Area 3.5

Make Retention and Deletion Reach Where AI Hides Data

When someone’s data must be deleted, does the deletion actually reach everywhere the AI put it?

Baseline: 1Enhanced: 2 3 controls

Who it applies to

Enter for every AI system. Some of its controls apply only where it processes personal data (B4). Go as deep as your system letter, and implement in your mode.

Why it exists

Set retention and deletion rules that reach the places AI quietly accumulates personal data — because the right to have data erased does not stop at the database, and neither should you. AI systems keep personal data in logs of every interaction, in stored conversation histories, and in the numeric representations they derive from data; and a deletion request that clears the original record while leaving these intact is an obligation unmet and a breach waiting to be found. Data-protection regimes in Europe and a growing number of jurisdictions give individuals the right to erasure; honour it everywhere the data went, not only where it started.

What to actually do

Set retention limits on logs and histories, and enforce them. Interaction logs and conversation histories accumulate personal data indefinitely unless you decide otherwise. Set how long each is kept, matched to why you need it, and delete on that schedule automatically — an indefinite log is an indefinite liability.
Make erasure reach the derived stores. When a person exercises the right to have their data deleted, ensure the deletion reaches the numeric representations and derived stores built from that data, not only the source record. This is the step almost every organisation misses, because those stores are invisible to a privacy process that was designed before AI created them.
Confirm your provider can delete on your behalf, and hold them to it. Where a third party processes your data, your deletion obligation extends to their systems too. Establish — in the due-diligence questionnaire and the contract from the Sourcing dimension — that the provider will delete your data and its derivatives on request, and that they can prove it.

Controls

Baseline 1 controls

Control 158

Set and enforce retention limits on inference logs and conversation histories.

BaselinePreventiveall rolesper systempersonal-data

What it is

Automatic deletion on schedule; no indefinite retention by default.

How to implement it · Minimum rules

Retention period set per store: inference logs, conversation histories, uploaded files, copies held by the provider
Period matched to purpose (e.g. 30 days for troubleshooting; longer only with a stated reason)
Automatic deletion on schedule, using the tool's or provider's retention setting where available
Conversation history switched off or shortened where the use does not need it
No indefinite retention by default; any exception recorded with owner and end date

How to start

Find the retention settings of the tool and its provider, set a period for each log and history, and switch on automatic deletion.

Evidence

The retention schedule and the configuration showing automatic deletion is active.

Mode: light / standard / full

light: Set the provider's built-in retention and history options and note the chosen periods in a spreadsheet.
standard: The system owner sets retention per store and IT checks each quarter that deletion actually runs.
full: Retention is enforced centrally in the logging platform, with deletion reports reviewed by the privacy function.

Enhanced 2 controls

Control 159

Extend erasure to derived numeric stores.

EnhancedPreventiveall rolesper system

What it is

Deletion reaches representations built from the data, not only the source record.

How to implement it · Steps

Locate: use the inventory to find every derived store built from the person's data
Delete or rebuild: remove the person's entries from each index, or rebuild it from cleaned sources
Cover copies of the same data in caches, logs and conversation histories
Verify: search the system for the person's identifiers after deletion and confirm nothing comes back
Record the request, stores covered, completion date and verification result

How to start

Add a derived-stores step to the erasure procedure, stating per system which indexes must be updated and how.

Evidence

Erasure records showing the derived stores covered and the verification search.

Mode: light / standard / full

light: Ask the provider or integrator how to remove a person's documents from the tool's index and write that into the procedure.
standard: The system owner carries out deletion in derived stores within the erasure deadline, tracked in the request log.
full: Erasure requests propagate to derived stores through recorded data lineage, with verification logged automatically.

Control 160

Secure provider deletion of your data and its derivatives, with proof.

EnhancedPreventiveall rolesper system

What it is

Established contractually and verified, not assumed.

How to implement it · Minimum structure

Due-diligence question: can the provider delete our data and its derivatives (logs, indexes, adapted model copies) on request, and how fast?
Contract clause: deletion on request and at contract end within a set period (e.g. 30 days), backups included once their cycle ends
Coverage of the provider's own subcontractors that hold the data
Proof: written deletion confirmation for each request
Verification: right to request evidence of deletion and to check it periodically

How to start

Check the current provider contract for a deletion clause that covers derivatives and, if it is missing, request an addendum or add it at renewal.

Evidence

The contract clause and the provider's deletion confirmations.

Mode: light / standard / full

light: Use the provider's standard data-processing terms, confirm they cover derivatives, and keep the deletion confirmation emails.
standard: Procurement includes the clause in every AI contract and the privacy lead requests a confirmation for each erasure.
full: Legal and supplier management maintain standard clauses and periodically verify provider deletion against the evidence received.

Evidence

The retention schedules and evidence of enforced deletion; records showing erasure reaching derived stores; provider deletion terms and confirmations.