What is zero data retention?
Zero data retention is a contractual and technical arrangement in which a system processes a request and keeps no copy of the input or the output once the response is returned. Prompts, transcripts, attachments, and generated text exist in memory for the duration of the call and are then discarded.
Under standard commercial terms, a model provider may hold prompts for around thirty days to support abuse monitoring and incident investigation. A zero data retention agreement removes that window: the same prompt is deleted at the end of inference, so there is nothing to subpoena, leak, or reuse.
How zero data retention works
Zero data retention is enforced at four layers, and a gap in any one of them defeats the other three.
The first layer is transport. Data moves over encrypted connections and is never written to an intermediate queue or cache that outlives the request. The second is inference, where the large language model endpoint is configured so prompts and completions stay in volatile memory for the duration of the call and are never persisted for abuse review or model training.
The third layer is the application. The orchestration service that assembled the prompt has to discard the assembled context, the tool responses it collected, and any AI agent memory it built during the conversation. Retention hides here more often than teams expect, because long-term memory features are designed to persist preferences across sessions, which puts them in direct conflict with the clause someone signed.
The fourth layer is analytics and logging. Observability platforms, error trackers, session replay tools, and scheduled transcript exports all copy conversation content into systems the retention clause never named, and they are usually owned by a different team.
Types of zero data retention
Provider-level retention controls: The model vendor commits that prompts and completions are discarded at the end of inference, usually through a flagged account or a dedicated endpoint.
Platform-level zero retention: The application keeps no transcript, ticket note, or session record once the conversation closes, which also removes the material most quality programs depend on.
Field-scoped retention: The system stores the conversation but strips identifiers first, using PII redaction so what persists carries no personal data, a weaker guarantee that still satisfies many reviewers.
Time-bounded retention: Records are held for a fixed short window and then hard-deleted, which teams frequently market as zero retention even though a live window exists.
Zero data retention vs data minimization vs data residency vs de-identification
Security questionnaires stack these four terms in the same paragraph, and buyers routinely accept one when they needed another. Data minimization limits which fields are collected in the first place. Data residency fixes the geography the data sits in without shortening how long it sits there. De-identification keeps the record and removes the identifiers that make it personal. Zero data retention takes the remaining position: collection scope and geography stay as they are, and the record itself stops existing once the response is delivered.
What it holds | Ownership | Who reads it | AI-retrievable | Choose it when | |
|---|---|---|---|---|---|
Zero data retention | Nothing once the response is returned | Contract between customer and provider | Security reviewers, legal, procurement | No, nothing survives the request | Conversation content must not exist afterward |
Data minimization | Only the fields a task genuinely needs | Product and engineering design decisions | Whoever the data was collected to serve | Yes, whatever remains is stored | Collection scope is the real exposure |
Data residency | Everything, inside one named region | Infrastructure and hosting agreements | In-region systems and in-region staff | Yes, within that region only | Cross-border transfer is the blocker |
De-identification | The full record, minus direct identifiers | Data platform or analytics engineering | Analysts, evaluators, model trainers | Yes, in stripped form | Analysis matters more than deletion |
Pick zero data retention when the answer to "what happens if this system is breached next year" has to be that there was nothing there. If you still need transcripts for quality review, field-scoped redaction and a short deletion window will carry you further.
Why zero data retention matters for customer experience
Support conversations are unusually rich targets. In a single thread a customer will paste an order number, a partial card reference, a diagnosis, a dispute history, and an address, because that is what resolving their problem takes. Every stored copy of that thread is an asset somebody has to defend for as long as it exists.
When retention is unbounded, the failure is quiet. A vendor keeps transcripts to improve its models, a subprocessor keeps them longer than the vendor does, and two years of conversations sit in a bucket nobody has an owner for. Nothing goes wrong until it does, and then the exposure is retrospective.
The tradeoff deserves plain statement: deleting everything also deletes the evidence teams use to review answer quality, investigate a complaint, and recognize a returning customer. That is a live case of the competing priorities of cost and quality, and the resolution is a deliberate choice about which risk you would rather carry.
How is zero data retention measured?
Retention is measured as a window, and the only passing value is zero. Three checks establish it: the documented deletion point for every system the data touches, the list of subprocessors that inherit the same clause, and audit evidence showing the control operated continuously across a review period.
The number worth watching sits on the risk side of the ledger. Verizon's 2025 Data Breach Investigations Report recorded third-party involvement in 30% of breaches, up from 15% the year before, and retained conversation data is precisely the kind of asset that travels down a supplier chain.
Verification then becomes an evidence exercise: the retention setting in the provider console, the clause in the data processing agreement, and an independent report naming the systems in scope. A vendor that produces the first two and never the third has described an intention.
