AI data retention and deletion: how do you account for every copy?
How to set practical rules for retaining and deleting AI-related material, including files, indexes, logs, exports, ownership and evidence.
Article language: English

In this article
The conversation has disappeared from the screen. That does not tell you what happened to the attachment, index, log or export. A retention plan sets rules for what is kept and for how long. It should cover the relevant places throughout the process and explain how deletion will be confirmed. Start by mapping the copies and the people responsible for them. A Delete button has a defined scope. Do not assume it removes material from every company storage location.
A plan should connect the reason for keeping material with its scope, owner and the event that triggers deletion. There is no single retention period suited to every conversation and every organisation. Legal, contractual and recordkeeping requirements need to be established with the people responsible for them. This guide covers how to carry out and check the decisions you make.
Start with the reason for keeping the material
Ask why the company needs a particular copy after a task is finished. A final result may form part of a case record, while a working file may serve only to prepare an answer. Instruction history may help reconstruct a decision, but that does not necessarily mean the entire attachment must be kept. Each purpose needs its own justification.
The process owner establishes the business need, and the appropriate people confirm the other requirements. Only then does the administrator choose the settings. Reversing that order can turn the default configuration into company policy without checking its effects. If it is unclear which document is the result and which is a working copy, first organise the knowledge architecture and the flow of materials. The guide to organising company knowledge (in Polish) explains how to distinguish document status and ownership.
List the places where copies are created
Follow a representative task from the source document to the result received. Check the file in the AI service, conversation content, task history, search index, intermediate folder and export. Establish what the integration stores and what the source service stores. Do not assume every category is present. Record confirmed locations and questions that still need answers.
Keep the location identifier, owner and means of access. A guide to mapping an AI data flow (in Polish) explains how to build the register needed at this stage. The retention plan adds rules for keeping and deleting the material. This keeps the list of copies tied to a particular process, instead of turning it into a general list of products that no one can put into action.
Separate training, feature storage and logs
A provider's statement that it does not use data for training does not settle how long a conversation, file or security log is kept. Treat these as separate questions when assessing a service. Check the plan, feature, configuration and any individual terms. Do not apply the rules for a business product to a personal account, or API terms to an intermediary's tool, without confirmation.
The OpenAI API data controls documentation distinguishes training use, logs and application state, as well as limits for particular features. It illustrates why the exact service variant matters. Record the date checked and the relevant document in your own assessment. The guide to assessing an AI provider (in Polish) covers questions about terms before purchase.
Write a rule people can carry out
For each category, specify the purpose, scope, location, owner, retention period or ending event, and deletion method. If the rule says "after the project ends", define who marks that point and how the system recognises it. If the administrator is not told that the project has ended, the planned action may never start. There needs to be a signal and someone responsible for passing it on.
Also define exceptions, how they are approved and when they will be reviewed again. Do not leave an exceptional copy without an owner: over time it may become hard to establish why it was kept. If a requirement to retain material conflicts with planned deletion, refer the discrepancy to the appropriate people. The administrator needs a decision for the specific category, rather than having to choose independently between general documents.
Check what the Delete button means
Read the documentation and run a controlled test with prepared material. Establish whether deletion applies to the conversation, attachment, sharing or the entire account. Check the confirmation available and the stated scope of the operation. If there is a trash or restore function, include it in the procedure. Do not describe deletion as immediate and complete without evidence for the specific service.
After the test, try access again through the interface, link and integration, if they belong to the process under review. The test confirms an observable effect; it does not automatically reveal the provider's internal systems. Describe the limits of the evidence honestly. Where provider confirmation is needed, record the question, document and answer that cover the relevant material.
Include indexes and derived data
A tool that searches documents may create an additional structure to find their contents. Check how it is updated after the source changes or is deleted. The assumption that deleting a document from a drive automatically removes everything from the AI service needs confirmation. Use test material with recognisable content and check the behaviour after withdrawing it.
Treat summaries, compilations and other outputs in the same way. They may contain significant parts of the source even after the input file has been deleted. A decision to keep them should depend on their content and purpose, not only on the label "AI output". Record who is responsible for withdrawing an outdated summary from the company knowledge base. If ready-made answers are stored, check AI cache invalidation after a source or access scope changes. The guide to AI cache invalidation (in Polish) explains what to examine.

Separate removing access from handling data
Disconnecting an integration limits further retrieval of material, but may not delete copies already created. The closure plan should therefore cover both actions and separate confirmation for each. First identify the authorised person, dependencies and tasks in progress. Then carry out the agreed actions in the source service and the AI product, recording their date and scope.
