have.

Government and defense-adjacent data

Public-sector and defense-adjacent organizations often operate under data-sovereignty and access-control requirements that treat any un-bounded retention by a commercial vendor as a non-starter, independent of whether the data would ever actually be misused.

Avoiding a training-data leakage path

Even where a provider states it will not train on customer data, retained data is a standing risk: a policy can change, an incident can expose a data store, or a future feature can be built on historical logs. Zero retention removes the data before any of those paths can open, rather than relying only on a promise about how stored data will be used.

Verify a vendor's ZDR claim on six points before you rely on it.

A zero data retention claim is only as strong as the contract behind it. A trust-center page is a starting point for the conversation, not the evidence itself.

Contractual, not just marketing

A blog post or trust-center page describing zero retention is not the commitment. Confirm the term is written into the data processing agreement or order form, with the scope, the deletion mechanism, and any exceptions stated explicitly, not implied by a product page.

Which tier it actually applies to

ZDR is usually an enterprise-only, negotiated capability, not a setting available on a standard or self-serve plan. Confirm which specific products, endpoints, and access method (an API key, a particular admin console, a specific deployment) the agreement covers, because a commitment for one surface does not automatically extend to another the organization also uses.

Prompts and completions, both directions

Check that the commitment covers both the request sent to the model and the response it returns, not just one side. Some retention policies distinguish inputs from outputs, or treat them differently for logging versus training purposes.

Named, narrow exceptions

Most zero-retention agreements still carve out a residual set of exceptions, for example a short-lived safety or abuse-monitoring classification result, or data retained to meet a legal obligation. A credible vendor names these exceptions rather than claiming an absolute, unqualified zero.

Sub-processors and the full data path

If the vendor's product itself calls out to an underlying model provider, or if the organization reaches the model indirectly through a cloud marketplace or a reseller, zero retention has to hold at every hop, not just the first one. Ask for the sub-processor list and confirm each party in the data path is bound by the same commitment.

How it is distinct from a no-training commitment

Retention and training use are two separate controls, and a vendor can offer one without the other. Confirm both explicitly: whether data is retained at all, and, separately, whether any retained or in-flight data is ever used to train or fine-tune a model.

Provider policies on retention, training use, and eligible products change over time and vary by account type. Confirm the current terms directly with the vendor's account or sales team and against the signed agreement, not a general web page, before making a decision.

Zero data retention is one control inside a broader AI governance program.

Treat ZDR as one line item in vendor data-handling due diligence, not the whole assessment, and keep it inside the same governance loop that owns every other AI control.

It is one input to a vendor data-handling assessment

Zero data retention answers one question in a much larger vendor evaluation: what happens to the data itself. Encryption, identity and access controls, network isolation, audit logging, and the applicable compliance attestations are separate questions an organization still has to answer, retention status aside.

It belongs inside AI governance, not outside it

Deciding which AI uses require a ZDR agreement, and confirming that a vendor's claim holds up, is exactly the kind of control an AI governance program assigns an owner and evidences over time. Treating it as a one-time procurement checkbox, rather than a control that gets revisited as vendors and agreements change, is how the assurance quietly lapses.

It is a fact an AI audit can test

A stated ZDR commitment is a claim. An audit is what verifies that the claim is backed by a signed agreement, that the covered products match what the organization actually uses, and that no undisclosed exception has crept in since the agreement was signed.

Lightbridge Automation is an independent AI-governance advisor. We help organizations decide which AI uses need a zero data retention agreement, evaluate whether a vendor's contract language actually delivers it, and fold that evaluation into a durable AI governance framework and a repeatable AI audit practice. We do not broker, hold, or represent any vendor's zero data retention agreement on a client's behalf: that agreement is always between your organization and your AI provider.