How AI agents change zero data retention
One AI agent turn touches more systems than a scripted chat widget ever did. The message is enriched with account data from a CRM, matched against a knowledge source, sent to a model, sometimes passed to a second model for a safety check, and used to write back to a system of record. Each hop is a candidate retention point, and a clause signed with the model vendor governs exactly one of them.
Agents also improve by learning from what they handled, which is a retention argument pointed the other way. Evaluation sets, failure analysis, and fine-tuning all want the transcript that zero retention destroys. Teams that need both outcomes split the paths, keeping the live conversation ephemeral while a consented, redacted sample feeds evaluation, an approach covered in this guide to anonymizing customer data for training.
What to look for in a zero data retention agreement
Scope is the first axis. Ask which named systems the clause binds and which are silently excluded, because most agreements cover inference and stop there. Then walk the subprocessor chain: telephony, transcription, vector storage, analytics, and any escalation destination each hold conversation content.
Integration surface decides feasibility. A CRM write-back, a ticket record, and a callback queue all persist by design, so decide which of them stay outside the boundary and say so in writing.
Governance matters next: name the person who signs the data processing agreement and the person notified when it changes. Regulated buyers want to know how the deletion control is evidenced, and a SOC 2 Type II report is the artifact they ask for by name.
The constraint that bites later is change control. Providers add subprocessors mid-term, and unless the agreement obliges advance notice, your retention posture can move without anything in your own stack changing.
Zero data retention and AI compliance
Zero data retention is one control inside AI compliance, the broader practice of governing how models are chosen, documented, and monitored. Deleting the transcript answers the storage question and leaves the answer-quality and accountability questions entirely open, which is why a retention clause alone rarely clears a governance review.
In healthcare support, HIPAA compliance is the frame buyers use, and their questionnaires trace where protected health information lands at every hop, including the hops that happen after the model vendor has finished with it.
What does zero data retention mean in plain terms?
Think of zero data retention as a whiteboard wiped the moment the meeting ends. Everyone in the room saw the numbers, the work got done, and nothing is left on the wall for the next group to read. ZDR, the abbreviation you will meet in vendor documentation, stands for zero data retention, and the full form is the entire promise.
Without it, someone photographs the board after every meeting and files the photos somewhere, and the person who set up that folder has since left. Two quiet years pass, then one folder is exposed and two years of meetings go with it.
What you give up is memory. A wiped board means nobody can check what was written last Tuesday, so when a customer insists the agent promised them a full refund, you have their account of it and a timestamp.
Common zero data retention mistakes
Signing with the model vendor and calling the stack covered is the most common error. The contract scope covers one hop in a pipeline with six, and the transcription service or the analytics tool is the one still holding conversation text months later.
Treating encryption as deletion is the second. Encrypted data at rest is retained data with a lock on it, and a key exists somewhere by definition, which is why breach notification obligations still attach to it.
Forgetting the escalation path is the third. The moment an agent hands a case to a person, the conversation is written into the helpdesk and inherits that system's retention policy, no matter what the AI layer promised.
The fourth is adopting zero retention with no incident plan. The first serious complaint arrives, nobody can reconstruct what the agent said, and retention gets switched back on under pressure with none of the controls that were meant to accompany it.
What does a zero data retention policy actually cover?
A zero data retention policy covers the specific systems named in the agreement, which is usually the model inference endpoint. Application logs, transcription services, analytics tools, vector stores, and the helpdesk a case escalates into are separate systems with separate retention settings, and each one has to be examined and documented on its own.
What is the difference between zero data retention and deleting data on request?
Zero data retention removes the data automatically at the end of every request, so no copy ever accumulates. Deletion on request means data is stored by default and erased when someone asks, which leaves a live window during which the record exists, can be copied downstream, and can be exposed in a breach.
Is zero data retention the same as anonymization?
Zero data retention and anonymization solve different halves of the problem. Anonymization keeps the record and strips the identifiers that link it to a person, preserving the content for analysis. Zero data retention preserves nothing at all. Anonymized data can still be re-identified under some conditions; deleted data cannot.
Does zero data retention mean my data is never used for training?
Zero data retention normally implies no training use, since training requires stored examples. Confirm it explicitly anyway, because the two commitments are separate contract clauses and some agreements permit derived artifacts such as embeddings, safety classifications, or aggregate telemetry to persist after the raw text is deleted.
Can an AI agent remember a customer under zero data retention?
An AI agent under zero data retention cannot recall past conversations, since recall depends on stored history. Within a single session it holds context in working memory, which disappears when the session ends. Teams that need continuity usually keep identity and entitlement data in their own CRM and pass it in per request.
How do you verify a vendor's zero data retention claim?
Verifying a zero data retention claim takes three artifacts: the retention clause in the data processing agreement, a current subprocessor list showing which parties inherit it, and an independent audit report covering the systems in scope. A configuration screenshot documents a setting on one day and proves nothing about the period.