AI data retention is the practice of keeping, and then deleting, everything an AI system touches: prompts, uploaded files, outputs, logs, search indexes and backups. To set rules, split that data into categories, give each a documented purpose, an owner and a deletion trigger, then check that your AI provider's terms match. Regulators on both sides of the Channel expect exactly this, and a single blanket period for everything will not satisfy them.

This guide is written for teams that run AI in daily operations and need a retention position their DPO, their clients and their auditors can accept.

TL;DR

AI retention covers more than chat history: files, outputs, embeddings, vendor logs, analytics records and backups all count.
UK GDPR and EU GDPR both bar keeping personal data longer than the purpose needs. There is no universal period, so you set one per category and document it.
Logs may be a legal duty rather than a choice: the EU AI Act sets minimum log retention for high-risk systems.
"Zero data retention" has no standard meaning. Read what it covers, and what it leaves out, in each provider's terms.
Choose a platform where retention, training use and data location are fixed by design, then govern the rest with a schedule.

What AI data retention actually covers
How long AI tools keep prompts, outputs and logs
What the law expects from your retention rules
How to set AI data retention rules, step by step
What to check in an AI provider's terms
Retention in agent workflows
Our perspective
Run agents on an EU-based workspace
Takeaway
FAQ
Sources

What AI data retention actually covers

Most teams think of retention as chat history. In practice an AI product can hold uploaded documents, generated summaries, conversation history and records of requests sent on to a model provider. A proper inventory also has to account for embeddings, hosting logs, analytics records, older file versions and backups.

A useful way to map it is by layer:

Layer	Typical contents	Who usually controls it
Inputs	Prompts, attachments, pasted text, connected inbox or drive content	You, through configuration
Outputs	Drafts, summaries, classifications, generated files	You
Derived data	Embeddings, search indexes, knowledge base chunks	Platform and you
Operational records	Run logs, approval records, audit trails	Platform, with your settings
Provider-side records	Model provider request logs, abuse monitoring data	The model provider
Backups	Snapshots of all of the above	Platform and hosting provider

Every row needs an answer to three questions: why is it kept, for how long, and what triggers deletion.

How long AI tools keep prompts, outputs and logs

There is no single answer, and anyone quoting one figure for "AI tools" is simplifying. Retention depends on the provider, the plan or API tier, the feature (stored conversations, file search and fine-tuning each behave differently), the contract and any legal obligation to preserve records. Consumer plans, business plans and API access from the same vendor often carry different terms.

Some model providers publish a short default window for abuse monitoring on API traffic, and some offer zero data retention to eligible customers on approved endpoints. Those terms change, so take the figure from the provider's current data-controls or enterprise privacy page and from your signed contract, not from a summary. Record the date you checked.

Two practical points follow:

Retention is layered. Your platform may delete a prompt immediately while the model provider's abuse-monitoring copy, or a backup, still exists. Ask for the whole chain.
Legal holds override schedules. If litigation or an investigation requires preservation, routine deletion pauses. Your policy needs a named person who can order that.

What the law expects from your retention rules

UK GDPR and the ICO

The storage limitation principle says personal data must not be kept in identifiable form for longer than necessary for its purpose, and the ICO has said there is no universal retention period for each data type. The ICO's AI toolkit warns that without reviews and documented retention schedules, personal information may be retained inappropriately or longer than necessary, becoming inaccurate and excessive, potentially breaching UK GDPR Article 5(1)(c).

The same toolkit gives a clear method. Review the relevance of personal information at each stage of development and training and justify keeping it. Classify personal information by sensitivity and set retention periods for each category. Document schedules based on business need and statutory requirements. Standardise regular weeding, such as rolling deletion.

EU GDPR and the CNIL

If you serve EU customers, the French regulator's 2026 recommendations say that you must not keep personal data indefinitely. You define a period after which data is deleted or, in some cases, archived, and you set it according to the purpose. For data used in the development phase, the periods must be pre-planned and monitored over time, and people must be told the retention period, for example in information notices.

EU AI Act logging duties

Retention is not only about deleting. For high-risk systems, one published analysis of the EU AI Act notes that Article 12 requires automatic recording of events over the system's lifetime, and Article 19 requires providers to keep those logs for at least six months. Another reports that technical documentation must be kept for ten years after the system is placed on the market or put into service. Whether your use case is high-risk is a legal question for your counsel or DPO. Our EU AI Act compliance guide covers what a platform needs to support.

Other regimes

Outside Europe the direction is the same. Singapore's PDPC guidelines on AI recommendation and decision systems say organisations should limit personal data to the attributes and time periods necessary, and encourage pseudonymisation and a data protection impact assessment. If you operate across several countries, attribute each rule to its jurisdiction and apply the strictest one that touches your data.

How to set AI data retention rules, step by step

Legal advice from a UK firm sums up the principle: an AI retention policy needs categories, reasons and review points rather than an invented rule such as "keep everything for 90 days". Prompts, logs, training material and backups serve different purposes, so each gets a justified purpose, an accountable owner and a workable review or disposal trigger.

1. Inventory every AI data flow

List each AI workflow, the data it receives, where that data is stored, which model providers process it and which integrations feed it. Include connected tools such as email, file storage and shop platforms, because they are the source of most personal data entering an agent.

2. Classify by sensitivity and purpose

Separate ordinary business content from personal data, special category data and confidential client material. The ICO recommends classifying by sensitivity and setting a period per category, so do that before you pick any numbers.

3. Assign a period, a reason and an owner

A workable schedule looks like this. The periods are illustrations of structure, not recommendations: set yours from your own business need and legal obligations.