The guide to AI integration permissions (in Polish) describes how to check permissions and disconnect an integration. Also review automated tasks that could retrieve the material again after deletion. Unless the source of that re-creation is disabled, deletion can appear ineffective. A guide to reviewing automation (in Polish) helps identify schedules, retries and responsibility for ending them.
Account for exports and employee copies
Check where the team moves results for further work: a project folder, email, a shared document or a case management system. Appropriate company rules should apply in those places. Employee instructions need to explain what constitutes the official record and what is a working copy that needs to be sorted and handled under the applicable rules. Without that distinction, material can be deleted correctly from the AI service while undocumented exports remain in many folders.
Do not try to solve this by collecting employees' private data broadly. Base the review on the approved scope and company ways of working. If not all copies can be confirmed, describe that limitation. An AI use policy should identify approved storage locations, while team training should demonstrate the appropriate workflow on a task. The guide to an AI use policy (in Polish) and the guide to team training (in Polish) discuss how to prepare those arrangements.
What about backups and exceptions?
Establish whether the provider's documentation describes backups, how they are retired and any limits on deleting individual items. Ask about the particular service rather than assuming the same mechanism applies to every product. In the plan, distinguish removal from current access from the remaining storage processes. If some terms are unconfirmed, show how that uncertainty affects the decision to use the service.
Material connected with an incident, dispute or recordkeeping obligation may need a separate decision. That decision should identify the scope, owner and method of later review. An ordinary automated rule should not independently delete data covered by such a decision. The guide to responding to an AI data incident (in Polish) describes how to handle a suspected improper disclosure.
Practice example: closing a research project
A team is finishing work in which AI helped organise approved material. The project folder contains an approved report, the tool holds conversation history and working files, and intermediate exports remain on a drive. The owner identifies the records needed after closure and directs the other categories to the agreed cleanup process. The rules follow the chosen purpose and requirements, without one arbitrary deadline for everything.
The administrator disconnects the sources, carries out the agreed deletion and keeps confirmation of its scope. The person leading the work checks that an authorised recipient can still find and open the final report. The exercise also raises the question of an exported copy. The team decides according to its contents, rather than assuming that every output file must be kept or that each one can be deleted.
Keep evidence of the action without duplicating the content
A deletion register can contain the material category, location, identifier, the basis for the rule, date, person carrying out the action and result of the check. It does not need a copy of the document being deleted. If an error occurs, record what remains to be done, who is handling it and whether access has been restricted. The difference between the plan and what was actually done should be visible, especially for manual actions.
Check the register against several representative entries. The person reviewing it should understand what was done and what the evidence does not cover. If the operation is automatic, arrange an error signal and an owner for exceptions. The passage of the agreed period alone does not prove successful deletion. You need an observable result or appropriate service confirmation, with its limitations recorded.
How do you find a particular piece of material?
The plan is also worth checking from the point of view of identification. An authorised person should be able to determine which conversations, files and outputs concern a given task. If names are arbitrary and exports are not connected to a project, even a sound deletion rule can be hard to apply. Process identifiers and explicitly agreed storage locations help, without putting excessive data in file names.
Run an exercise with test material: find its confirmed copies, apply the relevant rule and record the limits. If finding it requires a broad review of other people's documents, revisit the process organisation and permissions. The aim is a workable way to handle a defined category of information. Do not promise to locate every fragment if the architecture or available evidence cannot support that claim.
Include AI assistant memory and its actual scope in the copy plan. The guide to auditing assistant memory (in Polish) explains what to examine. Establish whether deleting a conversation affects stored context, files and connected sources in the particular product. A feature name in the interface is no substitute for checking documentation and available evidence.
How do you keep the plan current as things change?
A new feature, integration, plan or export may create more storage locations. Tie the retention review to decisions about process changes. Record the data owner and a link to the current rule in the use-case record. Periodically check retired solutions too, so an old trial does not leave an active account and a forgotten copy. Prepare the same map for a local AI solution and its actual dependencies, rather than leaving it out simply because of where it runs. The guide to local AI and its dependencies (in Polish) explains why the model location alone does not describe the whole data flow.
A plan is useful when it covers identified locations, has rules people can carry out and distinguishes a declaration from a confirmed action. If you cannot yet say where the life of material ends, start with one process. See the Aeneia Audit offer.