Zero data retention (ZDR): frequently asked questions

What does zero data retention (ZDR) mean?: Zero data retention means an AI provider has committed, contractually and technically, not to store, log, or otherwise retain a customer's prompts and the model's outputs beyond the processing needed to generate that specific response. Once the response is returned, the data is not kept in the provider's systems for later use, whether for debugging, abuse review, or model training. It is typically offered as a negotiated enterprise capability rather than a default setting, it applies to specific named products and access methods rather than a vendor's entire platform, and credible versions of it still name a small set of exceptions, such as a short-lived safety classification result or data kept briefly to meet a legal obligation. The point of ZDR is to remove the standing risk that comes with stored data existing anywhere outside the customer's control.
How is zero data retention different from a standard AI API's terms of service?: Under standard commercial terms, most major providers already exclude API inputs and outputs from model training by default, but that is a separate commitment from retention. Providers commonly keep API requests and responses for a limited window, often described as roughly 30 days, specifically to investigate abuse, enforce usage policies, and support debugging. That window applies automatically and is not something a customer can disable from a self-serve account. A zero data retention agreement replaces that default window with a documented commitment that data is not persisted at all beyond the immediate request, and it typically requires an enterprise agreement, provider approval, and a scope limited to the products and endpoints named in the contract.
Why do enterprises require zero data retention agreements?: The reasons cluster around four concerns. Competitive confidentiality: prompts can carry product roadmaps, source code, or deal terms that an organization does not want existing anywhere outside its own systems. Regulated data: healthcare workflows touching protected health information, financial workflows touching material non-public information, and legal workflows touching privileged material each carry handling obligations that un-bounded third-party retention puts at risk, and in healthcare a business-associate relationship under HIPAA often makes ZDR a condition of the deal. Data sovereignty: government and defense-adjacent organizations frequently treat any indefinite retention by a commercial vendor as unacceptable regardless of the vendor's intentions. And training-data leakage: even a provider that promises not to train on customer data still holds a standing risk while that data is stored, since a policy can change, an incident can expose a data store, or a future product can be built on historical logs. Zero retention removes the data before any of those risks can materialize.
Is zero data retention the same as a vendor not training on my data?: No, and treating them as the same thing is a common and consequential mistake. Not training on customer data is a commitment about how stored or in-flight data is used. Zero data retention is a commitment about whether the data is stored at all. A vendor can offer one without the other: a standard commercial plan frequently excludes customer data from training by default while still retaining it for a period for abuse monitoring and debugging. When evaluating a vendor, confirm both commitments explicitly and separately, and read the specific language in the data processing agreement rather than assuming a training exclusion implies zero retention or the reverse.
Is ZDR available on every AI plan, or only enterprise tiers?: In the well-established pattern across major LLM API providers, zero data retention is not a self-serve toggle available to every account. It is typically offered only to enterprise customers through a negotiated agreement, subject to the provider's approval, and scoped to specific eligible products or endpoints named in that agreement rather than the vendor's entire platform. An organization using a personal or standard commercial account should not assume ZDR applies just because the vendor offers it somewhere in its enterprise lineup; the commitment has to be confirmed for the specific access method actually in use.
What should we verify before trusting a vendor's zero data retention claim?: Verify six things rather than accepting a trust-center statement at face value. First, that the commitment is written into the data processing agreement or order form, not only described in marketing material. Second, exactly which products, endpoints, and access method it covers, since a commitment for one surface does not extend automatically to another. Third, that it covers both prompts and completions, not just one direction. Fourth, the specific, narrow exceptions the vendor names, such as short-lived safety classification results or legally mandated retention, since a credible claim names its exceptions rather than asserting an absolute zero. Fifth, whether any sub-processor or underlying model provider in the data path is bound by the same commitment, not just the vendor's own front end. Sixth, that retention and training-data use are confirmed as two separate, explicit commitments rather than assumed to travel together.
How does zero data retention fit into a broader AI governance program?: Zero data retention answers one question, what happens to the data, inside a much larger vendor data-handling assessment that also covers encryption, identity and access controls, network isolation, audit logging, and applicable compliance attestations. Deciding which AI uses require a ZDR agreement, tracking which vendors and products actually carry one, and re-checking the commitment as agreements and vendor products change are exactly the kind of ongoing controls an AI governance framework assigns an owner for. A stated ZDR commitment is also a claim an AI audit can test: whether a signed agreement exists, whether it covers the products actually in use, and whether any exception has crept in since it was signed.
Does Lightbridge Automation hold a zero data retention agreement with an AI vendor on our behalf?: No. Lightbridge Automation is an independent AI-governance advisor. This page is buyer education: it explains what zero data retention is, why organizations require it, and what to verify, so a team can evaluate its own vendor agreements with a clear checklist. Lightbridge Automation does not hold, broker, or represent a specific vendor's zero data retention agreement on a client's behalf, and any such agreement is a contract between the organization and its AI provider. Where Lightbridge Automation helps is in scoping which of your AI uses need a ZDR agreement, reviewing the contractual language against what your program actually requires, and folding that evaluation into a broader AI governance framework and AI audit practice.

Important notices

General advisory: this page is general buyer education from Lightbridge Automation as an independent AI-governance advisor. It is not a warranty and not a substitute for reviewing your own vendor agreements.

Not legal advice: this is not legal, audit, or accounting advice. It describes the general, well-established pattern of zero data retention offerings across major AI API providers, not the specific current terms of any single vendor. Consult your own counsel and the vendor's controlling data processing agreement before relying on any retention claim.

No Lightbridge vendor agreement claimed: nothing on this page should be read as a claim that Lightbridge Automation holds, brokers, or is a party to a zero data retention agreement with any AI vendor on behalf of a client, or as an org-level certification or compliance claim by Lightbridge Automation itself.

Freshness: vendor retention policies, eligible products, and contract terms change. This page reflects a general understanding as of mid-2026 and should be re-checked against current vendor documentation and your signed agreement before action.

Trademarks: Claude and Anthropic are trademarks of Anthropic, PBC. Other AI provider and product names referenced are trademarks of their respective owners. Lightbridge is not affiliated with, endorsed by, or sponsored by any of these vendors.

From understanding ZDR to a defensible AI vendor stack.

When the question shifts from what zero data retention means to which of your AI uses actually need it and whether your vendor contracts deliver it, Lightbridge Automation scopes the requirement, reviews the agreement, and builds it into your AI governance program.

Talk to Lightbridge Automation See AI governance