Data category	Purpose	Period basis	Owner	Deletion trigger
Prompts and attachments	Completing the task	Shortest period that lets the task finish and be reviewed	Workflow owner	Run completed and reviewed
Generated outputs	Business records, customer replies	Same as the record they become part of	Department head	Record schedule expiry
Run logs and audit trails	Accountability, incident review, legal duty	Longer of your policy need and any statutory minimum	DPO or compliance lead	Schedule expiry, unless on hold
Knowledge base documents	Grounding agent answers	While the source document is current	Content owner	Document withdrawn or superseded
Embeddings and indexes	Search	Tied to the source document	Platform administrator	Source deleted
Backups	Recovery	Fixed rotation cycle	IT lead	Rotation

Notice that derived data inherits the clock of its source. An embedding that outlives the document it came from is a retention failure.

4. Plan for deletion you cannot fully perform

The ICO notes that where permanent deletion is not possible because of system functionality, information should be stored securely out of reach, with access locked down, or anonymised, and training data no longer required should be removed or erased. Write down which systems fall into this group before a deletion request arrives, not after.

5. Build in erasure requests and legal holds

The EDPS generative AI orientations tell organisations to apply core data protection principles, including retention and security, and to uphold data subject rights with robust procedures for erasure. Test the procedure end to end: request, locate, delete, confirm, record.

6. Review on a fixed cycle

Set a review date for each category, and whenever you add a workflow, model or integration. Retention rules written once and never revisited drift out of line with how the tools are actually used.

What to check in an AI provider's terms

Your rules only work if the provider's behaviour matches them. The EDPS orientations say organisations deploying generative AI should perform third-party vendor due diligence and request documentation to verify compliance. Compliance depends on your internal controls and on what the vendor can evidence.

Put these questions to every provider in the chain:

How long are prompts, outputs, files and logs kept, per plan and feature?
Is customer content used to train or improve models?
Where are servers, models and integrations located?
Which subprocessors and model providers receive the data?
Can you configure or shorten retention, and can you export or delete on request?
Are audit logs available, and for how long?
What happens to backups after deletion?

Our AI vendor due diligence guide expands this into a full question set, and the guide to evaluating AI vendors under EU rules covers how to weigh the answers.

Reading "zero data retention" correctly

Zero data retention (ZDR) is a contractual or technical rule under which covered content is not kept persistently after processing. But ZDR says nothing conclusive about training, metadata, data location, third parties or features outside its scope, and each provider defines it differently, often applying it to certain accounts, projects, models or APIs only. No persistent retention does not mean no processing. Read the scope clause, then ask the training-use and data-location questions separately. Our explainer on no-training and zero-retention AI data goes through the distinctions.

Retention in agent workflows

Agents raise the stakes because they act on live business data: they read inboxes, draft replies, update records. Three design choices make retention easier to govern:

Approval before action. When a human approves anything an agent sends, books or changes, the approval itself becomes a record you can retain or delete on a defined schedule.
Audit logs by default. You cannot set a retention rule for logs that do not exist. Logs with a clear owner give you evidence for regulators and for your own incident reviews.
Narrow connections. Connect an agent only to the mailbox, folder or store it needs. Less data in means less data to retain and delete.

For a full view of the controller and processor questions around agents, read our GDPR and AI agents playbook, and use the enterprise AI data security checklist to check the surrounding controls.

Our perspective

Retention is a governance problem that product design can make smaller. Most of the pain we see comes from data that nobody decided to keep: copies in provider logs, indexes that outlive their documents, exports in personal drives.

That is why we build AgentWorks so that servers, AI models and integrations run exclusively within the EU, and so that your data is never used to train models. It is also why human approval is required for any agent action that sends, books or changes data, and why audit logs and sector-ready templates are part of the platform rather than add-ons.

Our advice: decide your categories first, pick a platform whose defaults already match them, and then spend your effort on the schedule, not on chasing copies. Governance multiplied by throughput is the real measure of an AI rollout.

Erwin

Run agents on an EU-based workspace

AgentWorks is an EU-based AI workspace where you configure agents for recurring tasks, connect tools such as Outlook, Shopify and Google Drive, and keep a human approval step on actions that change data. Pricing is usage-based: the Free plan costs 0 € one-off, the Business plan is 29 € per month per person or 290 € per year per person, and extra workflow runs are available from 5 € to 400 € one-off. See the pricing page for current terms, or look at how workflows with agents and approvals are set up.

If data location is part of your retention case, our guides on GDPR-compliant AI and EU data residency and data residency versus sovereignty explain the difference between where data sits and who can compel access to it. The AI compliance checklist for EU businesses pulls the wider obligations into one list.

Takeaway

Write a schedule by category, give every category a reason, an owner and a trigger, and make your provider's terms prove they match. Document it, review it on a cycle, and test erasure before anyone asks for it.

FAQ

How long should a business keep AI prompts and outputs?

Only as long as the purpose needs. The ICO has said there is no universal retention period for each data type, so set one per category and document the reason. Outputs that become business records follow the schedule of those records.

Does zero data retention mean my data is never processed?

No. ZDR usually means covered content is not stored persistently after the request completes. It does not on its own settle training use, metadata, data location, third parties or features outside its scope, so read the exact terms.

Do I have to keep AI logs?

Sometimes. For high-risk systems the EU AI Act requires automatic event logging and sets a minimum retention for providers. Whether your system is high-risk is a question for your DPO or legal adviser. Even where it is not mandatory, logs support accountability.

What if an AI system cannot delete data completely?

The ICO says to store it securely out of reach, with access locked down, or to anonymise it, and to remove training data you no longer need. Record those systems in advance so your response to an erasure request is already planned.

AI Data Retention: How to Set Rules That Hold Up

TL;DR

Table of Contents

What AI data retention actually covers

How long AI tools keep prompts, outputs and logs